CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
A Critical Analysis on Prosthetic Arm
Anil kumar
B.Tech (EI 3 yr)
Galgotia‟s college of Eng. & Tech.
Greater Noida
anil00699@gmail.com
Abstract – Prosthetic limbs are a real
time challenging and demanding
biomechatronics field due to more and
more day-to-day basis work objectives for
amputees. This paper investigates the
design of various prosthetic arms to
mimic its human counterpart. The
comparative analysis of available
technology and research work in field of
bionic arms are being done. In this
analysis, the merits and demerits of
various prosthetic hand for real time
application is being highlighted. In-depth
analysis shows that further enhancement
could be done in order to augment DOF,
grasp control, available strength of the
artificial arm and signal processing
techniques.
Keywords- Prosthetic Arm, Robotic Arm,
Artificial hand
INTRODUCTION
Prosthesis is an artificial extension that
replaces a missing body part. Prostheses are
typically used to replace parts lost by injury
or missing from birth or to supplement
defective body parts. In addition to the
standard artificial limb for every-day use,
many amputees have special limbs and
devices to aid in the participation of sports
and recreational activities. Myoelectric
Dr. Monika Jain
Professor-Electronics & Instrumentation
Galgotia‟s college of Eng. & Tech.
Greater Noida
monika_jain24@rediffmail.com
prosthetic hand is used as natural hand by
amputees (who lost their natural hand) by
amputation. Initially, hooks are being used
for this purpose. In recent time, with the
advancement in technology, artificial smart
arm are being used. In myoelectric hands,
electromyography (EMG) signals are used
as control input. EMG signals are
bioelectrical signal released by muscles.
They are normally of 10-500 Hz frequency
and up to 10 mV. The main part of signal is
near 450Hz. conventionally, surface
myoelectrodes are being used in myoelectric
prosthetic arms, however with help of SRI
EMG technique, signal strength and
accuracy of signal would increase. This can
be use to control individual fingers. If
prosthetic hand can be made of same
mechanical properties as a neuromuscular
control system of hand then amputee can use
it more effectively. A lot of work is done in
this field to improve prosthetic arms but
feedback is not used effectively. By using
feedback connected at artificial fingers like
pressure sensor, accelerometer, temperature
sensor, hand‟s working can be improved.
Also DOF and grasping capacity of artificial
arm can be improved. There is no doubt that
microelectronics offers numerous
opportunities for improvement in the
engineering of substitutes for the human
limb. The possibilities become exciting as
soon as we move away from the all-
mechanical prostheses which have been the
COPO1O1-1
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
conventional prescription until today. We
have described a relatively simple
application of hybrid microelectronics in
this novel field.
The purpose of this project is to develop a
myoelectric controlled individual finger,
multi-DOF adaptive grasping technique
equipped commercial hand capable to fast
response. Also amputee can feel touch,
temperature and force applied by prosthetic
hand on the joint of prosthetic and natural
hand.
LITERATURE SURVEY
In the field of prosthetic arm, lots of
research work is in progress. Maximum
prosthetic hands are simple grippers with
one or two degrees of freedom (DOF). They
are using smart hooks (passive fingers and
thumb) like Otto Bock [1] they have 2 or 3
point of contact so more force is required for
grasping. Also they are capable of gripping
as in myohand variplus speed [2] but not
capable of some more works that include
finger, like door opening, rotating car key
etc. Kenzo Akazawa [3] developed a hand
using dynamic behaviour of antagonist
muscles, flexor muscles and extensor
muscles but is similar as in [1]. Also its
response is slow and a large training session
is required to train the amputee. For fast
processing, Isamu [4] has presented
evolvable hardware chips with genetic
algorithm. It takes less time to train
amputee. But it restricts the DOF of hand.
Also dedicated chip is required for
prosthetic hand which reduces versatility to
use high speed microcontrollers.
Microprocessor and high torque motor based
hand is being made by Ryuhei [5], however,
this uses only two surface EMG signals, due
to which less DOF is obtained. Also, more
training required for amputee. Another main
drawback in above is that they uses EMG
signal as stationary signal but for multi DOF
it is required to process it as non stationary
signal. Real time EMG pattern recognition
[6] is done to improve the hand functioning.
But still signal is not sufficient so fingers
can‟t be controlled separately. For this
purpose SRI EMG (surface recorded
intramuscular EMG [7] signals can be used
they are more accurate than surface EMG
signals. Prosthetic hand with individual
finger movement will be more similar to
natural hand. Prosthetic hand approximately
similar to natural hand is made by Bryan
Christie [8] and Dean Kamen [9] with the
help of DARPA. But it is connected by
nerve system so not easy to use also it‟s still
in experimental state. Chappell [10] has
presented an approach having an artificial
hand with sensors allowing for the inclusion
of automatic control loops, freeing the user
from the cognitive burden of object holding
which is similar to the natural low level
spinal loops that automatically compensate
for object movement. Force, object slip and
finger positions are variables that need to be
measured in a hand designed for the
physically impaired person. It shows that a
high specification sensor is required for
designing an arm. Also it must be designed
separately for each amputee. But latest
technology provides adaptive signal
processing helping to adapt the amputees‟
requirements. An electrically driven locking
mechanism has been built by Law and
Hewson [11], which is controlled by the
electromyogram (EMG) of the surviving
muscles in the upper arm. Hybrid
technology is used for the construction of
the associated electronic circuitry. Many
similar applications are now being
considered in an attempt to improve the
performance of upper-limb prostheses using
latest researches. Development, testing and
experimentation work of a device for the
hand rehabilitation was done by Mulas et al
COPO1O1-2
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
[12]. The system designed is intended for
people who have partially lost the ability to
control correctly the hand musculature, for
example after a stroke or a spinal cord
injury. Based on EMG signals the system
can "understand" the subject volition to
move the hand and actuators can help the
fingers movement in order to perform the
task. This paper describes the device and
discusses the first results conducted on a
healthy volunteer. It requires a number of
actuators to increase the DOF of prosthetic
arm. Also EMG processing done is not
sufficient to provide significant
performance. Massa et al [13] designed a
hand to augment the dexterity of traditional
prosthetic hands while maintaining
approximately the same dimension and
weight. This approach is aimed at providing
enhanced grasping capabilities and natural
sensory-motor coordination to the amputee,
by integrating miniature mechanisms,
sensors, actuators, and embedded control. A
biomechatronics hand prototype with three
fingers and a total of six independent DOFs
has been designed and fabricated. This
research work is focused on the actuators
system, which is based on miniature
electromagnetic motors. However, still not
using better EMG processing technologies
which can dramatically increase
performance of the hand also the grasping
force is low because of the limited torque
generated by the miniature actuators used
for the hand (which are among the best
available on the market in that range of
size). Embedded control architecture for the
action and perception of an
anthropomorphic 16 degree of freedom, 4
degree of actuation prosthetic hand for use
by transradial amputees has also been
reported. The prosthetic hand is provided
with 40 structurally integrated sensors useful
both for automatic grasp control and for
biofeedback delivery to the user through an
appropriate interface (either neural or non-
invasive). Cipriani et al [14] briefly
describes the mechatronic design of the
prosthesis, the set of sensors embedded in
the hand and finally focuses on the design of
the control architecture that allows action
and perception for such a sophisticated
device. It consists of 8-bit architecture
microcontrollers, however, not using the
available signal processing techniques.
Herrera et al [15] designed and constructed
a prosthesis that will be strong and reliable,
while still offering control on the force
exerted by the artificial hand. The design
had to account for mechanical and electrical
design reliability and size. These goals were
targeted by using EMG in the electrical
control system and a linear motion approach
in the mechanical system. The prosthetic
gripper uses EMG to detect the amputee's
intended movement. The control system
requires an adaptation mechanism for each
amputee's characteristics. Gordon et al [16]
used proportional myoelectric control of a
one-dimensional virtual object to investigate
differences in efferent control between the
proximal and distal muscles of the upper
limbs. Restricted movement was allowed
while recording EMG signals from elbow or
wrist flexors/extensors during isometric
contractions. Subjects used this proportional
EMG control to move the virtual object
through two tracking tasks, one with a static
target and one with a moving target (i.e., a
sine wave). Eriksson et al [17] studied the
neural network feasibility for categorizing
patterns of EMG signals. The signals
recorded by the surface electrodes are
sufficient to control the movements of a
virtual prosthesis. The presented method
offers great potential for the development of
future hand prostheses. A signal processing
system based on RAM as a look-up table
(LUT) has been presented by Torresen et al
[18]. This provides a fast response besides
being compact in size. Several algorithms
for programming it have been proposed For
COPO1O1-3
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
the given data set used in the following
experiments, the time used for programming
the RAM was approximately equal to the
time used for training a feed-forward neural
network – solving the same problem in the
best possible way. However, the main
advantage of this scheme is the fast runtime
speed. Ferguson [19] described the
development of a system that will allow
complex grasp shapes to be identified based
on natural muscle movement. The
application of this system can be extended to
a general device controller where input is
obtained from forearm muscle, measured
using surface electrodes. This system
provides the advantage of being less
fatiguing than traditional input devices. V.
Tawiwat et al. [20] applied a mouse„s roller
with a gripper to increase the efficiency of a
gripper could lead to material handling
without slipping. To apply a gripper, the
optimization principle is used to develop
material handling by use of a signal for
checking a roller mouse that rotates. In case
the roller rotates, meaning that the material
slips. A gripper will slide to material
handling until the roller does not rotate. In
an attempt to improve the functionality of a
prosthetic hand device, a new fingertip has
been developed, that incorporates sensors to
measure temperature and grip force and to
detect the onset of object-slip from the hand.
The sensors have been implemented using
thick-film printing technology and exploit
the piezoresistive characteristics of
commercially available screen printing
resistor pastes and the piezoelectric
properties of proprietary lead-zirconate-
titanate (PZT) formulated pastes. The force
sensor exhibits a highly linear response to
forces. The force sensor response is also
extremely stable with temperature. The
ability of the piezoelectric PZT vibration
sensor to detect small vibrations of the
cantilever, indicative of object slip, has also
been demonstrated [21]. Externally powered
upper extremity prosthesis has been
considered as a system in which the
necessary components to design a better
prosthetic arm are viewed and divided into
four subsystems: input, effecter, feedback
and support. Current research is reviewed in
terms of these subsystems. Each subsystem
performs its own task, but they are related to
each other and together they function to
make up a prosthetic upper extremity, which
provides the movement to the amputee [22].
Hands such as the All Electric Prosthetic
Hand utilize a series of gears to transmit the
motion of motors housed in the forearm to
the relevant fingers [23]. Other designs have
the actuators transmitting the power directly
to the joint. An example of this is the
Anthroform Arm, which uses pneumatic
„muscles‟ mimicking the muscles of the
human arm connected directly to the „bones‟
it moves [24]. Shape Memory Alloy wires
are also used, to both provide the force and
transmit the motion. SMA wires contract
when heated and return to their initial shape
when cooled [25]. This method of actuation
is utilized in the Shape Memory Alloy
Activated Hand constructed by DeLaurentis
et al [26].
CONCLUSIONS
A lot of work is done for developing the
prosthetic arm but still more precision work
could be done. Grasping technique can be
enhanced so that amputees can work more
effectively, response time of available arms
is not sufficient. Still the available prosthetic
arms are not comparable with natural hand
in terms of DOF. The main limitations are
less space available for motors and their low
size to torque ratio. Power consumed is also
more if number of motors increases. The
lack of enhanced signal processing
techniques in artificial hands and associated
controllers has limited their functionality
COPO1O1-4
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
and technological progress. Still a lot of
work can be done to improve the available
prosthetic arms. Use of hydraulics in place
of motors can provide great amount of
controlled force in lesser space. Due to this
available control of force applied by the
hand on the object, amputees can handle soft
or brittle items easily. By effective use of
hydraulics, DOF can also be increased
without any increased power consumption.
Also use of adaptive signal processing
techniques can improve overall performance
of artificial hand.
REFERENCES
[1]Available on:
http://www.ottobock.com.au/cps/rde/xchg/o
b_au_en/hs.xsl/384.html
[2Available on:
http://www.ottobock.com.au/cps/rde/xchg/o
b_au_en/hs.xsl/19932.html
[3] Kenzo Akazawa, Ryuhei Okuno and
Masaki Yoshida “Biomimetic EMG-
prosthetic-hand” 18th Annual International
Conference of the IEEE Engineering in
Medicine and Biology Society Amsterda,
pp. 535-536.
[4] Isamu Kajitani and Tsukuba Masahiro
Murakawa “An Evolvable Hardware Chip
for Prosthetic Hand Controller”
Microelectronics for Neural, Fuzzy and Bio-
Inspired Systems, 1999. MicroNeuro' 99.
Proceedings of the Seventh International
Conference on pp. 179-186.
[5] Ryuhei Okunol, Masahiro Fujikawa',
Masaki Yoshida2, Keno Akazawal
“Biomimetic hand prosthesis with easily
programmable microprocessor and high
torque motor” Engineering in Medicine and
Biology Society, 2003. Proceedings of the
25th Annual International Conference of the
IEEE Vol.2 pp.1674-1677.
[6] Jun-Uk Chu, Inhyuk Moon, Member,
IEEE, and Mu-Seong Mun “A Supervised
Feature Projection for Real-Time
Multifunction Myoelectric Hand Control”
International Conference of the IEEE
Engineering in Medicine and Biology
Society pp.282-290.
[7] Nikolay S. Stoykov, Madeleine M.
Lowery, Charles J. Heckman, Allen Taflove,
and Todd A. Kuiken “Recording
Intramuscular EMG Signals Using Surface
Electrodes” Proceedings of the 2005 IEEE
9th International Conference on
Rehabilitation Robotics June 28 - July 1,
2005, Chicago, IL, USA pp. 291-294.
[8] Available on:
http://spectrum.ieee.org/robotics/medical-
robots/winner-the-revolution-will-be-
prosthetized/2
[9] Available on:
http://spectrum.ieee.org/biomedical/bionics/
dean-kamens-luke-arm-prosthesis-readies-
for-clinical-trials
[10] P H Chappell, “A fist full of sensors”
Journal of Physics: Conference Series, vol.
15, 2005, pp. 7-12.
[11] H.T. Law and J. J. Hewson, “An
Electromyographically Controlled Elbow
Locking Mechanism for an Upper Limb
Prosthesis”, Electro component Science and
Technology, vol. 10, 1983, pp. 87-93.
[12] Marcello Mulas, Michele Folgheraiter
and Giuseppina Gini, “An EMG-controlled
Exoskeleton for Hand Rehabilitation”, IEEE
9th International Conference on
Rehabilitation Robotics, Chicago, IL, USA,
June 28 - July 1, 2005, pp 371-374.
[13] M. C. Carrozza, S. Micera, R.
Lazzarini, M. Zecca and P. Dario,
COPO1O1-5
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
“Development of a Novel Prosthetic Hand”
Ongoing Research and Preliminary
Results‖ IEEE/ASME Transactions on
Mechatronics, vol. 7, 2002, pp. 108-114.
[14] Christian Cipriani, Marco Controzzi,
Fabrizio Vecchi and M. Chiara Carrozza,
“Embedded Hardware Architecture Based
on Microcontrollers for the Action and
Perception of a Transradial Prosthesis”,
International Conference on Biomedical
Robotics and Biomechatronics Scottsdale,
October 19-22, 2008, AZ, USA, pp. 848-
853.
[15] Andres Herrera, Andres Bernal, David
Isaza and Malek Adjouadi, “Design of an
Electrical Prosthetic Gripper Using EMG
and Linear Motion Approach”, National
Science Foundation Grants EIA-9906600
and HRD-0317692.
[16] Keith E. Gordon, Daniel P. Ferris,
“Proportional Myoelectric Control of a
Virtual Object to Investigate Human
Efferent Control”, Exp Brain Res, vol. 159,
2004, pp 478–486.
[17] Lars Eriksson, Fredrik Sebelius,
Christian Balkenius, “Neural Control of a
Virtual Prosthesis”, IEEE 8th International
Conference on Artificial Neural Networks,
vol. 2, 1998, pp. 905-910.
[18] Jim Torresen, Vidar Engh Skaugen, “A
Signal Processing Architecture Based on
RAM Technology”, 16th European
Simulation Multiconference on Modelling
and Simulation, ISBN:90-77039-07-4, pp.
317-319.
[19] Simon Ferguson, G Reg Dunlop,”Grasp
Recognition from Myoelectric Signal”,
Australasian Conference on Robotics and
Automation, Nov 27-29, 2002, Auckland,
pp.83-87.
[20] V. Tawiwat, and S. Sarawut, “Optimal
Design for Grip Force of Material
Handling”, Proceedings of World Academy
of Science, Engineering and Technology,
July, 2008, pp 864-868.
[21] A. Cranny , D.P.J. Cotton, P.H.
Chappell, S.P. Beeby, N.M. White, “Thick-
film Force and Slip Sensors for a Prosthetic
Hand”, Sensor and Actuator 2005, pp. 162-
171.
[22] HaeOck Lee, Dawnlee J. Roberson, “A
Systemic View of an Upper Extremity
Prosthesis”, IEEE International Conference
on System of Systems Engineering, April
2007, pp. 1-6.
[23] I. Kato and K. Sadamoto (ed.),
“Mechanical Hands Illustrated”, Survey
Japan (1982).
[24] G. K. Klute, J. M. Czerniecki and B.
Hannaford, “McKibben Artificial Muscles:
Pneumatic Actuators with Biomechanical
Intelligence”, [Online, accessed 26 March
2001]
URL:
http://brl.ee.washington.edu/BRL/devices/m
ckibben/aim99v6.pdf
[25] J. L. Banks, “Anthropomorphic Robotic
Finger Platform Based on Shape Memory
Alloy”, Master‟s Thesis, MIT Artificial
Intelligence Laboratory, [Online, accessed 4
May 2001]
URL:
http://www.ai.mit.edu/projects/humanoid-
robotics-group/Abstracts2000/banks.pdf
[26] K. J. DeLaurentis, C. Mavroidis and C.
Pfeiffer, “Development of a Shape Memory
Alloy Activated Robotic Hand”, [Online,
accessed 19 March 2001] URL:
COPO1O1-6
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

http://www.cronos.rutgers.edu/~mavro/pape
rs/act2000.pdf

COPO1O1-7
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Analysis of Bio-inspired Algorithms–

BFO, PSO and FFA
Vipul Singhal
Vipul.vanshi@gmail.com

Delhi College of Engineering, Delhi University, Delhi, India

ABSTRACT emergence. Advancements and research areas then

Engineering problems can be defined as search for
one near optimal description among many
possibilities, under real time constraints. Search
techniques such as Bacterial foraging, Particle swarm
and Firefly based optimization algorithms are now
among the latest research topics. This paper throws
light on these bio-inspired evolutionary algorithms
that are used to optimize the objective function, given
constraints. Optimization as an engineering problem
has been exposed as their one application along with
the several advancements made in this field. These
techniques are explained in detail along with work
and research that has been done in that area. It also
provides the insight of modification and
improvements, these algorithms continue to enjoy
performing better for specified problem.
1. INTRODUCTION
exposed.
2. BACTERIAL FORAGING OPTIMIZATION:
Bacterial foraging algorithm is inspired by the
pattern exhibited by bacterial foraging behaviours
[1]. They tend to maximise energy per unit time.
Bacteria have the tendency to gather to the nutrient-
rich areas by an activity called “chemotaxis”. It is
known that bacteria swim by rotating whip-like
flagella driven by a reversible motor embedded in
the cell wall. E. coli has 8-10 flagella placed
randomly on a cell body. When all flagella rotate
counter clockwise, the move is called Run. When
the flagella rotate clockwise, they all pull on the
bacterium in different directions, which causes the
bacteria to Tumble.

Optimization has been an active area of research

for several decades. As many real-world
optimization problems become increasingly
complex, better optimization algorithms are
needed. Over the past few decades, many
biologically inspired computational methodologies
have been invented, such as evolutionary
algorithms which include genetic algorithms,
bacterial foraging, particle swarm firefly algorithms
and many others. These tend to enjoy a regular
attention by scientist in several research areas of FIG. 1.1 SWIM AND TUMBLE OF A BACTERIUM [1]
.
soft computing and others as well.
The foraging algorithm is based on the behaviour
Here in this paper we have taken review of three
of the E. coli bacteria, which has tree fundamental
basic algorithms and have seen how they can be
steps: chemotaxis, reproduction, and elimination
applied to several problems. Section 1 contributes
and dispersion. Chemotaxis is used in order to
to explain basic and classic algorithm of BFA
move and search for food, this step has much
which is then extended to provide its further
resemblance with a biased random-walk model [2].
applications and research areas. It tells about how
In order to maximise energy in minimum time
the bacteria forage by maximising their energy
bacteria tend to follow different search strategies as
intake in minimum span of time. This concept then
described in fig. 1.2
is used to keep them moving and later reproduce to
give next generation. It also provides that those
with poor foraging strategy tend to eliminate.
Section 2 explains basics and classic particle
swarm optimization technique. It also throws light
on the work done in that area and its applications.
Section 3 then explains firefly algorithm and its

COPO1O2-1
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
FIGURE 1.2 SEARCH STRATEGIES FOR FORAGING
BACTERIA [1]
The chemotactic operator employed in BFOA is
supposed to guide the swarm to converge toward
optima. Reproduction is used to make an identical
copy; and elimination and dispersion is used to
avoid noxious elements that can kill or spread a
group of bacteria in the search space.
2.1. CLASSICAL ALGORITHM:
When bacteria get food in sufficient amount, they
increase in length, and in presence of suitable
temperature, they break in the middle to form an
exact replica of themselves. This phenomenon
inspired Passino[1] to introduce an event of
reproduction in BFOA. Due to the occurrence of
sudden environmental changes or attack, the
chemotactic progress may halt, and a group of
bacteria may move to some other places. This
constitutes the event of elimination–dispersal in the
real bacterial population, where all the bacteria in a
region are killed or a group is dispersed into a new
part of the environment. As said, bacteria try to
move to areas which are rich in nutrients and free
of noxious substances [4]. If θ is a position and J(θ)
is a function which is the more negative the more
the nutrients prevail over the noxious substances,
then we can imagine that bacteria naturally try to
find those, where J (θ) has its minima.
A bacterium position after a tumble can be
determined through (1), where the position at a
given instant is calculated in terms of the position
at the previous instant and the step size C(i) applied
in a random direction ϕ (j), generated by the
bacterium tumble
θi(j+1,k,l)= θi(j,k,l)+C(i)* ϕ(j,k,l) (1)
Where i indicate the bacterium in a total population
of S individuals, j the chemotactic step, the
reproductive step k, and the elimination and
dispersal step l.
The health function is calculated at these positions
of bacteria as (2)
J(i,j,k,l)=J(i,j,k,l)+Jcc(θ’(j,k,l),P(j,k,l)). (2)
Here second expression shows the effect of
attraction and repulsion among the bacteria. The
exact expression of Jcc(θ’(j,k,l),P(j,k,l)) is in [1].
And this then is compared with the previous health
function as
J (θi(j+1,k,l))<J(θi(j,k,l))
If it’s true then the bacterium will keep moving in
the same direction (“run” behaviour) for a given
maximum number of steps, after which a “tumble”
will occur anyway. The least healthy bacteria
eventually die while each of the healthier bacteria
(those yielding lower value of the objective
function) asexually split into two bacteria, which
are then placed in the same location. This keeps the
swarm size constant. This is done through
reproduction step. Some bacteria are liquidated at
random with a very small probability while the new
replacements are randomly initialized over the
search space.
2.2. PSEUDOCODE:
1. Initialize parameters n, N, Nc, Ns, Nre, Ned,
Ped,
C(i)(i=1,2…N), i =1,2,---
Where,
n: Dimension of the search space,
N: The number of bacteria in the population,
NC : Chemotactic steps,
Nre: The number of reproduction steps,
Ned : the number of elimination-dispersal events,
Ped: Elimination-dispersal with probability,
C (i): the size of the step taken in the random
direction specified by the tumble.
2. Elimination-dispersal loop: l=l+1
3. Reproduction loop: k=k+1
4. Chemotaxis loop: j=j+1
[a] For i =1,2…N, take a chemotactic step for
bacterium i as follows.
[b] Compute fitness function, J (i, j, k, l).
Let,J (i, j, k, l) =J (i, j, k, l)+Jcc(θi( j, k, l), P( j, k, l))
(i.e. add on the cell-to cell attractant–repellent
profile to simulate the swarming behaviour )
.Where, Jcc is defined in .
[c] Let Jlast=J (i, j, k, l) to save this value since we
may find a better cost via a run.
[d] Tumble: generate a random vector Δ(i)€ Rn
with each element ∆m(i),m =1,2,..., p, a random
number on
[-1, 1].
[e] Move:
Let
COPO1O2-2
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
This results in a step of size C(i) in the direction of
the tumble for bacterium i.
[f] Compute J (i, j +1, k, l) and
Let J(i, j,k,l)=J(i, j,k,l)+Jcc( θi( j,k,l),P( j,k,l))
[g] Swim
i) Let m=0 (counter for swim length).
ii) While m<Ns(if have not climbed down too long).
• Let m=m+1.
• If J (i, j +1, k, l)<Jlast( if doing better), let
Jlast= J (i, j+1, k, l) and
Let
And use this θi(i+1, j, k) to compute the new J (i,
j+1, k, l) as we did in [f]
Else, let m= Ns. This is the end of the while
statement.
[h] Go to next bacterium (i, 1) if i≠N (i.e., go to [b]
to process the next bacterium).
5. If j<Nc, go to Step 3. In this case, continue
chemotaxis, since the life of the bacteria is not
over.
6. Reproduction:
[a] For the given k and l, and for each i =1,2,...,N,
Let
be the health of the bacterium i (a measure of how
many nutrients it got over its lifetime and how
successful it was at avoiding noxious substances).
Sort bacteria and chemotactic parameters C(i) in
order of ascending cost health J (higher cost means
lower health).
[b] The Sr bacteria with the highest Jhealth values die
and the remaining Sr bacteria with the best values
split (this process is performed by the copies that
are made are placed at the same location as their
parent).
7. If k<Nre, go to Step 3. In this case, we have not
reached the number of specified reproduction steps,
so we start the next generation of the chemotactic
loop.
8. Elimination-dispersal: For i=1,2...,N, with
probability Ped, eliminate and disperse each
bacterium, and this result in keeping the number of
bacteria in the population constant. To do this, if a
bacterium is eliminated, simply disperse one to a
random location on the optimization domain. If
l<Ned , then go to Step. 2 ; otherwise end.
2.3. ADVANCEMENTS IN BFO AND ITS APPLICATION AND
RESEARCH AREAS:
Vast applications have been found where BFO has
shown remarkable results and has been modified
for different problems according to the objective
function. Initial applications of evolutionary
algorithm were meant for static optimization
problems but in recent years the emergence of
another member of the EA family[5]– bacterial
foraging algorithm (BFA), the self adaptability of
individuals in the group searching activities has
attracted a great deal of interests including dynamic
problems. W. J. Tang and Q. H. Wu have
contributed their work by proposing DBFA, which
is especially designed to deal with dynamic
optimization problems, combining the advantage of
both local search in BFA and a new selection
scheme for diversity generating. They used the
moving peaks benchmark (MPB) [6] as the test bed
for experiments. The performance of the DBFA is
evaluated in two ways. The first is concerned with
the convergence of the algorithm in random
periodical changes in an environment, which are
divided into three ranges from a low probability of
changes to a higher one. The second is testing a set
of combinations of the algorithm parameters which
are largely related to the accuracy and stability of
the algorithm. All results are compared with the
existing BFA [1], and show the effectiveness of
DBFA for solving dynamic optimization problems.
It is worth mentioning that the diversity of DBFA
changes after each chemotactic process rather than
the dispersion adopted by the BFA after several
generations. The DBFA utilizes not only the local
search but also applies a flexible selection scheme
to maintain a suitable diversity during the whole
evolutionary process. It outperforms BFA in almost
all dynamic environments. The results are shown in
[5]. They have further given solution for global
optimization given in [7].
The novel BSA has been proposed for global
optimization. In this algorithm, the adaptive tumble
and run operators have been developed and
incorporated, which are based on the understanding
of the details of bacterial chemotactic process. The
operators involve two parts: the first is concerned
with the selections of tumble and run actions, based
on their probabilities which are updated during the
searching process; the second is related to the
length of run steps, which is made adaptive and
independent of the knowledge of optimization
problems. These two parts are utilized to balance
the global and local searching capabilities of BSA.
Beyond the tumble and run operators, attraction
and mutation operations have also been developed.
A.ABRAHAM, A. BISWAS, S. DASGUPTA AND S.
DAS have shown [8] that the major driving forces
of Bacterial Foraging Optimization Algorithm
(BFOA) is the reproduction phenomenon of virtual
bacteria each of which models one trial solution of
the optimization problem.
BFO and PSO have been used in combination and
their combined performance has been utilised to
incorporate the merits [9] of two bio-inspired
algorithms to improve the convergence for high-
dimensional function optimization. It is assumed
COPO1O2-3
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
that the bacteria have the similar ability like birds
to follow the best bacterium (bacterium with the
best position in the previous chemotactic process)
in the optimization domain. The position of each
bacterium after every move (tumble or run) is
updated according to (3)
θi(j+1,k,l)=θi(j+1,k,l)+Ccc(θb(j,k,l)-θi(j,k,l), (3)
if Ji(j+1,k,l)>Jmin(j,k,l)
where θb(j,k,l) and Jmin(j,k,l) are the position and
fitness value of the best bacterium in the previous
chemotactic process respectively, Ccc is a new
parameter, called attraction factor, to adjust the
bacterial trajectory according to the location of the
best bacterium.
3. PARTICLE SWARM OPTIMIZATION:
Particle swarm optimization is a high-performance
optimizer that is very easy to understand and
implement. It is similar in some ways to genetic
algorithms or evolutionary algorithms, but requires
less computational bookkeeping and generally only
a few lines of code [10]. Particle swarm
optimization originated in studies of synchronous
bird flocking and fish schooling, when the
investigators realized that their simulation
algorithms possessed an optimizing
characteristic[11]-[13]. As the particles traverse the
problem hyperspace, each particle remembers its
own personal best position that it has ever found,
called its local best. Each particle also knows the
best position found by any particle in the swarm,
called the global best. Overshoot and undershoot
combined with stochastic adjustment explore
regions throughout the problem hyperspace,
eventually settling down near a good solution. This
process can be visualized as a dynamical system,
although the behaviour is extraordinarily complex
even when only a single particle is considered with
extremely simplified update rules. This new
optimization technique has much promise, and
electromagnetic researchers are just beginning to
explore its capabilities.
3.1. CLASSICAL ALGORITHM:
The particle swarm concept originated as a
simulation of a simplified social system. The
original intent was to graphically simulate the
graceful but unpredictable choreography of a bird
flock. Initial simulations were modified to
incorporate nearest-neighbour velocity matching,
eliminate ancillary variables, and incorporate
multidimensional search and acceleration by
distance (Kennedy and Eberhart 1995, Eberhart and
Kennedy 1995). At some point in the evolution of
the algorithm, it was realized that the conceptual
model was, in fact, an optimizer. Through a process
of trial and error, a number of parameters
extraneous to optimization were eliminated from
the algorithm, resulting in the very simple original
implementation (Eberhart, Simpson and Dobbins
1996).
PSO emulates the swarm behaviour and the
individuals represent points in the -dimensional
search space. A particle represents a potential
solution. The velocity Vid and position Xid of the dth
dimension of the ith particle are updated as follows
(4) & (5).
Vid ← Vid + C1* rand1id *(pbestid - Xid) + C2*
rand2id *(gbestd - Xid) (4)
Xid ← Xid + Vid (5)
Where Xi = (Xi1, Xi2, … , XiD) is the position of ith
particle Vi = (Vi1, Vi2, … , ViD) represents velocity
of the particle i. pbest =( pbesti1, pbesti2, … ,
pbestiD) is the best previous position yielding the
best fitness value for their ith particle; and gbest =
(gbesti1, gbesti2, … , gbestiD) is the best position
discovered by whole population[14]. C1 and C2 and
are the acceleration constants reflecting the
weighting of stochastic acceleration terms that pull
each particle toward pbest and gbest positions,
respectively. rand1id and rand2id are two random
numbers in the range [0, 1].
3.2. PSEUDOCODE:
1:Generate the initial swarm by randomly
generating the position and velocity for each
particle;
2: Evaluate the fitness of each particle;
3: repeat
4: for each particle i do
5: Update particle i according to (1) and (2);
6: if (xi) <f(xpbesti) then
7: xpbesti:=xi;
8: if f (xi) <f(xgbest) then
9: xgbest := xi
10: end if
11: end if
12: end for
13: until the stop criterion is satisfied
3.3. ADVANCEMENTS IN PSO AND ITS APPLICATION AND
RESEARCH AREAS:
APPSO (Agent based parallel PSO) is based on two
types of agents: one coordination agent and several
swarm agents. The swarm is composed of various
COPO1O2-4
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
sub-swarms, one for each swarm agent. The
coordination agent has administrative and
managing duties. All the calculations are done by
the swarm agents (see Figure 2).
In order to gain benefit from the large knowledge
and insights achieved in the research field of
sequential PSO it is important to modify the
swarm’s behavior as little as possible. The
inevitable changes to the algorithm due to the
parallelization should also lead to positive effects
FIG. 2 STRUTURE OF APPSO NETWORK SYSTEM [14]
with respect to solution quality. This is realized by
using strategically niching where according to the
requirements the sub-swarms either work together
searching for one global optimum or spread over
the search space to find various global/local
optima, e.g. on multimodal functions.
To address the problem of space locus searching, a
slowdown particle swarm optimization (SPSO) is
proposed to improve the convergence performance
of particle swarm from the position viewpoint. The
particle swarm in SPSO is divided into many
independent sub-swarms to guarantee that particles
converge to different position, since space locus
has multiple optimal solutions and requires the
convergence of both fitness and position of particle.
Furthermore, particle velocity is updated by half
according to fitness to achieve the position
convergence [14].
Several PSO algorithms have been recently
proposed to address DOPs of which using multi
swarms seems a good technique. The multi swarm
method can be used to enhance the diversity of the
swarm, with the aim of maintaining multiple
swarms on different peaks.
One of the reasons that particle swarm optimization
is attractive is that there are very few parameters to
adjust. One version, with very slight variations (or
none at all) works well in a wide variety of
applications. Particle swarm optimization has been
used for approaches that can be used across a wide
range of applications, as well as for specific
applications focused on a specific requirement. In
this brief section, we cannot describe all of particle
swarm’s applications, or describe any single
application in detail. Rather, we summarize a small
sample.
Generally speaking, particle swarm optimization,
like the other evolutionary computation algorithms,
can be applied to solve most optimization problems
and problems that can be converted to optimization
problems. Among the application areas with the
most potential are system design, multi-objective
optimization, classification, pattern recognition,
biological system modelling, scheduling
(planning), signal processing, games, robotic
applications, decision making, simulation and
identification. Examples include fuzzy controller
design, job shop scheduling, real time robot path
planning, image segmentation, EEG signal
simulation, speaker verification, time-frequency
analysis, modelling of the spread of antibiotic
resistance, burn diagnosing, gesture recognition
and automatic target detection, to name a few.
4. FIREFLY OPTIMIZATION
Fireflies, also called lighting bugs, are one of the
most special and fascinating creatures in nature.
These nocturnal luminous insects of the beetle
family Lampyridae (order Coleoptera), inhabit
mainly tropical and temperate regions, and their
population is estimated at around 1900 species.
They are capable of producing light thanks to
special photogenic organs situated very close to the
body surface behind a window of translucent
cuticle. Bioluminescent signals are known to serve
as elements of courtship rituals, methods of prey
attraction, social orientation or as a warning signal
to predators (in case of immature firefly forms
commonly referred to as glow worms). The
phenomenon of firefly glowing is an area of
continuous research considering both its
biochemical and social aspects.
Mechanisms of firefly communication via
luminescent flashes and their synchronization has
been imitated effectively in various techniques of
wireless networks design, dynamic market pricing
and mobile robotics.
4.1. CLASSICAL ALGORITHM
The flashing light of fireflies is an amazing sight in
summer sky in the tropical and temperate regions.
There are about two thousand firefly species, and
most fireflies produce short and rhythmic flashes.
The pattern of flashes is often unique for a
particular species. The flashing light is produced by
COPO1O2-5
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
a process called bioluminescence. However, two
fundamental functions of such flashes are to attract
mating partners and to attract potential prey. In
addition it may also serve as a protective warning
mechanism to remind potential predators of the
bitter taste of fireflies.
The rhythmic flash, the rate of flashing and the
amount of time form the part of the signal system
that bring both sexes together. We know that the
light intensity at a particular distance r from the
light source obeys the inverse square law. That is to
say, the light intensity I decreases as the distance r
increases in terms if I . Furthermore, the air
absorbs light which becomes weaker as the
distance increases. These two combined factors
make most fireflies visual to a limit distance,
usually several hundred meters at night, which is
good enough for fireflies to communicate.
The flashing light can be formulated in such a way
that it is associated with the objective function to
be optimized, which makes it possible to formulate
new optimization algorithms.
Assumptions:
1. All fireflies are unisex so that one firefly will be
attracted to the other fireflies regardless of their
sex.
2. The brighter one will move towards the less
bright one. The attractiveness is proportional to the
brightness and they both decrease as the distance
increase. And if there is no brighter one than a
particular firefly, it will move randomly;
3. The brightness of a firefly is affected or
determined by the landscape of the objective
function.
4.1.1 LIGHT INTENSITY AND THE ATTRACTIVENESS:
In the firefly algorithm, there are two important
issues: the variation of light intensity and
formulation of the attractiveness. For simplicity, we
can always assume that the attractiveness of a
firefly is determined by its brightness which in turn
is associated with the encoded objective function.
In the simplest case particular location x can be
chosen as I(x) . However attractiveness β is
relative.
In simplest form, the intensity I(r) varies according
to inverse square law
I(r) =
Where Is is the intensity at the source. For a given
medium with light absorption coefficient , the light
intensity I varies with the distance r. That is
,
where I0 is the original light intensity. In order to
avoid the singularity at r=0 in the expression Is/r2,
combined effect of both the inverse square law and
the absorption can be approximated as the
following Gaussian form.
I(r) =
Thus of the firefly can be defined as
Where β0 is attractiveness at r=0
The movement of a firefly i is attracted to another
more attractive firefly j is determined by
= + (xi - xj) +α ,
where second term is due to the attraction. The
third term is randomization with α being the
randomization parameter, and is a vector of
random numbers drawn from a Gaussian
distribution or uniform distribution. For eg. The
simplest form is and can be replaced by
rand(1/2).
The algorithm presented here makes use of a
synergic local search. Each member of the swarm
explores the problem space taking into account
results obtained by others, still applying its own
randomized moves as well. The influence of other
solutions is controlled by value of attractiveness. It
can be adjusted by modifying two parameters: its
maximum value β0 and an absorption coefficient γ.
The first parameter describes attractiveness at rj = 0
i.e. when two fireflies are found at the same point
of search space S. In general β0 [0; 1] should be
used.
SPECIAL CASES:
For γ→0, the attractiveness is constant β =β0, this is
equivalent to say that the light intensity does not
decrease in an idealized sky.
When γ→∞ leads to β(r)→δ(r), which means that
the attractiveness is almost zero in the sight of
other fireflies.
.4. 2. PSEUDOCODE:
1. Input:
f(z), z = [z1; z2; :::; zn]T
m, β0, γ, min ui, max ui (parameters)
2. Output:
Ximin
COPO1O2-6
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
Begin
for i=1 to m do
xi← Generate Initial Solution ()
end
3. repeat
imin← arg mini f(xi)
ximin← arg minxi f(xi)
4. for i=1 to m do
for j=1 to m do
if f(xj) < f(xi) then
rj← Calculate Distance (xi , xj)
β←β0e-γrj
ui← Generate Random Vector (min ui, max ui)
for k=1 to n do
xi,k ←(1- β)xi,k + βxi,k + ui,k
end
end
end
end
5. uimin← Generate Random Vector (min ui, max ui)
for k=1 to n do
ximin,k← ximin,k + uimin,k
end
until stop is condition true
6.end
4.3. ADVANCEMENTS IN FFA AND ITS APPLICATION AND
RESEARCH AREAS:
Many bio-inspired algorithm exist that use simple
rules and local information to create a global
consensus of a single parameter, such as clock time
in the flashing firefly algorithm, or location in the
slime algorithm. In this paper we identify a control
loop as the core structure of several bio-inspired
algorithms. Two ways of combining these loops are
identified, serially, and nested. These methods of
combination can be used to combine algorithms to
allow the control of multiple parameters (one per
algorithm) on a global level without the use of a
centralised point of control.
Thresholding is one of the most important
techniques for performing image segmentation. It is
generally simple and computationally efficient. The
MEFFT algorithm simulates the behaviour of
fireflies and the phenomenon of bioluminescent
communication to develop the algorithm to select
the adequate thresholds for image segmentation.
The segmentation results of MEFFT algorithm are
promising and it encourages further researches for
applying this algorithm to complex and real-time
image analysis problems such as target recognition,
complex document analysis and biomedical image
application [16]. The other application include
synchronisation of wireless network which be seen
in [17]. An application in pulse-coupled oscillator’s
model addresses the issue of synchronization of the
oscillators with different frequencies. As a
continuation of this research, it is important to
study how the topology and communication delay
of the distributed oscillators system affect the
synchronization process.
5. CONCLUSION
While studying and working on these algorithms
we have seen that performance of any technique is
objective dependent. There are various parameters
that define the aptness of the algorithm. So it
depends on the requirements of the objective
function and on the constraints imposed. For
example if we take several benchmark problems
and compare the results of three algorithms then it
has been seen each would perform differently. Not
only their individual performance vary even they
tend to become less or more stable in terms of the
parameters defining the quality of performance
such as elapsed time, convergence rate, optimum
value (maximised/minimized), mean and standard
deviation etc. So depending upon our requirement
we may define our objective function and the
performance of algorithm for that particular
function is based on how aptly the required
objective is received, which could be in terms of
final value, elapsed time, mean etc.
Thus having witnessed these variations one can
generalise the performance on overall terms and
state any of the algorithm as best among the three
taken.
6. REFERENCES
[1] K. M. PASSINO, “Biomimicry of bacterial foraging for
distributed optimization and control,” IEEE Control Syst., vol.
22, no. 3, pp. 52–67, Jun. 2002.
[2]. S. DAS, S. DASGUPTA, A. BISWAS and A. ABRAHAM,
“On Stability of the Chemotactic Dynamics in Bacterial-
Foraging Optimization Algorithm.” IEEE Trans. on Systems,
Man, and cybernetics, part: A: Systems and Humans, vol. 39,
no. 3, pp. 670-679, May 2009.
[3]. B. D. Hughes, Random Walks and Random Environments.
London, U.K.: Oxford Univ. Press, 1996.
[4]. L. DOS SANTOS COELHO, C. DA COSTA SILVEIRA,
C.A. SIERAKOWSKI and P. ALOTTO, “Improved Bacterial
Foraging Strategy Applied to TEAM Workshop Benchmark
Problem.” IEEE Trans. On Magnetics, Vol. 46, no. 8, pp. 2903-
2906, Aug. 2008.
[5] W. J. TANG and Q. H. WU, “Bacterial Foraging Algorithm
For Dynamic Environments.” Proc. of IEEE Congress on
Evolutionary Computation Sheraton Vancouver Wall Centre
Hotel, Vancouver, BC, Canada, pp. 1324-1330, July 16-21,
2006.
[6] R. W. MORRISON and K. A. DE JONG, “A test problem
generator for non stationary environments”, Proc. of IEEE
Congress on Evolutionary Computation, IEEE Press, pp. 2047–
2053, 1999.
[7]. W. J. TAN and Q. H. WU, “A Bacterial Swarming
Algorithm For Global Optimization.” Proc. of IEEE Congress
on Evolutionary Computation, pp. 1207-1212, 2007.
COPO1O2-7
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
[8]. A. ABRAHAM, A. BISWAS, S. DASGUPTA AND
SWAGATAM DAS, “Analysis of Reproduction Operator in
Bacterial Foraging Optimization Algorithm.” Proc. of IEEE
Congress on Evolutionary Computation, pp. 1476-1483, 2008.
[9]. YING CHU, HUA MI, HUILIAN LIAO, ZHEN JI, AND
Q. H. WU, “A Fast Bacterial Swarming Algorithm For High-
dimensional Function Optimization.” Proc of IEEE Congress on
evolutionary computation, pp. 3135-3140, 2008.
[10]. DANIEL W. BOERINGER AND DOUGLAS H.
WERNER, “Particle Swarm Optimization Versus Genetic
Algorithms for Phased Array Synthesis.” Trans. Of IEEE on
Antennas and Propagation, vol. 52, no. 3, pp. 771-779,March
2004.
[11] R. EBERHART AND J. KENNEDY, “A new optimizer
using particle swarm theory,” in Proc. 6th Int. Symp. Micro
Machine and Human Science (MHS ’95), pp. 39–43, 1995.
[12] J. KENNEDY AND R. EBERHART, “Particle swarm
optimization,” in Proc. IEEE Int. Conf. Neural Networks, vol. 4,
pp. 1942–1948, 1995.
[13]. J. J. LIANG AND A. K. QIN, “Comprehensive Learning
Particle Swarm Optimizer for Global Optimization of
Multimodal Functions.” Trans of IEEE on Evolutionary
Computation, vol. 10, no. 3, June 2006.
[14]. LI ZHI-JIE, LIU XIANG-DONG, DUAN XIAO-DONG,
WANG CUN-RUI, “An Improved Particle Swarm Algorithm
for Search Optimization.” Proc. of IEEE Global Congress on
Intelligent System, pp.154-158, 2009.
[15]. MICHAEL BREZA, JULIE MCCANN, “Can Fireflies
Gossip and Flock?: The possibility of combining well-knowbio-
inspired algorithms to manage multiple global parameters in
wireless sensor networks without centralised control.
[16]. MING-HUWI HORNG AND TING-WEI JIANG,
“Multilevel Image Thresholding Selection based on the Firefly
Algorithm.” Symposia and Workshops on Ubiquitous,
Autonomic and Trusted Computing, pp. 58-63, 2010.
[17]. LIN CUI, HONGPENG WANG, “Reachback Firefly
Synchronicity with Late Sensitivity Window in Wireless Sensor
Networks.” Proc. of IEEE Ninth International Conference on
Hybrid Intelligent Systems, pp. 451-456, 2009.
COPO1O2-8
 BIOMETRIC AUTHENTICATION USING IMAGE PROCESSING TOOLS
Neha Mittal Ruchi Gupta
Department of Electronics and communication
J.P Institute of Engineering & Technology, Meerut
Email: nehamittal2008@gmail.com
Abstract
This paper presents the study of biometric
authentication using various digital image
processing techniques and some future
possibilities in this field. The position of
biometrics in the current field of security has
been depicted in this work. For biometrics
recognition the Region of interest (ROI) is
extracted from the biometric modalities.
Various different features are extracted from
ROI. These features can be fuzzy features,
wavelet features, Energy and entropy
features. These sets of features are used for
the authentication of users using Euclidean
Distance, Chi square measure and fuzzy
Support Vector Machines as classifiers.
Recognition score of 82% is obtained with
wavelet feature on palmprint database of
PolyU. We have also outlined the
comparative study and usability of biometric
authentication systems in this paper.
Keywords: wavelet features, multimodal
biometric, ROI, Euclidean Distance, chi-
square,
1. INTRODUCTION
Computer aided personal authentication is
becoming more and more reliable in this
information era. As in order to deal with
security, Authentication plays an important
role. Biometrics is one of the most reliable in
this field. A wide variety of systems require
reliable personal recognition schemes to
either confirm or determine the identity of an
individual requesting their services. The
purpose of such schemes is to ensure that the
rendered services are accessed only by a
genuine user, and not anyone else. By using
Pratibha Singh
biometrics it is possible to confirm or
establish an individual’s identity.
The most widely used biometric feature is the
fingerprint while most reliable one is the
iris[1].Palmprint, as a biometric feature
gaining acceptance in the field of biometric
recognition compared with other modalities:
low cost capture device, easy self positioning,
cannot be easily stolen, large area to get large
information.
In this paper, an authentication of biometrics
is studied. Selection of region of interest,
feature extraction, and classification of these
features is explored.
The organization of the paper is as follows:
Section 2 presents the extraction of ROI. The
extraction of features is described in Section
3. Results are described in Section 4 and
finally multimodal biometrics systems and
conclusions are drawn in Section 5 and 6.
2. ROI EXTRACTION
The useful information content of an image of
any biometric modality can be abstracted by
different image parameters; this is nothing but a
region of interest. ROI is simply a subimage of
interest. An algorithm for abstracting ROI of
Palmprint is given in [8].
COPO1O1-1
 BIOMETRIC AUTHENTICATION USING IMAGE PROCESSING TOOLS
Fig. 1: A typical Palmprint Fig. 2: The extracted ROI of
palmprint
3. FEATURE EXTRACTION
The issue of choosing the features to be
extracted should be guided by some concerns
like the features should carry enough
information about the image and should not
require any domain-specific knowledge for
their extraction. They should be easy to
compute in order for the approach to be
feasible for a large image collection and rapid
retrieval.
Finger print matching techniques can be
placed into two categories. One of them is
Minutiae based and the other one is
Correlation based. Minutiae based techniques
find the minutiae points first and then map
their relation placement on the finger.
Correlation based techniques require the
precise location of a registration point and are
affected by image translation and rotation [6,
11,12].
The iris of the eye is colored area surrounded
by the pupil. Iris patterns are unique and are
obtained through video based image
acquisition system. Each iris structure is
featuring a complex pattern. This can be a
combination of specific characteristics known
as corona, crypts, filaments, freckles, pits,
furrows, striations and rings [4].
The palmprints possess random textures. A
palmprint is a rich source of texture. It
consists of zigzag and crisscrossing lines in
addition to three principal lines. Palmprint
also contains creases, wrinkles and ridges.
Texture features can be extracted using
statistical measures (entropy, homogeneity),
wavelet, and fractals.
In [8] the ROI is divided in to non-
overlapping windows. First the membership
function is calculated and from this
membership function different features (i.e.
energy feature, fuzzy feature, Entropy feature)
can be calculated. Some features are as
follows:
3.1 Wavelet Features
Wavelets are immensely suitable for texture
analysis. In [1] and [8] Haar wavelets are
effective on the palmprints. Simplicity of
these wavelets is an overriding factor in the
choice of these wavelets though other
celebrated wavelets such as Daubechies
wavelets, Symlets are also tried but the
improvement accrued from the latter ones is
not proportionate with the effort.
A two-dimensional discrete wavelet transform
is applied on the ROI of a palmprint resulting
in four subbands: HL1, LH1 and HH1, which
correspond to fine-level coefficients and LL1,
which corresponds to coarse-level
coefficients. The HL1, LH1 and HH1together
contribute to a detail image. Each detail
image represents a high frequency structure,
i.e. lines, creases, wrinkles, ridges on a
palmprint. Fig. 3 shows one level of
decomposition. LL1 can be further
decomposed into four sub bands and the
process is repeated till the information from
LLn at the nth decomposition is insignificant.
In any decomposition, the detail images
include horizontal, vertical and diagonal parts
COPO1O1-2
 BIOMETRIC AUTHENTICATION USING IMAGE PROCESSING TOOLS
of the decomposed image. These three detail
images are superimposed in [8] to yield
composite mage. It may be noted that as the
decomposition level increases the size of the
detail images also decreases.
The results of authentication correspond to
the composite image.
From each detail images, energy feature is
calculated as [1]:
M N
(1)
Eid ( Si ( x, y )) 2 , i 1, 2,..., 5
x 1 y 1
Fig. 3: Two dimensional one level DWT
decomposition
Fig. 4: Three-level DWT decomposition
where i is the decomposition level. Hi,Vi and
Di are the detail images in horizontal, vertical
and diagonal directions respectively. Other
features are under investigation.In [8] all the
detail images are super imposed and then
energy is calculated.
4. Results and implementation
In [8] a 100% recognition score is obtained
using fuzzy feature. A simple Euclidean
distance measure is used to find out the
recognition rate. In [8] the database used, is
created in the biometrics lab of IIT Delhi.
ROI is divided in to non-overlapping
windows. Features are calculated from these
windows. The size of the window is varied
and the recognition score is obtained.
Given two data sets of features corresponding
to the training and testing samples, a
matching algorithm determines the degree of
similarity between them. A Euclidean
distance is adopted as a measure of
dissimilarity for the palmprint matching using
both wavelet and fuzzy features.
The wavelet feature in [8] is applied on polyU
database [15] on 50 users and 5 images of
each user (total 250 images).4 images are
taken for training data and 1 image for testing
data. A recognition score of 82% is obtained.
Euclidean distance measure is taken as a
classifier.ROC plot is shown in fig.5
COPO1O1-3
 BIOMETRIC AUTHENTICATION USING IMAGE PROCESSING TOOLS
Fig 5: ROC plot for PolyU database
5. MULTIMODAL BIOMETRIC
SYSTEM
Multimodal biometric systems are capable of
utilizing more than one physiological or
behavioral characteristic for authentication. In
order for the biometrics to be ultra-secure and
to provide more-than-average accuracy, more
than one form of biometric identification is
required. Hence the need arises for the use of
multimodal biometrics. This uses a
combination of different biometric
recognition technologies.
In certain situations, the user might find one
form of biometric identification is not exact
enough for identification. This can be the case
with fingerprints, where at least 10% of the
population have worn, cut or unrecognizable
prints.
Multimodal biometric technology uses more
than one biometric identifier to compare the
identity of the person. Therefore in the case of
a system using say three technologies i.e. face
mimic and voice. If one of the technologies is
unable to identify, the system can still use the
other two to accurately identify against.
By using more than one means of biometric
identification, the multimodal biometric
identifier can retain high threshold
recognition settings. The system administrator
can then decide the level of security he/she
requires. For a high security site, they might
require all three biometric identifiers to
recognize the person or for a lower security
site, only one or two of the three. With this
methodology, the probability of accepting an
impostor is greatly reduced.
The unimodal biometric systems can have the
variety of limitations which are as follows
[13]:
(a) Noise in sensed data: A fingerprint
image with a scar or a voice sample
altered by cold, are examples of noisy
data. Noisy data could also result from
defective or improperly maintained
sensors (e.g., accumulation of dirt on a
fingerprint sensor) or unfavorable ambient
conditions (e.g., poor illumination of a
user’s face in a face recognition system).
(b) Intra-class variations: These variations
are typically caused by a user who is
incorrectly interacting with the sensor
(e.g., incorrect facial pose), or when the
characteristics of a sensor are modified
during authentication (e.g., optical versus
solid-state fingerprint sensors).
(c) Inter-class similarities: In a biometric
system comprising of a large number of
users, there may be inter-class similarities
(overlap) in the feature space of multiple
users.
(d) Non-universality: The biometric system
may not be able to acquire meaningful
biometric data from a subset of users. A
fingerprint biometric system, for example,
may extract incorrect minutiae features
COPO1O1-4
 BIOMETRIC AUTHENTICATION USING IMAGE PROCESSING TOOLS

from the fingerprints of certain Multimodal biometric system also prevent

individuals, due to the poor quality of the from spoofing as it is difficult for an impostor
ridges. to spoof multiple biometric modalities of a
(e) Spoof attacks: This type of attack is genuine user simultaneously.
especially relevant when behavioral traits
such as signature or voice are used. References
However, physical traits such as
fingerprints are also susceptible to spoof [1] Xiang-Qian Wu, Kuan-Quan Wang, David
attacks. These systems are able to meet Zhang, ”Wavelet Based Palmprint
the rigorous performance requirements Recognition”, First international
imposed by various applications. They conference on Machine learning
address the problem of non-universality, andcybernetics, Beijing, pp-1253-1257, 4-
since multiple traits ensure sufficient 5 Nov 2002.
population coverage. They also deter [2] R. C. Gonzalez, R. E. Woods, Digital
spoofing since it would be difficult for an Image Processing, Addison Wesley
impostor to spoof multiple biometric traits publishers, 1993.
of a genuine user simultaneously. [3] L. Hong and A. K. Jain, “Integrating Faces
and Fingerprints for Personal
Identification”, IEEE Trans. on Pattern
Analysis and Machine
Intelligence, Vol. 20, No. 12, pp. 1295-
1307, Dec1998.
[4] Sanjay R. Ganorkar, Ashok A. Ghatol,
“Iris Recognition: An Emerging Biometric
Technology”, In Proc. of the 6th WSEAS
International Conference on Signal
Fig 6: A multimodal biometric system Processing, Robotics and Automation,
Greece, Feb. 2007, pp. 91 – 96.
Limitations of unimodal biometrics are [5] N. Duta, A.K. Jain, and K.V. Mardia,
overcome by the multimodal biometric “Matching of Palmprint,” Pattern
system. Recognition Letters, Vol.23, No. 4, pp.
477–485, 2001.
6. CONCLUSIONS [6] David Zhang, Wai-Kin Kong, Jane You,
and Michael Wong, “Online Palmprint
Biometric recognition can offer a high degree Identification”, IEEE Trans. Pattern
of security, they are far from perfect solution. Analysis &Machine Intelligence, Vol. 25,
As the fingerprint can give the high accuracy No. 9, pp. 1041 – 1050,September 2003.
of recognition but it have draw backs as [7] Ying-Han Pang, Andrew Teoh Beng Jin,
discussed. So one solution to this problem is David Ngo Chek Ling, "Palmprint based
to integrate two or more modalities, to Cancelable Biometric Authentication
provide high security.
COPO1O1-5
 BIOMETRIC AUTHENTICATION USING IMAGE PROCESSING TOOLS

System", Int. J. of Signal Processing, http://www4.comp.polyu.edu.hk/~biometri

Vol.1, No.2, pp.93-99, 2004. cs/.
[8] M. Hanmandlu, H M Gupta, Neha Mittal,
S. Vasikarla,”An authentication system
based on palmprint”,ITNG-2009,pp-399-
404
[9] Lei Zhang, ZhenhuaGuo, Zhou Wang,
DavidZhang, "Palmprint Verification
Using Complex Wavelet Transform", IEEE
Intl. Conf. on ImageProcessing, San
Antonio, TX, Sept.16-19,2007.
[10] Li Fang, Maylor K.H. Leung , Tejas
Shikhare, Victor Chan, Kean Fatt Choon ,
“Palmprint Classification”, IEEE Int. Conf.
on Systems, Manand Cybernetics, Vol. 4,
8-11 October 2006, pp.2965-2969.
[11] L. Hong and A. K. Jain, “Integrating
faces and fingerprints for personal
identification”, IEEE Trans. Pattern
Anal. Mach. Intell., Volume 20, No. 12,
Dec. 1998, pp. 1295–1307.
[12] R. Cappelli, D. Maio, D. Maltoni, J. L.
Wayman, and A. K. Jain, “Performance
evaluation of fingerprint verification
systems”, IEEE Trans. Pattern Anal. Mach.
Intell., Volume 28, issue 1, Jan. 2006, pp.
3–18.
[13] M. Indovina, U. Uludag, R. Snelick,
A. Mink, and A. K. Jain, “Multimodal
biometric authentication methods: A
COTS approach,” in Proc. of Workshop on
Multimodal User Authentication, (Santa
Barbara, CA), pp. 99–106, Dec 2003.
[14] Wen-Shiung Chen, Yao-Shan Chiang
and Yen-Hsun Chiu, ”Biometric
Verification by Fusing Hand Geometry and
Palmprint”, IIH-MSP 2007
[15] H. K. Polytechnic University,
“Palmprint database”, Biometric Research
Center Website. 2005.

COPO1O1-6
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Medical Image Enhancement In Spatial Domain: Statistical Analysis

Sanjeev Kumar Gupta1, Sonika Singh2& Rahul Singh3
1
Assistant Professor , 2,3 B.Tech Scholar, ECE Department, Faculty of Engineering and
Technology, R.B.S. College, Bichpuri Campus, Agra

ABSTRACT

Image quality and utility become crucial issues for engineers, scientist, doctors, patients,
insurance companies and lawyer whenever there are changes in the technology by which
medical images are achieved. Examples of such changes include analog to digital conversion,
lossy compression for transmission and storage, image enhancement, and computer aided
methodology for diagnosis of disease in medical images. Edit an image so that it is more suitable
for a specific application than the original image is termed as image enhancement technique.
Image is defined as a two dimensional function f(x, y), where x and y are spatial coordinates,
that bears information, which can be generated in any form such as visual, x-ray and so on. X-
rays are the oldest source of electromagnetic radiation used for medical imaging. Medical image
enhancement methods are used like all other methods and algorithm in image processing, as
chain subsequent edits aimed at achieving a suitable result. Improving one function in the chain
is only useful if the end result is really unproved, and that does not solely depend on that
particular function; it also depends on the quality of the first image. In this paper we have
compared different types of image enhancement technique of medical image in spatial domain
and also presented a statistical analysis.

Medical Image Enhancement domain methods, which operate directly on

Digital image processing is a technique
using which the data from an image is
digitized. With the help of computer
algorithms, various mathematical operations
are applied to the data to create an enhanced
image. In medical image enhancement we
use the input, which is the form of X-ray
film. In the input image we perform some
mathematical operation and the gray values
of the input image are mapped to another set
of gray values, so that the image is made
suitable for specific application. it is called
Image Enhancement method.
Image enhancement technique can be
divided in two broad categories spatial
pixel and frequency domain methods, which
operate directly on the Fourier, transform of
an image. In this present paper we focus our
attention on spatial domain. There are
several enhancements in spatial domain such
as-
i. Gray Scale manipulation.
ii. Histogram equalization.
iii. Image smoothing-Neighborhood
averaging and Edge preserving
smoothing.
The term spatial domain refers to the
aggregate of pixels composing an image.
These methods are procedures that operate

COPO1O5-1
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
directly on these pixels. Spatial domain
process will be denoted by the expression
g(x>y) =T[f(x, y)] (1)
Where f(x, y) is the input image, g(x, y) is
the processed image, and T is an operator on
f, defined over some neighborhood about a
(x, y). The principle approach in defining a
neighborhood about a point (x, y) is to use a
square or rectangular sub image area
centered at (x, y), as shown in fig.1.
Fig.1.
The center of the sub image is moved from
pixel to pixel starting from the top left
corner to the bottom most right corner. The
operator T is applied at each location (x, y)
to yield the output at that location. The
process utilizes only the pixel hi the area of
the image spanned by the neighborhood.
Although other neighborhood shapes, such
as approximation to a circle, square and
rectangular arrays are by far the most
predominant because of their ease of
implementation. The simplest form of T is a
single pixel when the neighborhood is of
size 1X1.In this case, g depends only on the
value of fat(x, y), and t becomes a gray-level
(also called intensity or mapping)
transformation function of the form:
s=T(r) (2)
Where, r and s are variables denoting, the
gray level of f(x, y) and g(x, y) at any
point(x, y) respectively.
The effect of thresholding would be to
produce an image of higher contrast than
original by darkening the levels below m
(threshold) and brightening the level above
in the original image, this technique is
known as contrast stretching.
The gray scale modification method
includes Log transformation, Power law
transformation and Piecewise transformation
and Histogram equalization method include
local histogram equalization and global
histogram equalization. The third method
and many more methods are possible but we
have not discussed in this paper. Gray-level
transformation is very simple and powerful
processing method of radiographic image
enhancement. The log transformation is used
to compress dynamic range. In this
transformation function large numbers of
gray level values are mapped in to narrow
range. The general form of log
transformation is:
S=c log(1+r) (3)
Where c is a constant and it is assumed that
r>=o
Power-law transformation is used to expand
a certain range of intensity values. In these
methods the exponent of the power-law
equation is called gamma f. By varying
gamma, a family of power-law
transformation can be obtained. Power law
transformations have the basic form:
S=crγ (4)
Where c and γ are positive constants.
The simplest form of Piecewise linear
function is a contrast stretching
COPO1O5-2
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
transformation. Low-contrast image can
result from poor illumination, lack of
dynamic range in the imaging sensor, or
even wrong setting of lens aperture during
image acquisition. The idea behind contrast
stretching is to increase the dynamic range
of the gray levels in the image being
processed.
Global histogram equalization method is a
simple and computationally inexpensive
implementation, but it treats the entire
content image uniformly, regardless of the
local variation of gray levels. This may lead
to loss of visual quality of images for certain
areas and increased visibility of noise.
Discrete domain
Sk= T(rk)= = (5)
Analog Domain
s= T(r) = (6)
Local histogram equalization enhances the
small region in the image features. A square
or rectangular neighborhood is defined and
the centre of this area from pixel to pixel is
moved. At each location, the histogram of
die points in the neighborhood is computed
and either a histogram equalization or
histogram specification transformation
function is obtained. Fig.3 shows various
outputs.
Medical X-ray images have narrow latitude
or dynamic range. The film’s ability to
record and view all the anatomy within the
x-ray field is limited by this narrow dynamic
range. Image enhancement is the process of
applying these techniques to facilitate the
development of a solution to a computer
imaging problem. Consequently, the
enhancement methods usually are
application specific and are often developed
empirically. Actuality of image
enhancement problem conditioned by
necessity of it’s solving in medical
diagnostics, pattern recognition,
enhancement amateur photography and etc.
Since radiation oncologists believe that final
images have a higher contrast, and from the
statistical assessment it can be concluded
that an increase of average numbers and
deviation of images value will increase the
contrast. As a result processed images via
this program will lead to have a higher
diagnostic ability.
Result and discussion
Here we have presented a statistical analysis
in table form. The variation of the mean
value and standard deviation value of input
radiographic images and enhanced
radiographic images are calculated in this
present paper. We have worked on many
radiographic images but only three images
chest, hand and leg are shown in this paper.
In global histogram equalization method
enhances the whole area of image features.
Therefore for achieving uniform distribution
of gray scale values for histogram
equalization the intensity are increased. This
raises the mean value of enhanced image.
With the increase in mean value, the average
distribution of intensity values also
increases. Consequently the standard
deviation increases.
In local histogram equalization method,
main focused area is selected regions
determined by window. The entire image
need not be enhanced. So, mean values and
standard deviation values of the enhanced
COPO1O5-3
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
images are not affected as much as that of
global histogram method. If window size
increases then proportionately larger region
is chosen out of entire image region causing
mean values to decrease. If value of constant
k changes or increases, more uniform
distribution is seen in the histogram it
indicates that more equalization is achieved.
It is observed that in log transformation
method for different value of c, if gray value
increases then statistical values (mean,
standard deviation) of enhanced image also
increases which in turn increases the image
brightness. Similarly in power law
transformation using fixed values of gamma
and different value of c, if the gray value is
increased, the statistical values are also
increased proportionately. In power law
transformation with gamma correction
approach, if the value of gamma increases,
the value of mean of enhanced image
decreases while the value of standard
deviation increases, as result the brightness
of enhanced image is decreased.
Similar observation have been found in
piecewise contrast stretching method i.e.
enhanced image is dark for gamma=0.1 and
contrast increase as the value of gamma
increases from gamma-0.2 onwards.
Table 1, 2, 3 shows the comparison of
different method of Radiographic Image
Enhancement.
Because of flexibility of changing window
size and equalization factor k an auxiliary
region can be enhanced to the extent
required in the local histogram equalization
method. Therefore this method is also
sometimes referred to as adaptive method.
In the present enhancement approaches,
power law transformation using gamma
correction performs much better as
compared to all other methods addressed.
This conclusion is on the basis of the change
in standard deviation and mean value on the
histogram of the enhanced image.
Figure 4, 5, 6 shows plot of mean value and
standard deviation value of four methods of
Leg1, Hand1, and chest 1 image. In bar
graph the global histogram method is not
shown.
Conclusion
In this present paper we have compared
different techniques of image enhancement
in spatial domain. They have been evaluated
in terms of standard deviation, mean values
as statistical measures. As medical image
has only a finite number of gray scales, an
ideal equalization is not possible. Image
enhancement technique can be improved if
the enhancement criteria are application
dependant and can be illustrated precisely.
References
Millener AE, AubinM, Pouliot J,
ShinoharaK, Roach HIM (2004)
“Daily electronic portal imaging
from morbidly obes men undergoing
radiotherapy for localized prostate
cancer”. Int J Radiat Oncol
Bioiphys, 6.10.
Van Denberg DL, De Ridder M,
Storm GA (2004) “Imaging in
radiotherapy” European Journal of
radiology, 41-48.
Shi X.Q., Sallstrom P, and Welander
U. “A Color Coding Method for
Radiographic Image and vision
COPO1O5-4
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

computing” Vol.20, pp:761- Advances in Image Guided

767.20Q2. Radiotherapy”. Euspico 2007,
Pznon, Poland, Sep2007.
Gonzalez RC and Woods R.E.
“Digital Image Processing”, S.S.Y. Lau, „Global image
Pearson Education Singapore, enhancement using local
Fourth Indian Reprint, 2nd edition information”. Electronics letters. Pp.
2003. 122-123, January 1994.

Jain A. “ fundamentals of Digital Jha Vandana, Thakur Kavita and

Image Sinha G.R. “Image Enhancement
Processing”, Englewood Cli.s Techniques in spatial domain”. A
Printice Hall, 1989. case study, J. of Acta Ciecia Indica,
Vol.XXXHI p, No.2, pp:201, 2007.
Pratt W. “Digital Image
Processing”, New York, Wiley JohnB. Zimmerman. et al., “An
Interscience, 1991. evaluation of efectiveness of adaptive
histogram equalization for cantrast
Andrews H.C., Monochrome enhancement”. IEEE Trans. On
“Digital Image Enhancement”, Medical Imaging, pp.304-312.
Appl., p.495, 1976. Deember 1988.
Gonzalez RC and Woods R.E., and www.wikipedia.com.
Eddins SL. “Digital Image
Processing using MATLAB”. http://www.freepatentsonline.com/E
Pearson Edition. P 1397780.html.

R.S.Gaborskl Kodak Health Imaging

Syst. Inc. Rochester, NY, USA,
…BK_ Jang, Kodak Health Imaging
Syst. Inc. Rochester, NY, USA,
“enhanceme… for Computed
Radiographic Images”.

Joung-youn Kirn; Lee-Sup Kim;

Seung-Ho Hwang. “An advanced
contrast enhancement using partially
overlapped subblock histogram
equalization” Circuits and
Systems.2000. Proceedings. ISCAS
2000. Geneva.

Matuszewski Bogdan J.. Moore

Christopher J.. “Deformable Models
in Medical Image Processing,
COPO1O5-5
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Figure 3: Original Image and Enhanced

images using different Methods

Log Transformed

Original Image

Power Law Transform γ=1

Global Histogram Equalized

Power Law Transform γ=0.9

Adaptive Histogram Equalized

Power Law Transform γ=0.8

COPO1O5-6
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Power Law Transform γ=0.7

Figure-4: Mean values and Standard

Deviation of Medical Image Enhancement
for Leg

Figure-6: Mean values and Standard

Deviation of Medical Image Enhancement
for Chest

Figure-5: Mean values and Standard

Deviation of Medical Image Enhancement
for Hand

COPO1O5-7
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Table 1: Comparison of different method of medical image enhancement for chest

Chest

Method Input Image Enhanced Image

Mean Std Mean Std

1.Global Histogram 116.0488 93.1951 127.3704 74.9504

Equalization

2.Adaptive histogram 116.0488 93.1951 114.8210 84.0122

equalization

3.Log transformation 116.0488 93.1951 180.8969 77.4127

4.Power law 116.0488 93.1951 116.0488 93.1951

transformation γ=1

5. γ =0.9 116.0488 93.1951 120.6711 93.2783

6. γ =0.8 116.0488 93.1951 125.9139 93.0551

7. γ =0.7 116.0488 93.1951 132.0142 92.2453

Table 2: Comparison of different method of medical image enhancement for Hand

Hand

Method Input Image Enhanced Image

Mean Std Mean Std

1.Global Histogram 68.9736 80.6995 127.6161 74.2843

Equalization

2.Adaptive histogram 68.9736 80.6995 72.7664 72.7077

equalization

COPO1O5-8
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

3.Log transformation 68.9736 80.6995 157.7583 60.5003

4.Power law 68.9736 80.6995 68.9736 80.6995

transformation γ =1

5. γ =0.9 68.9736 80.6995 73.9827 81.1636

6. γ =0.8 68.9736 80.6995 80.1207 81.0990

7. γ =0.7 68.9736 80.6995 87.7081 80.3351

Table 3: Comparison of different method of medical image enhancement for Leg

Leg

Method Input Image Enhanced Image

Mean Std Mean Std

1.Global Histogram 60.4228 70.2646 149.6279 50.3502

Equalization

2.Adaptive histogram 60.4228 70.2646 72.8500 73.9090

equalization

3.Log transformation 60.4228 70.2646 170.7315 108.0530

4.Power law 60.4228 70.2646 60.4228 70.2646

transformation γ =1

5. γ =0.9 60.4228 70.2646 64.7082 73.3188

6. γ =0.8 60.4228 70.2646 69.4427 76.5475

7. γ =0.7 60.4228 70.2646 74.7634 79.9895

COPO1O5-9
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

AN INTELLIGENT ALGORITHM FOR TASK ALLOCATION IN DISTRIBUTED

NETWORK USING STATIC APPROACH
1
Kapil Govil, 2Dr. Avanish Kumar and 3Rakesh Kumar Dwivedi
1,3
Teerthanker Mahaveer University, Moradabad – 244 001, (UP), INDIA
2
Bundelkhand University, Jhansi – 284 126, (UP), INDIA
1
drkapilgovil@gmail.com, 2dravanishkumar@yahoo.com, 3r_dwivedi2000@yahoo.com

ABSTRACT suitable for arbitrary number of processor in

Distributed Network is one the appropriate
network for providing the optimal solution in
real-time applications problems. The
systematic allocation of tasks needs to plays
the key role to optimize the overall processing
cost of the distributed network. Distributed
network consists of multiple
autonomous computers that communicate
through a common communication channel.
The computers interact with each other in
order to achieve a common goal. A computer
program that runs in a distributed system is
called a distributed program, and distributed
programming is the process of writing such
programs. The allocation problems in any
computer system play the key role for
deciding the performance of the system. The
allocation put the direct impact of software
resources as well as hardware resources. In
distributed network, partitioning of the
application software in to module and the
proper allocation of these modules dissimilar
processors are important factors, which
determine the efficient utilization of resources.
The static model discussed in this paper
provide an optimal solution using algorithm of
Assignment problem for assigning a set of
“m” tasks of a task to a set of “n” processors
where m > n in a computer system for
evaluation for optimal processing cost of the
system. The several sets of input data are
considered to test the efficiency of the
algorithm and it is found that the algorithm is
all the cases.
Keywords: Distributed Network, Processor,
Task, Allocation, Processing Cost.
INTRODUCTION
Distributed network refers to the means by
which a single computer program runs in more
than one computer at the same time. In
particular, the different elements and objects
of a program are being run or processed using
different computer processors. In a distributed
networking, the program runs like it would in
a single computer even when it is, in fact,
using different computer processors. This
means that no single computer carries the
entire burden on system resources that running
a computer program usually implies.
One of the major research problems for
distributed networks is the allocation problem,
in which tasks are assigned to various
processors of the network, in such a way that
processing cost is to be minimized as per the
requirement. These problems may be
categorized as static (AVAN2001,
KUMA1995a, KUMA1995b, KWON2006)
and dynamic (JACQ2005, AVAN1999,
KUMA1996, KWON2006, PALM2005) types
of task allocation. Some of the other related
methods have been reported in the literature,
such as, Integer Programming (DESS1980,

COPO3O1-1
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

BRIA2003, MISR1991, MUHA2002), Load compute total processing cost by adding total
Balancing (JACQ2005, DANI2005, processing cost of task, which are assigned at
SAEE2005, SHU2009) and Modeling specified processor.
(RENE2002, JAVI2000, KENT2002). Tasks
are allocated to various processors of the ALGORITHM
distributed network in such a way that overall
processing cost of the network should be Start Algorithm
minimized. As it is well known that the tasks Read the number of task in m
are more than the number of processors of the Read the number of processor in n
network. For i = 1 to m
For j = 1 to n
OBJECTIVE
Read the value of
processing cost (c) in
In the Distributed Processing Environment
Processor Cost Matrix
(DPE), it is the common problem to allocate
tasks where the number of tasks is more than namely PCM(,)
j=j+1
the number of processors. The objective of the
Endfor
present research paper is to enhance the
i=i+1
performance of the distributed networks by
Endfor
using the proper utilization of its processors
Calculate the sum of each row and
and as well as proper allocation of tasks. In
column and store the results in
the present research paper the type of
Modified Processor Task Matrix
allocation of task to the processor is static in
MPTM(,)
nature. As in this paper the performance is
By arranging the MPTM(,) in
measured in terms of processing cost, so we
ascending order of their row_sum and
have to minimize the processing cost to obtain
column_sum, we get Arranged
the best performance of the processors. To
Processor Task Matrix APTM(,)
overcome from the problem we have designed
i=1
an intelligent algorithm for task allocation.
While all tasks != SELECTED
Select the biggest possible
TECHNIQUE
square matrix from left upper
corner and store it into SMi(,)
In order to evaluate the overall optimal
Apply algorithm of Assignment
processing cost of a distributed network, we
Problem [KANT2002] on SMi(,)
have chosen the problem where a set P = {p1,
i=i+1
p2, p3, …….pn} of „n‟ processors and a set T =
Endwhile
{t1, t2, t3, …….tm} of „m‟ tasks, where m>n.
Club processor-wise overall optimal
The processing cost of each task to each and
processing cost
every processor is known and it is mentioned
State the results
in the Processing Cost Matrix of order m x n.
End Algorithm
After making a matrix of same order taking in
ascending order of its sum of row and sum of
column, we apply the algorithm of assignment IMPLEMENTATION
problem on it. For each processor we evaluate
In the present research paper, the distributed
the overall allocation of each task; and
allocation of the task on the processor which network consist a set P of 4 processors {p1, p2,
p3, p4} and a set T of 10 tasks {t1, t2, t3, t4, t5,
has the minimum processing cost. Finally we

COPO3O1-2
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

t6, t7, t8, t9, t10}. It is shown in the figure1. The in the Processor Cost Matrix PCM(, ) of order
processing cost (c) of each task to each and 10 x 4.
every processor is known and it is mentioned
p1 p2 p3 p 4 Row_Sum
t1 t2 t3 t1 t2 t3 t1 600 200 900 300 2000
t4 t5 t6 t7 t4 t5 t6 t7
t2 500 300 200 100 1100
t8 t9 t 10 t8 t9 t 10

t3 800 700 300 400 2200

p1 p2
t4 100 400 600 300 1400
t5 400 500 600 200 1700
MP T M(,)
t6 200 100 800 900 2000
t7 200 800 600 700 2300
t8 600 700 800 900 3000
p3 p4

t9 400 500 600 200 1700

t1 t2 t3 t1 t2 t3 t10 300 700 400 200 1600
t4 t5 t6 t7 t4 t5 t6 t7 Column_Sum 4100 4900 5800 4200
t8 t9 t 10 t8 t9 t 10

Figure1: Processor Task Combination By arranging the MPTM(,) in ascending order

of their row_sum and column_sum, we get
p1 p2 p3 p4 Arranged Processor Task Matrix APTM(,) of
t1 600 200 900 300 order 10 x 4.
t 2 500 300 200 100 p1 p 4 p2 p3 Row_Sum
t 3 800 700 300 400 t2 500 100 300 200 1100
t 4 100 400 600 300 t4 100 300 400 600 1400
t10 300 200 700 400 1600
t 5 400 500 600 200
P CM(,) t5 400 200 500 600 1700
t 6 200 100 800 900 t9 400 200 500 600 1700
t 7 200 800 600 700 AP T M(,)
t1 600 300 200 900 2000
t 8 600 700 800 900 t6 200 900 100 800 2000
t 9 400 500 600 200 t3 800 400 700 300 2200
t7 200 700 800 600 2300
t10 300 700 400 200
t8 600 900 700 800 3000
Column_Sum 4100 4200 4900 5800
Now, calculate the sum of each row and
column and store the results in Modified
Processor Task Matrix MPTM(,) of order 10 x Now, selecting the biggest possible square
4. matrix from left upper corner, we get Selected
Matrix SMi(,) of order 4 x 4.
p1 p 4 p2 p3
t 2 500 100 300 200
t 100 300 400 600
SM 1 (, ) 4
t10 300 200 700 400
t 5 400 200 500 600

COPO3O1-3
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Now, by using algorithm of Assignment Processor Task

Problem [KANT2002] on SMi(,), we get task p1 t7
allocation as – p2 -
Processor Task p3 -
p1 t4 p4 t8
p2 t2
p3 t10 The overall processor-wise task allocation
p4 t5 along with optimal processing cost is
mentioned in table1.
Now, we eliminate allocated tasks (i.e. t2, t4,
t5, t10) from APTM(,); and again selecting the Table 1: Processor-wise Task Allocation
biggest possible square matrix from left upper Processor Assigned Task Processing Cost
corner, we get selected matrix SM2(,) of order p1 t4*t6*t7 500
4 x 4. p2 t1*t2 500
p1 p 4 p2 p3 p3 t3*t10 700
t 9 400 200 500 600 p4 t5*t8*t9 1300
t 600 300 200 900 Optimal Processing Cost 3000
SM 1 (, ) 1
t 6 200 900 100 800
The graphical representation of the optimal
t 3 800 400 700 300 allocation is shown in figure 2.
t7 t1
t6
t2
Now, on applying the algorithm of t4
Assignment Problem [KANT2002] on matrix p1 p2

SM2(,), we get task allocation as –

Processor Task
p1 t6
p2 t1
p3 t3
p4 t9
p3 p4
Again, we eliminate allocate tasks (i.e. t1, t3,
t3 t5
t6, t9) from APTM(,) and again selecting the
biggest possible square matrix from left upper t 10
t8
t9

corner, we get Selected Matrix SM3(,) of order

2 x 2. Figure 2: Optimal Allocation
p1 p 2
t 2 7 The processor-wise processing cost graph is
SM 3 (, ) 7 shown in the figure 3.
t8 6 9

Now, on applying the algorithm of

Assignment Problem [KANT2002] on matrix
SM3(,), we get task allocation as –

COPO3O1-4
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Processing Cost Processing Cost Table 2: Optimal Results

No. of No. of tasks Optimal
1400
1200 processors (n) (m) Results
1000 3 4 12
800
600
3 5 15
400 3 6 18
200
0
3 7 21
t4*t6*t7 t1*t2 t3*t10 t5*t8*t9 3 8 24
p1 p2 p3 p4 4 5 20
4 6 24
Figure 3: Processor-wise Processing Cost 4 7 28
Graph 4 8 32
4 9 36
CONCLUSION 5 6 30
5 7 35
In this research paper we have chosen the 5 8 40
problem, in which the number of the tasks is 5 9 45
more than the number of processors of the 5 10 50
distributed network. The model mentioned in
this paper is based on the consideration of The graphical representation of results of table
processing cost of the tasks to various 2 are shown by figure 4, 5 and 6 as mentioned
processors. The method is presented in pseudo below,
code and implemented on the several sets of m
No. of Processors = 3
input data to test the performance and Complexity

8 25
effectiveness of the pseudo code. It is the 7
common requirement for any allocation 6
20

problem that the tasks have to be processed

No. of tasks

5
15

with minimum processing cost. Here, 4

10
3
performance is measured in terms of 2
processing cost of the task that has been 1
5

processed by the processors of the network 0 0

1 2 3 4
Example
and also these tasks have been processed
optimally. As we know that, the analysis of an Figure 4: Complexity Graph
algorithm is mainly focuses on time
m
complexity. Time complexity is a function of No. of Processors = 4
Complexity

input size „n‟. It is referred to as the amount of 9 35

8
time required by an algorithm to run to 7
30

completion. The time complexity of the above 6

25
No. of tasks

20
mentioned algorithm is O(mn). By taking 5

4 15
several input examples, the above algorithm 3
10
returns following results as in table 2. 2
5
1

0 0
1 2 3 4
Example

Figure 5: Complexity Graph

COPO3O1-5
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

m
No. of Processors = 5 No. of Processors = 3 Algorithm SAGA1991
Complexity
Present algorithm
10 50
9 45 250
8 40
7 35
200
No. of tasks

6 30
5 25
150
4 20
3 15
2 10 100

1 5
0 0
50
1 2 3 4
Example

Figure 6: Complexity Graph 0

1 2 3 4 5 6

The performance of the algorithm is compared Figure 7: Comparison Graph

with the algorithm suggested by Sagar et. al. No. of Processors = 4 Algorithm SAGA1991
Present algorithm
[SAGA1991]. Table 3, shows the time 350

complexity comparison between algorithm 300

[SAGA1991] and present algorithm. 250

200
Table 3: Comparison Table 150
Time Time 100
Complexity Complexity 50
n m of algorithm of present
0
(SAGA1991) algorithm 1 2 3 4 5 6 7

O(m2n) O(mn) Figure 8: Comparison Graph

3 4 48 12 No. of Processors = 5 Algorithm SAGA1991

3 5 75 15 600
Present algorithm

3 6 108 18
500
3 7 147 21
400
3 8 192 24
4 5 100 20 300

4 6 144 24 200

4 7 196 28 100

4 8 256 32 0

4 9 324 36 1 2 3 4 5 6 7

5 6 180 30 Figure 9: Comparison Graph

5 7 245 35
5 8 320 40 REFERENCES
5 9 405 45
5 10 500 50 AVAN1999 Avanish Kumar,
Optimizing for the Dynamic
Task Allocation,
From the table 3, it is clear that present Proceedings of the „III
algorithm is much better for optimal allocation Conference of the
of tasks that upgrade the performance of International Academy of
distributed network. Figures 7, 8 and 9 show Physical Sciences,
the pictorial representation between algorithm Allahabad; (1999), 281-
SAGA1991 and present algorithm.

COPO3O1-6
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

294. Environments, Elsevier

AVAN2001 Avanish Kumar, An Inc., Vol. 60(2000) No. 6,
Algorithm for Optimal 741 – 763.
Index to Tasks Allocation KANT2002 Kant Swaroop, P. K. Gupta
Based on Reliability and and Man Mohan,
cost, Proceedings of Operations Research,
„International Conference Sultan Chand & Sons,
on Mathematical Modeling, ISBN No.: 8180540200,
Roorkee; (2001), 150-155. 2002.
BRIA2003 Brian Ensink, Joel Stanley, KENT2002 Kent Fitzgerald, Shahram
and Vikram Adve,. Latifi, and Pradip K.
Program Control Srimani, Reliability
Language: a programming Modeling and Assessment
language for adaptive of the Star-Graph
distributed applications; Networks, IEEE
Elsevier Inc., Vol. Transactions on Reliability,
63(2003), No. 12, 1082 – Vol. 51(2002), 49-57.
1104. KUMA1995a A. Kumar, M. P. Singh and
DANI2005 Daniel Grosu, and Anthony P. K. Yadav, An Efficient
T. Chronopoulos, Algorithm for Allocating
Noncooperative load Tasks to Processors in a
balancing in distributed Distributed System,
systems; Elsevier Inc., Vol. Proceedings of the „19th
65(2005), No. 9, 1022- National System
1034. Conference, SSI,
DESS1980 O.I. Dessoukiu-EI, and W. Coimbatore, (1995), 82-87.
H. Huna, Distributed KUMA1995b A. Kumar, M. P. Singh, and
Enumeration on Network P. K. Yadav, A Fast
Computers; IEEE Algorithm for Allocating
Transactions on Computer, Tasks in Distributed
Vol. 29(1980), 818-825. Processing System,
JACQ2005 Jacques Bahi, Raphael Proceedings of the „30th
Couturier, and Flavien Annual Convention of CSI,
Vernier, Synchronous Hyderabad, (1995), 347-
distributed load balancing 358.
on dynamic networks. KUMA1996 A. Kumar, M. P. Singh, and
Elsevier Inc., Vol. P. K. Yadav, An Efficient
65(2005), No. 11, 1397- Algorithm for Multi-
1405. processor Scheduling with
JAVI2000 Javier Contreras, Arturo Dynamic Reassignment,
Losi, Mario Russo, and Proceedings of the „6th
Felix F. Wu, DistOpt: A National seminar on
Software Framework for theoretical Computer
Modeling and Evaluating Science, Banasthally
Optimization Problem Vidyapeeth, (1996), 105-
Solutions in Distributed 118.

COPO3O1-7
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

KWON2006 Yu-Kwong Kwok, Anthony Vol. 51(2002), 133-140.

A. Maciejewski, Howard SAEE2005 Saeed Iqbal, and Graham F.
Jay Siegel, Ishfaq Ahmad, Carey, Performance
and Arif Ghafoor, A semi- analysis of dynamic load
static approach to mapping balancing algorithms with
dynamic iterative tasks onto variable number of
heterogeneous computing processors. Elsevier Inc.,
systems. Elsevier Inc., Vol. Vol. 65(2005), No. 8, 934-
66(2006) No. 1, 77-98. 948.
MISR1991 K. B. Misra, and U. SAGA1991 G. Sagar and A. K. Sarje,
Sharma, An Efficient Task Allocation Model for
Algorithm to solve Integer Distributed System,
Programming Problem International Journal of
arising in System System Science, Vol.
Reliability Design, IEEE 22(1991), 1671-1678.
Transactions on Reliability, SHU2009 Shu Wanneng, Wang
Vol. 40(1991), 81-91. Jiangqing, Min-Min
MUHA2002 Muhammad K. Dhodhi, Chromosome, Genetic
Imtiaz Ahmad, Anwar Algorithm for Load
Yatama and Ishfaq Ahmad. Balancing in Grid
An Integrated Technique Computing, International
for Task Matching and journal of Distributed
Scheduling onto sensor network, Vol.
Distributed Heterogeneous 5(2009). No. 1, 62-63.
Computing Systems
Elsevier Inc., Vol. 62(2002)
No. 9, 1338 – 1361.
PALM2005 J. Palmer, and I. Mitrani,
Optimal and heuristic
policies for dynamic server
allocation. Elsevier Inc.,
Vol. 65(2005) No. 10,
1204-1211.
RENE2002 Rene L. Bierbaum, Thomas
D. Brown, and Thomas J.
Kerschen, Model-Based
Reliability Analysis, IEEE
Transactions on Reliability,

COPO3O1-8
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Improvement in Dijkstra algorithm

Ashutosh Singh1, Sandeep Rathour2
Department of Electronics Engineering, H.B.T.I., Kanpur, India
1
singhash@iitk.ac.in, 2ersandeep.hbti@gmail.com

Abstract 2. Storage structure[1]

To improve the efficiency of a
network, many people have conducted many
studies. For finding the shortest path between
two points, Dijkstra Algorithm is a research
hotspot. The Dijkstra Algorithm has its own
merits and demerits when finding the shortest
path between two points. The two main
disadvantage of Dijkstra Algorithm can be
summarized as (i) Storage structure (ii)
Searching area. One more process by which
Dijkstra Algorithm can be improved. It can be
improved by reducing the no. of repeated
operation and can find the shortest path and
path length from the start point to all the
other remaining nodes by a concept of
Feature matrix.
Keywords: Dijkstra Algorithm , Storage
structure, Searching area, Shortest Path.
The storage structure of Dijkstra
Algorithm is large i.e. it takes large space for
storing it.
Let take a example of 10 nodes. For the
classical Dijkstra Algorithm it requires
10x10[100 values] for its storage. The figure
1 shows the example for 10 nodes. Since it
require 10x10 matrix for storage, the matrix
shown in figure 2 have following properties.
(a) if the start point and the end point are
same,
i.e. i=j, then aij=0
(b) aij=aji
(c) if there is no any direct path between
two nodes , then it has ∞(large value) value.
Example
3
4 3

2 3
2 9 4
1. Introduction 2
5

In the graph theory, Dijkstra 5

2 5
1
Algorithm is the most classical algorithm for 3
10
4
finding the shortest path in the graph. But the 3
1 6
8
time complexity of the Dijkstra Algorithm is
3 3
higher and also it takes larger space for 7

storage. To overcome these shortcomings of

Dijkstra Algorithm many scholars have done
many researches and their result have its own
merits and demerits. In this paper there are
some improvement in classical Dijkstra
Algorithm to reduce the time complexity of
the Dijkstra Algorithm and also reduce the
searching area of the algorithm. And also for
improvement in Dijkstra Algorithm, if the no.
of repeated operation for searching the
shortest path is reduced, then the efficiency of
the Dijkstra Algorithm is also get improved.
Figure.1 nodes arrangement(10 nodes)
For reducing the storage structure of the
above network shown in figure.1, in the
matrix shown in figure.2, the places where the
values are 0(zero) and ∞(infinity) can be
removed, only non-∞ elements are stored in
the simple linked list format. The improved
structure is shown in figure.3. in this form
each node is having only two elements, first
one is the column value and second one is its
weight.

COPO3O2-1
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

4. Searching Area
0 2 ∞ ∞ ∞ ∞ ∞ 3 ∞ ∞ Since the searching area of a classical
2 0 4 ∞ ∞ ∞ ∞ 3 2 ∞
Dijkstra Algorithm is large, there are some
∞ 4 0 3 ∞ ∞ ∞ ∞ ∞ 5
process by which the searching area of the
∞ ∞ 3 0 5 2 ∞ ∞ ∞ ∞
∞ ∞ ∞ 5 0 4 ∞ ∞ ∞ ∞
classical Dijkstra algorithm can be reduced.
∞ ∞ ∞ 2 4 0 3 ∞ ∞ ∞ Since the shortest path between two points is
∞ ∞ ∞ ∞ ∞ 3 0 3 ∞ 1 a straight line. The direction from the start
3 3 ∞ ∞ ∞ ∞ 3 0 ∞ ∞ point to the is generally strike of the shortest
∞ 2 ∞ 3 ∞ ∞ ∞ ∞ 0 5
∞ ∞ ∞ ∞
path when plan the route of the real road
∞ 5 ∞ 1 5 0
network. The shortest path between two
points is generally the both side of connection
line of the start point to the destination point
Figure.2 adjancy matrix(10x10) and usually it is near to the connection line. If
there are only one edge between the start
The storage format of the above matrix get point and the destination point, the edge itself
reduced and given as- is the shortest path.

G1 2 2 8 3 ^ 4.1 Searching Area is rectangle

G2 1 2 3 4 8 3 9 2 ^ The author[1] shows in figure 4 to
G3 2 4 4 3 ^
find the shortest path from the source S to the
destination D, the theoretical searching area
G4 3 3 5 5 6 4 9 3 ^
of classical Dijkstra algorithm is a circle
G5 4 5 6 4 ^ whose center is source S and radius is R. the
G6 4 2 5 4 ^ restricted searching area algorithm, the
G7 6 3 8 4 10 1 ^ theortical searching area is small rectangle. It
G8 1 3 2 3 7 4 ^
may be considered that there may exist a
reserve path, each side of the small rectangle
G9 2 2 4 3 10 5 ^
can be increased to threshold T2 , and it make
G10 7 1 9 5 ^ a large rectangle. So the restricted searching
area is formed by the two lines L1 and L2
figure.3 improved storage strucuture which cuts the larger rectangle, so the final
searching area is shown by dark line in figure
From the above figure it can be told that this 4.
structure takes less space in comparison to the
classical Dijkstra algorithm.

3. Complexity Analysis
Let for the list array, the space
complexity is given as O(T). where T is the
edgentain one or more tables to no. of the
directed graph.
In the worst case T=n2 , then the space
complexity will be O(n2).

COPO3O2-2
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

the distance between the neighbours, the time

L1
complexity difference is more significant. So
this improved algorithm restricting the
T1 L2 searching area reasonably reduce the scale of
D
searching, so it minimizes the complexity.
T2
And also it improves the efficiency of the
system. The improved algorithm itself restrict
the searching area, unfortunately it may be
S
the algo. Sometimes can not search the actual
R2 shortest path and it can only be the an
approximate shortest path, it may be a
defective shortest path algo and also it reduce
the accuracy.
Figure.4 rectangular searching area

4.2 Searching area is an ellipse

In the figure, S is the start point, When searching area is an ellipse
D is the destination point, shown in figure 5, the following parameter
T1 & T2 are threshold. are used for finding the searching area. In
R is the distance between source and case of ellipse,
destination. SD=R=2a [major axis of ellipse],
(i) for the classical Dijkstra Algorithm the Let the eccentricity e=0.7 [<1],
searching area is a circle of radius R , then the b [semi minor axis] can be obtained.
will be In this case the searching area is shown in
A1=π*R2 figure 5 by a thick line, and can be obtained
(ii) for the restricted searching area algorithm as,
the searching area is given by A2 = T1(4a+πT2)
A2=2T1[R+ T2] Now the time complexity ratio can be given
Where T1/T2= as
(iii) since the searching node is the directly Time complexity ratio=
proportional to the searching area, so the time
complexity can be given as O(A12) and = [ ]2
O(A22).
Commonly two threshold T1 & T2 are
relatively small constant. SD=2a(major axis)
Eccentricity(e)=.7
L1
T2
So the ratio of the time complexities can be D L2
T2
expressed as –
T1
Time complexity ratio= T1

S
T2
= [ ]2

From the above equation it can be compared

that the searching area algorithm is smaller
than the classical Dijkstra algorithm. If there
is increase in the scale of road network and figure 5 when searching area is an ellipse

COPO3O2-3
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

from the above equation, it can be compared The abutment matrix for the example is
that the ratio of time complexities of, when given as A=
searching area is an ellipse is less in
comparison to the ratio of time complexities
0 2 1 ∞ 7 12
of , when searching area is rectangular. 2 0 ∞ 2 ∞ ∞
Elliptical searching area gives better result. 1 ∞ 0 1 2 ∞
∞ 2 1 0 3 ∞
5. Feature matrix[2] 7 ∞ 2 3 0 4
Another way for the improvement of 12 ∞ ∞ ∞ 4 0
classical dijkstra algorithm. Since if we to
find the shortest path between two points, for
this purpose there are a no. of operations by There are following steps which from which
which it can be done. But if in any way it can feature matrix can be obtained.
reduce these no of operations, then the 1. source S={v1}, D={0,2,1,∞,7,12}
efficiency of the classical dijkstra algorithm 2. firstly find the shortest distance from the
will get increases. For this purpose there is a source i.e. D[3]=1, so S={v1,v3}
concept of feature matrix. From this feature 3. now the nodes connected via 3 to the 1 ,
matrix we can draw the shortest path tree and distance can be obtain by
can get the shortest path and shortest path D[3]+A[3][4]=2<D[4]=∞
length from source node to all other & D[3]+A[3][5]=3<D[5]=7
destination node. So that modification will D[4]=2 ,D[5]=3 and
For understand the concept of feature matrix the D matrix will become D={0,2,1,2,3,12}
let’s take a example of 6 node which are 4. iterate operation as follows,
connected as shown in figure.6. The second time we get that,
S={v1,v3,v2,v4}
& D={0,2,1,2,3,12}
The third time we get S={v1,v3,v2,v4,v5}
1
12 & D={0,2,1,2,3,7}
6

2 7
The fourth time S={v1,v3,v2,v4,v5,v6}
1 4 & D={0,2,1,2,3,7}
3
2
2 5 5. compare the following D and A matrix we
2
1 will get the following matrix called feature
4
3
matrix(F).

0 0 0 0 0 0
Figure.6 nodes arrangement
2 0 0 0 0 0
1 0 0 0 0 0
0 0 1 0 0 0
0 0 2 0 0 0

0 0 0 0 4 0

COPO3O2-4
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
6.From the feature matrix it can be seen that
v1 is the precursor node of the node v2 and
v3. v3 is the precursor node of the node v4
and v5, and v5 is the precursor node of the
node v6.
7. now shortest path tree can be obtained like
shown in figure.7.
1
1
2 to any other node by reducing the no. of
3 2
1 2
4 5
4
6
Figure.7 shortest path tree
From the shortest path tree we can find the
shortest path from the source to the all other
nodes and also can find the path length.
6. Conclusion
Finally after studying the above
algorithm, it can concluded that the restricted
searching area algorithm has lower time
complexity than classical dijkstra algorithm
and also when searching area is an ellipse has
lower time complexity when searching area is
a rectangular.
By using the concept of feature matrix, it can
be found the shortest path from source node
searching operation which increases its
efficiency.
7. References:
[1]. Dong Kai fan, Ping Shi, “Improvement of dijkstra
algorithm and its application in route planning”,
Shandong University of technology, China, IEEE,
International conference on FSKD, 2010, pp 1901-
1904.
[2]. Ji-Xian Xiao, Fang-Ling Lu, “An improvement of
the shortest Path algorithm based on Dijkstra
algorithm”, IEEE, ICCAE, 2010, pp. 383–385.
COPO3O2-5
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
EXECUTED WRIST MOVEMENT CLASSIFICATION FOR
BRAIN COMPUTER INTERFACE
Mariam Siddiqui, Ayman Nidal Khalidi, Omar Farooq,Yusuf U Khan
Department of Electronics Engineering, Department of Electrical Engineering
Zakir Hussain College of Engineering & Technology, Aligarh Muslim University, Aligarh (India)
Abstract—The aim of our study is to
investigate best features for the
classification of pronation and extension
of wrist movements and further use the
best possible classifier to achieve accurate
classification of movements. 10-20
International Montage was used for the
recording of EEG data on 16 channels.
Data was processed, filtered and
windowed. Histogram plots were used for
selection of features. Standard statistical
features were employed in the
investigation and the result suggests that
they provide discriminatory conditions.
While the scope and depth of the study
was limited, the results do suggest that
variance, mean, skewness and kurtosis
can be used for designing the classifier for
both the movements. Different Standard
classifiers are finally designed and trained
to distinguish between different motor
movements.
Keywords:BCI, EEG, Pronation, Extension, Variance,
Skewness, Kurtosis, Mean
I. INTRODUCTION
A brain-computer interface (BCI) is a
system which translates a subject’s
intentions into a control signal for a device,
e.g., a computer application, a wheelchair or
a neuroprosthesis[1]. BCI has given us hope
where an alternate medium can be created
for people with severe motor disability so
that they can somehow work over their
disability to communicate with the external
world in a non-biological and easier way.
Encouraged by this understanding , a
growing interest has developed in recording,
detecting and analyzing brain (neural)
signals to investigate, explore and
understand human motor control systems
with the aim of attempting to build
interfaces that use real– time brain signals to
generate commands to control and/or
communicate with the environment [1].
We mainly focus on noninvasive,
electroencephalogram- (EEG) based BCI
systems which can be used as tools of
communication for the disabled or for
healthy subjects who might be interested in
exploring a new path of human-machine
interfacing. When measured non-invasively,
brain activity is acquired by scalp-recorded
EEG from a subject that tries to convey its
intention by reacting to well-defined tasks or
instructions [1]. Research in developing
successful brain–computer interfaces (BCIs)
endeavors to improve the speed and
accuracy of communication of these
interfaces. This is achieved by improving the
EEG signal processing techniques and
selection of features that are translated into
commands to control and communicate with
augmentative and assistive devices [3].
In this paper we propose and experimentally
verify different features that can be used for
training the classifier for wrist movement in
two directions. A histogram-based analysis
was applied to the EEG signals from
COPO303-1
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27
2011
different channels, to determine combination
of features and channels that yielded the best
class separability. Standard classifiers were
then designed and trained to achieve the best
possible classification accuracy
II. BACKGROUND RESEARCH
This domain is still quite fresh and
acqiuring different shapes ,since it emerged
40 years back in 1970.At that time, the
research subjects were initially animals(rats
and monkeys)but then it took a dramatic
shift in 1990s when the first successful
working implant in humans occured .Now
not only damaged eyesight, hearing,
impaired movements’ etc. can be restored
but researchers could now conceivably
attempt to produce BCI that augment human
functions rather than simply restoring
them.[1]
In order to meet the requirements of the
growing technology expansion,some sort of
standardization was required not only for the
guidance of future researchers but for the
validation and checking of new
developments with other systems as
well,thus a general purpose system was
developed called BCI2000 which made
analysis of brain siganl recording easy by
defining the output formats and operating
protocols to facilitate the researchers in
developing any sort of application.This
made it easier to extract specific features of
brain activity and translate them into device
control signals. [4]
.
III. METHODOLOGY
Fig1: Block diagram for feature extraction of wrist movement
A.Experimental Setup and Data Acquisition
For this study, data was recorded from three
healthy subjects, who signed consent before
their participation in the study. EEG was
recorded at the biomedical signal processing
lab, department of electronic engineering,
AMU.16 channels of EEG data were
recorded using a Brain Tech clarity TM
system with the electrodes positioned
according to the standard 10-20 system.
Sampling rate was 256 Hz. The subject was
seated on an armchair with eyes closed. The
subject was instructed to avoid speaking and
to ensure least body movement. The
subjected was instructed to move the wrist
from neutral position on an audio prompt
from the experimenter. The series of trials
recorded consisted of following sequences:
First of all the subject was asked to relax for
some time then followed by pronation, after
reaching the extreme position the subject was
asked to hold for some time which was
followed by extension.
Fig2 Sequence followed during
experimental recording
Times of recording of different movements
were different.
B.Data Processing
EEG data was notch filtered to remove 50Hz
frequency from each channel present in the
data due ac power supply of the EEG
machine. Harmonics of 50Hz frequency
were also removed from the data. IIR second
order notch filter with the quality factor (or
Q factor) of 3.55 was used to remove the
desired frequency.
COPO303-2
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27
2011

Magnitude Response (dB) then averaged over all possible channels

5
.These features could further be used to
classify between the given movements and
0
between movements and neutral state.
-5
Magnitude (dB)

-10

-15
D. CLASSIFIER DESIGN:
Filter #1
-20 Motor movements are finally differentiated
-25 by using different classifiers on the recorded
-30
data
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Normalized Frequency ( rad/sample) Table1: percentage values of classification for motor
Fig3: Magnitude response of a notch filter designed at 50Hz (or positions
.39∏ radian per sample) frequency

Power spectral density of filtered and

Classifier Extension Pronation
unfiltered EEG signal of all the 16 channels
stacked over one another is shown below SVM 77 80
1200 Quad 90 88.8
1000 Linear 80 77.7
800 Diaglinear 80 66.6
600
amplitude

400
From the observations in Table1 we see that
Quad classifier gives us the best possible
200 results with high classification percentage
0
accuracy for both extension and pronation
movements.
-200
0 20 40 60 80 100 120 140
frequency

Fig4: PSD plot of 16 channel filtered and unfiltered EEG signal

IV. RESULT AND DISCUSSION

EEG by nature is non stationary signal. So it
was fragmented into frames so that it can be Histogram plots for 16 channels for different
assumed stationary for small segment. EEG movements pave way for demarcating
data is divided into frames of 1s duration i.e. different movements. From Fig 5 which
frame size of 256 samples /s. shows the histogram plot for electrode F3 for
extension shows that the data has long tail,
hence kurtosis could be one of the features.
C.Feature extraction Further, we can see in the same figure that
The recorded data was performed for both the data is asymmetric about the mean
pronation and extension wrist movement position which shows that it skewed towards
related activity and neutral state of right right. Moreover, if we compare Fig 5 & 6,
hand. Histogram of each channel was plotted which is plot for pronation shows that both
and compared for each of the three different dynamic range and have different
movements and the best possible plots which mean values. Likewise, compare Fig 5, 6,
showed maximum demarcation were used to and 7 they differ in their dynamic ranges. Fig
extract possible features. Further,from16 7 which shows the histogram plot of neutral
channel filtered and segmented data for each state doesn’t have long tail and it doesn’t
of motor activity and non-motor activity’s have much variation about the mean position
basic statistical features [2] mean, variance, thus statistical features can be used to easily
skewness and kurtosis were extracted and

COPO303-3
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27
2011

differentiate between motor and non-motor 4000

F3-neutral state

activity.
3500

Thus from the above discussion we conclude 3000

that mean, variance, kurtosis and skewness

frequency of occurance
2500
can prove to be good features for
classification of wrist movement. 2000

1500

1000

500

F3-Extension 0
2500 -40 -30 -20 -10 0 10 20 30 40
Range of values

Fig7: Histogram plot of frontal electrode (F3) for Neutral

2000 state.
frequency of occurance

1500

features Neutral Extension Pronation

1000 position
Variance 120.16 586.85 299.06
500
Mean -0.07 0.33 0.24
kurtosis 4.77 6.44 5.75
0 skewness 0.13 0.42 0.28
-120 -100 -80 -60 -40 -20 0 20 40
range of values

Fig5: Histogram plot of frontal electrode (F3) for Table2: Statistical features values for different
extension.
positions

F3-Pronation
2500 From the result tabulated in Table2
Extension has larger variance, mean, kurtosis
2000 and skewness values while pronation has
smaller values. Also for both the motor
frequency of occurance

1500 activity values of features are far away from

the non-motor activity. Thus both the
1000 movements can be easily demarcated from
each other.
500

0
From Table1, we justify our findings by
-60 -40 -20 0 20 40 60
range of values designing different classifiers and training
Fig6: Histogram plot of frontal electrode (F3) for
them for the two class movements to get the
pronation. best possible accuracy from the 16 channel
recorded sample values

COPO303-4
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27
2011

V. CONCLUSION
EEG data class separation was investigated
for two wrist movement classes, pronation
and extension using a 16 channel setup. Four
features were extracted from the histogram
plots of signal for both the movement and
neutral state. Variance, mean, skewness and
kurtosis are the features which easily
demarcates between the two movements. It
was a three class problem since three data
sets were used. Different classifiers were
designed and trained based on the recorded
data sets on the basis of which we classified
between motor movement

VI. ACKNOWLEDGEMENT
The authors are indebted to the UGC.This
work is a part of the funded major research
project CF.No 32-14/2006(SR)

REFERENCES

1. Wolpaw, J.R., Birbaumer, N.,

McFarland, D.J., Pfurtscheller, G., Vaughan,
T.M.: Brain–computer interfaces for
communication and control. Clinical
Neurophys. 113 (2002) 767–791
2. Vaughan, T.M.: Brain-computer
interface technology: a review of the second
international meeting IEEE Trans. on Neural
Systems and Rehab. Eng. 11(2) (2003) 94–
109
3. Syed M.Siddique, laraib Hassan
Siddique: EEG based Brain computer
Interface:Journal of software.4(6)(2009)550-
555
4. F. Sepulveda, M. MeckesCluster
:.Separation Index Suggests Usefulness of
Non-Motor EEG Channels in Detecting
Wrist Movement Direction Intention:
Proceedings of the 2004 IEEE Conference
on Cybernetics and Intelligent Systems

COPO303-5
 Color Images Enhancement Based on Piecewise
Linear Mapping
Anubhav kumar 1, Awanish Kr Kaushik 1 , R.L.Yadava 1, Divya Saxena 2
1
Department of Electronics & Communication Engineering
Galgotia’s College of Engineering & Technology, Gr.Noida, India
2
Department of Mathematics
Vishveshwarya Institute of Engineering and Technology , G.B.Nagar,India
rajput.anubhav@gmail.com

illumination and to enhance contrasts in the image. It’s a

Abstract- Image enhancement is a basic process for all vision and
image processing function in various areas like computer vision,
biomedical image analysis, forensic video/image analysis, remote
sensing and fault detection image enhancement technology plays
a very important role in image processing. By enhancing some
information and restraining other information selectively, it can
improve image visual effect. For a digital image enhancement
process, the digital image is represented in the YCbCr color
space. In this paper, an enhancement arithmetic is used for color
images enhancement based on piecewise linear mapping is
discussed. The method can improve the image sharpness and
make the image characteristics more evident. The enhancement
result is useful for further analysis. Experimental results tested
from a data set of images have demonstrated that the proposed
method is effective and practical. Some parameters like as
RFSIM (Image Quality Assessment) are analyzed for both
existing and proposed method to determine the success and
limitation of our method.
frequency filtering, preferred to others techniques because it
corrects non uniform lighting and sharpens the edges at the
same time [2]. Another method for color image enhancement
is Bright stretching, also used auto scaling method which is a
common linear mapping function to enhance the brightness
and contrast level of an image [10].
Color image processing and enhancement is a more
complicated process than its counterpart of black-and-white
images [4] due to the involvement of multiple color channels
and the need to preserve the color information content [5]
while enhancing the contrast. Novel techniques that address
these problems are in great demand. For instance, in [6], Chen
et al. proposed to enhance the image quality by making use of
local contrast information and fusing the morphologically
enhanced image with the original.

Keywords - Color image enhancement, RGB color space, YCbCr
color space, Piecewise linear mapping, RFSIM.
I.INTRODUCTION
Image enhancement processes consist of a collection of
techniques that seek to improve the visual appearance of an
image or to convert the image to a form better suited for
analysis by a human or machine. Nowadays there is a rapid
increase in the application of color video media. This has
resulted in a growing interest in color image enhancement
techniques.
Some other common technique to enhance the contrast of
images is to perform histogram equalization and
Homomorphic methods. The advantage of histogram
equalization technique is that it works very well for grayscale
images, however, when histogram equalization is used to
enhance color images, it may cause shift in the color scale
resulting in artifacts and an imbalance in image color. The
homomorphic filtering is used to correct non uniform
Similarly Fairweather [8] have used techniques such as
contrast stretching and Markov Random field. They applied
bimodal histogram model to the images in order to enhance
the underwater image. Yoav [9] have used a Physics-based
model. They developed scene recovery algorithm in order to
clear underwater images/scenes through polarizing filter. This
approach addresses the issue of backscatter rather than blur.
In this paper color images enhancement based on piecewise
linear mapping has been proposed .In Section II, an illustration
of theoretical foundations of piecewise linear mapping
function are presented. The proposed enhancement algorithm
is developed in Section III. Experiments conducted using a
variety of color images are described in Section IV and results
are discussed. A conclusion is drawn in Section V.
II. THEORETICAL FOUNDATIONS
Piecewise linear Mapping Function –
The enhancement is done by first transforming the intensity
values using a piecewise linear mapping function for the
intensity component. The piecewise linear mapping function
consists of three line segments as indicated in Figure 1, where
vmin and vmax denote the minimum and maximum intensity
levels in the original image, respectively. This type of
mapping function permits proper allocation of the dynamic
range to different ranges of intensity levels. The actual values
of vlower and vupper determine the dynamic range allocation
for lower, intermediate and higher ranges of intensity levels.
Fig. 1 The piecewise linear mapping function for intensity
Component.
The first moment of the intensity histogram serves as an
indicator about the exposure of the image. The first moment is
closer to vmin for an under exposed image, and closer to vmax
for an over exposed image. Thus based on the first moment of
the histogram, two intensity levels, vlower and vupper are
determined such that the closer the first moment is to vmzn,
,  < 1
the larger the vupper is, and the smaller the dower is, and vice
original image are transformed to create the optimally
enhanced color image in NTSC YCbCr space directly.
Fig. 2 Flow chart of Proposed Algorithm
Step 4 - In the last contrast increases by adjusting image
intensity values. It suppresses eventually outliers’ pixels to
improve contrast stretching. It then stretches contrast to use
the whole range of intensity channel and if necessary it
saturates some low or high values. Contrast stretching
attempts to improve the contrast in an image by stretching’ the
range of intensity values it contains to a desired range of
values, e.g. the full range of pixel values that the image
concerned allows.
 0 <

,
 

 
g,  = 0  0 >
,  
versa.

1  1 <
, 
(1)

III. PROPOSED ALGORTHEM

The color image enhancement based on piecewise linear
mapping function is specified as follows.
Step 1 – In the first step, firstly read the input image and
reshape the input color image (m*n size) in to (m*n)*1
column vector using array. After this change the Sort array
elements in ascending or descending order in the image has
been done.
Step 2 – In the second step, Employ ceil and array function
and find out Vmin and Vmax.Let the full dynamic range for the
display device be [VMIN, VMAX], where VMIN and VMAX
correspond to the minimum and maximum displayable
intensities, respectively.
Step 3 - Using the luminance and chrominance mapping
functions, the luminance and chrominance values of the
where, g( x,y ) represents the images, on the left side it
represents the output image, while f(x,y) represents input
image. In this equation the min and max are the minimum
intensity value and the minimum intensity value in the current
image.
Enhanced Image = g( x,y ) * No. of intensity levels (2)
Here No. of intensity levels shows the total number of
intensity values that can be assigned to a pixel. For example,
normally in the gray-level images, the lowest possible
intensity is 0, and the highest intensity value is 255. Thus "no.
of intensity levels" is equal to 255.
IV. RESULT AND ANALYSIS
On our personal laptop of Pentium 4 the experiment has been
carried out at 2.4 GHz. The test set for this evaluation
experiment randomly selected from the internet with 256*256
and 512*512 sizes. Figures(3a,4a,5a,6a ) show the original
color images, figure (3b,4b,5b,6b) show the enhanced image
by using Murtaza method [1], and the figure (3c,4c,5c,6c)
show the enhanced images with the help of our proposed .

(a ) (b)

(a ) (b)

(c)
Figure-4 –Watch image (a) Original image (b) Enhanced from
Murtaza method [1] (c) Enhanced from Proposed Method

(c)
Figure-3 -Lena image (a) Original image (b) Enhanced from Murtaza
method [1] (c) Enhanced from Proposed Method

(a ) (b)

(a ) (b)

(c)
Figure-4 –Play image (a) Original image (b) Enhanced from Murtaza
method [1] (c) Enhanced from Proposed Method

Table 1 show the comparison of performance of our proposed

(c)
Figure-4 –Flower image (a) Original image (b) Enhanced from
Murtaza method [1] (c) Enhanced from Proposed Method
method for enhancement with existing method, where our
proposed method has a better performance in contrast
enhancement and Image Quality Assessment (RFSIM). A
novel IQA model, namely Riesz-transform based Feature
Similarity index (RFSIM) method used to evaluate the existing
and proposed experiment result.
This naturally leads to the following formula to define the transforms", in: Proc. IEEE International Conference on Image
similarity between the feature maps fi and gi by considering Processing, IEEE 2010.
only the key locations marked by mask M, [4] N. M. Kwok, Q. P. Ha, D. Liu, and G. Fang, “Contrast

∑ ∑  ,.  ,
enhancement and intensity preservation for gray-level images

∑ ∑  ,
using multi objective particle swarm optimization,” IEEE
Di = (3) Trans. on Automation Science and Engineering, vol. 6, no. 1,
pp. 145–155, 2009.
The similarity between two feature maps fi (i = 1~5) and gi at the [6] S. K. Naik and C. A. Murhty, “Hue-preserving color image
corresponding location (x, y) is defined as enhancement without gamut problem,” IEEE Trans. on Image

 ,.!,"#
Processing, vol. 12, no. 12, pp. 1591–1598, 2003.

 $ ,.! $ ,"#
[7] Q. Chen, X. Xu, Q. Sun, and D. Xia, “A solution to the
di (x,y) = (4) deficiencies of image enhancement,” Signal Processing, vol.
90, pp. 44–56, 2010.
[8] A J R Fairweather, M A Hodgetts, A R Greig, “Robust

= ∏'() &
Then, we compute the RFSIM index between f and g as scene interpretation of underwater image sequences”, In 6th
International Conference on Image Processing and its
RFSIM [3] (5) Applications, 1997, pp. 660 -664, ISBN: 0 85296 692 X
[9] Schechner, Y and Karpel, N., “Clear Underwater Vision”.
Proceedings of the IEEE CVPR, Vol. 1, 2004, pp. 536-543.
Table-I [10] M.Isa,M.Y.Mashor,N.H.Othman,"Contrast Enhancement
Image Processing on Segmented Pap smear Cytology Images",
RFSIM (Image Quality Assessment) Prof. of Int. Conf. on Robotic, Vision, Information and Signal
Images Murtaza et al. [1] Proposed Method Processing.pp. 118 – 125, 2003.
Lena Image 0.3445 0.7261
Flower Image 0.4183 0.8321
Watch Image 0.4604 0.8953
Play image 0.4577 0.7614

The proposed model was implemented using five images. The

implementation results were compared with other results and
the performance of our proposed method is better than that of
existing color image enhancement algorithms.

V .CONCLUSION

In this paper color images enhancement based on piecewise

linear mapping has been presented. From the above results and
the Table -I, it is clear that the RFSIM (feature based image
quality assessment) of the proposed algorithm is high.
Experiments show that our proposed method obtains better
enhancement result, and it is more effective for color image
enhancement in normal and noisy environment

V. REFERENCE

[1]. Syed Mashhood Murtaza, Jahanzeb Ahmad, Usman Ali

‘Efficient Generalized Colored image Enhancement’ in 2006
IEEE conference.
[2]. C.M. Sujatha, K. Navinkumar, K. S. Arunlal, S. A. Hari,
Performance evaluation of Homomorphic filtering,
Anisotrophic filtering and Autocontrast algorithm,
International Conference on Advances in Computing, Control,
and Telecommunication Technologies, IEEE 2009.
[3] Lin Zhang, Lei Zhang, and Xuanqin Mou, "RFSIM: a
feature based image quality assessment metric using Riesz
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
Information Fusion Based on Multiplicity of
Data Preprocessing Improves AdaBoost
Classification Efficiency
Divya Somvanshi and R. D. S. Yadava
Abstract—The paper considers the possibility of boosting
efficiency of the AdaBoost algorithm by introducing diversity in
the information extraction procedures and subsequent
information fusion. A given raw data set is transformed by two
well known methods known as vector autoscaling and
dimensional autoscaling. These methods transform data by mean-
centering and variance-normalization with respect to the
measurement variables and the measured samples, respectively.
The feature (information) extraction is implemented by the
principal component analysis. Final data representation is
obtained by two alternate fusion strategies. In one, feature
extraction is done immediately after the scaling transformations
and the two feature sets are fused by a simple procedure of
concatenation. In the other, data vectors obtained by the two
scaling transformations are first fused by simple concatenation
and then the feature extraction is done. The data representation is
thus by fused feature vectors of dimensionality twice that of the
original data. The AdaBoost algorithm by Freund and Schapire
is implemented by using a simple threshold base classifier. The
classification efficiency of these procedures is compared with that
obtained by using single preprocessor based AdaBoosting. Four
benchmark data sets are used for validation. The analysis
demonstrates that the performance of the AdaBoost algorithm is
enhanced further by multiple preprocessor based fusion of data
space, particularly if the original variables are of same kind.
Index Terms—Boosting the AdaBoost algorithm, data
preprocessing, information fusion for pattern recognition,
machine learning.
I. INTRODUCTION
T he success in any pattern recognition problem depends on
three factors: quality of raw data, efficiency of information
extraction procedures, and classifier algorithm in
succession. An improvement in each boosts the performance
of its successor. The overall performance of the pattern
recognition algorithm depends on the optimization of all steps
Manuscript received March 5, 2011. This work was supported by the
Government of India, Defence Research and Development Organization
Grant No. ERIP-ER-0703643-01-1025.
Divya Somvanshi is with the Department of Electronics Engineering,
Institute of Technology, Banaras Hindu University, Varanasi 221005, India
(e-mail: divya.jrf@gmail.com).
R. D. S. Yadava is with the Department of Physics, Faculty of Science,
Banaras Hindu University, Varanasi 221005, India (phone: 91- 9450246572
mobile; fax: 91-542-2368390; e-mail: ardius@ gmail.com,
ardius@bhu.ac.in).
doing their operations in tandem. Given a raw data, the
objective of information extraction procedures is to represent
the data in a form that optimally helps the classifier. The
information extraction consists of cleaning up the data for
noise, removal of outliers, shifting and scaling, and
transformation to generate a new representation of data which
is potentially more helpful for pattern classification. The data
transformation involves combining raw data variables (often
linearly) to create new set of variables by focusing on certain
statistical attributes such as variance, statistical independence,
or separability. This is referred to as feature extraction. The
procedures of data preparation for feature extraction are
referred to as preprocessors. Some commonly used feature
extraction methods are principal component analysis,
independent components analysis and linear discriminant
analysis. The new variables (feature components) define the
feature space, and the representation of a sample vector in this
space is called the feature vector or pattern. The feature
vectors are representation of raw data vectors, and are input to
the classifier for discrimination of patterns of different classes.
The pattern classifier algorithms map feature space (in effect
input data space) to different class labels or object identities. A
number of pattern classification algorithms are developed that
use various strategies for creating input data space to output
class space mapping functions. For example, neural network
methods produce nonlinear mapping by optimizing feature
weights and node activation functions, support vector
machines create large dimensionality hyperspace using a
kernel function, and discriminant analysis seeks separation
boundaries in the feature space. An overview of these methods
can be found in [1], [2].
A pattern classification algorithm is basically a decision
function which computes decision values by taking feature
vectors as input. The construction of decision functions has
been inspired by a variety of paradigms from statistical and
biological domains. For example, the clustering and
discriminant function algorithms implement statistical models,
while the neural network and genetic algorithms are bio-
inspired [2]. Given a pattern classifier, weak or strong, its
performance can be further enhanced by methods of so called
„ensemble learning‟ [3]. The ensemble learning methods use a
number of classifiers and combine their decision functions in
some way to produce a new classifier with enhanced
capabilities. It was proved by Schapire [4] that a number of
weak learners trained on the same set of examples can be
COPO305-1
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

combined to yield a strong learner with enhanced predictive The class identities were given numeric representation of 1
capability. The ensemble learning methods produce a fused and 1 . That is, yi Y 1, 1 for all x i X for all
decision fusion. In principle the performance of any classifier i 1, 2,..., N .
can be enhanced. However, usually the risk of overfitting Step 1. Define input. It consists of N training examples, a
prompts to use an ensemble of weak learners [5]-[7]. There is base learning algorithm, and number of training runs T .
no limit on the number of learners in the ensemble. The Step 2. Initialize weight distribution over training examples
performance of an ensemble learning method can be optimized
according to w1 (i) 1 / N for i 1,2,..., N where i stands for
by proper selection of ensemble size and fusion method to suit
any specific application [5], [8]. the training examples and subscript 1 denotes the first round
Most popular among many ensemble learning algorithms is of T runs to determine weights wt (i ) . In the first round all
the AdaBoost (nickname for adaptive boosting) algorithm training examples are assigned equal weights. This assignment
described by Freund and Schapire [9]. A summary of the ensures that weight distribution is normalized, that is,
AdaBoost algorithm is given in section II. The objective of N
present study is to show that the predictive ability of the w1(i) 1.
i 1
AdaBoost algorithm can be further boosted by varied
Step 2. Start a loop „for t 1 to T ‟ that will create
representation of the same raw data by employing several
sequential base classifiers according to the following substeps
preprocessing and feature extraction methods. Different
in succession:
preprocessing and feature extraction procedures reveal hidden
Train the base classifiers ht based on the example
data structure from different perspectives, and yield alternate
sets of features to represent the same example. By combining weight distribution wt (i ) .
these sets in some way can, in principle, provide a more Determine the training error defined as t wt i
reliable and accurate representation. Motivated by this idea, i:ht xi yi
we report here a study on enhancing performance of the which is the sum of weights for all misclassified
AdaBoost algorithm by using a simple model of feature fusion examples by ht .
based on the two common preprocessors and one feature Assign weight to the t-th classifier ht according to:
extractor combination. The procedure is described in section
III. Using a linear threshold classifier for the AdaBoost 1 1 t
t ln .
ensemble generation, section IV presents validation results 2 t
based on some benchmark data sets available from open Update the example weights according to:
sources. The paper concludes with a discussion in section V.
wt i e t for ht xi yi
wt 1 i
zt e t for ht xi yi
II. THE ADABOOST ALGORITHM where z t is the normalization factor for making
The AdaBoost [9] is a supervised learning boosting wt 1 (i) to be a probability distribution. That is, for
algorithm that produces a strong classifier by combining N
several weak classifiers from a family. It needs a set of training making wt 1 (i) 1.
examples and a base learning algorithm as input. Let i 1
X x1 , x 2 ,..., x N be the set of N training vectors drawn End loop.
Step 3. Output the boosted classifier as weighted sum of
from various target classes, and let Y y1, y2 ,..., y N denote
the T base classifiers
their class labels. The base learner is chosen such that it T
produces at least more than 50% correct classification on the H x sign t ht x .
training set. Let base learner be denoted by ht x . The basic t 1

idea of ensemble generation in the AdaBoost algorithm is to

assign a distribution of weights on the training examples
according to the error rate in such a manner that, in
sequentially repeated application of the base learner, the
misclassified examples get more weight in the current
application than the preceding one. The base classifier in each
application is assigned a weight value t according to its error
rate. The process is iteratively repeated until the error rate is
minimized. The ensemble consists of all base classifiers
generated thus in successive runs until the error rate minimum.
The boosted classifier, say H x , is obtained as weighted-sum
of all the base classifiers in the ensemble.
The steps of the AdaBoost algorithm are reproduced here
for the sake of clarity in the presentation of present analysis.
We consider the AdaBoost solution for a two-class problem.
The Base Classifier
In this work, we have taken the base learning algorithm to
be a simple threshold classifier which uses only one feature
dimension for class discrimination. The selection of that
particular feature dimension is made by searching over all the
feature dimensions while monitoring the rate of
misclassification according to certain threshold criterion. The
dimension with minimum rate of misclassification is selected
for defining the base classifier. The error rate for each
dimension is calculated as the sum of weights of all
misclassified training samples as defined in Step 2 above. The
threshold is a moving hyperplane which is perpendicular
bisector to feature components of consecutive training
samples. The selection of minimum error dimension with

COPO305-2
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

associated threshold definition and samples weight distribution

Feature Set 1
specify the base classifier in each round of the AdaBoost run.
In the present calculations we used the AdaBoost package Principal
available in MatLab [10]. Component
Analysis

Vector Autoscaling
III. MULTIPLE PREPROCESSORS AND INFORMATION FUSION Data Set 2

Fusion of information means to combine or assemble

Data Space Input Feature Space
information measures obtained from different sources, or Fusion Data Fusion
extracted by different methods from the same data source, or
both, in such a way that the fused information becomes Dimensional Autoscaling
Principal
potentially more efficient for target representation. In this Component
Data Set 1
work, we analyze the effects of information fusion on the Analysis
Principal
performance of AdaBoost algorithm as outlined schematically Component
in Fig. 1. The information fusion is implemented in two Feature Space 1
Analysis
Feature Space 2
alternate ways which have been referred to in the schematic as
Feature Set 2
data space fusion and feature space fusion. The two most
AdaBoosting 1 AdaBoosting 2
commonly used mean centering and data scaling methods
called the vector autoscaling and the dimensional autoscaling
have been used. These are defined as follows. Let M denote Error Minimization Error Minimization

the dimensionality of raw data, and N before the number of

training examples of samples. The training data set will thus be Fig. 1. Schematic of data preprocessing, feature extraction and adaboosting.
a N M matrix with samples in rows and variables in
columns.
dimensional autoscaled data separately. Then, the fusion is
Vector autoscaling done by a similar concatenation of the feature components as
The matrix elements are mean-centered and variance- explained above. The final feature spaces created by the
normalized for each dimension separately (column wise) as methods are of the same size, N 2M matrix with N training
xij xi examples with increased dimensionality equal to 2M .
xij
i

1 N 1 N
where xi xij and ( xij xi ) 2 . IV. VALIDATION
i
N j 1 N j 1 Four data sets of two class problems have used in the
validation analysis. These data were collected from the UCI
Dimensional autoscaling machine learning repository. When analyzed by a single strong
The matrix elements are mean-centered and variance classifier (backpropagation neural network) in combination
normalized for each sample separately (row wise) as with dimensional autoscaling and PCA feature extraction, the
xij x j classification rate for the two data sets (sonar and heart) were
xij typically <60% and that for the other two (Haberman breast
j cancer and Pima Indian diabetes) were typically >70%. The
1 M 1 M analysis of the same data sets were done by the proposed
where x j xij and j ( xij x j ) 2 . method consisted of the weak threshold classifier based
Mi1 M i 1 AdaBoost algorithm and the two methods of feature extraction.
The feature extraction has been done by the principal The division of the available data between the training and the
component analysis (PCA). test sets were done by random selection in nearly 50-50 ratio.
In data space fusion, the sample vectors transformed by the The description of the data sets is given Table 1.
two methods were fused by simple concatenation of the vector Table 2 presents the best classification results obtained by
components. That is, if the i-th training vector processed by the AdaBoost algorithm with the linear threshold base
the vector autoscaling is x1i {x1ij } {x1i1, x1i 2 ,..., x1iM } and that classifiers as described in the preceding. The data are
processed by four combinations of preprocessing, fusion and
processed by the dimensional auto scaling is
feature extraction strategies before AdaBoosting. These
xi2 {xij2} {xi21, xi22 ,..., xiM
2
} then the i-th fused data vector is combinations are: vector-autoscaling + PCA; dimensional-
autoscaling + PCA; (vector-autoscaling + dimensional-
defined by z i {x1i1, x1i 2 ,..., x1iM , xi21, xi22 ,..., xiM
2
} . The feature
autoscaling) data space fusion + PCA; and, (vector-autoscaling
extraction is done by PCA of the new N 2M dimensional + PCA) + (dimensional-autoscaling + PCA) feature space
data matrix. fusion. It can be seen that in all cases the performance of the
In feature space fusion, two alternate feature spaces are AdaBoost algorithm has improved after the multiple
created first by the PCA of the vector autoscaled and the

COPO305-3
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

preprocessor based data space or feature space fusions. The vector autoscaling dimensional autoscaling
amount of improvement however depends on the type of data. 1 1

In the Haberman breast cancer data and the pima-indian

error rate

error rate
diabetes data the improvements are very marginal. However, 0.5 0.5
in case of sonar data and heart data the improvements are
substantial after data space fusion.
0 0
20 40 60 80 20 40 60 80
TABLE I
DATA SETS USED IN PRESENT ANALYSIS data space fusion feature space fusion
1 1
Data Classes Sample Attribute Remark
s s

error rate

error rate
Sonar 2 208 60 Classes: sonar returns from a 0.5 0.5
metal cylinder and from a
similarly shaped rock.
Attributes: integrated energy
within a frequency band. 0 0
20 40 60 80 20 40 60 80
Heart 2 267 22 Classes: cardiac normal and X - axis: number of threshold classifiers in AdaBoost ensemble
abnormal condition.
Attributes: SPECT images. Fig. 1. Variation of error rate for sonar test data with ensemble size in
Haberman‟s 2 306 3 Classes: patients‟ survival after AdaBoost algorithm for linear threshold classifiers.
Breast-Cancer 5 years and death within 5 years
of breast cancer surgery. vector autoscaling dimensional autoscaling
1 1
Attributes: Age, year of
operation, positive axillary
nodes.

error rate

error rate
Pima-Indian 2 768 8 Classes: signs or no-sign of 0.5 0.5
Diabetes diabetes pima-indian females
above 21 years age.
Attributes: patient‟s history and 0 0
physiological parameters 20 40 60 80 20 40 60 80

TABLE II data space fusion feature space fusion

ADABOOST CLASSIFICATION RESULTS OBTAINED UNDER DIFFERENT 1 1
CONDITIONS OF PREPROCESSING AND FUSION
error rate

error rate
Classification Rate (%) 0.5 0.5

Data Data Feature

Vector Dimensional
Space Space
Autoscaling Autoscaling 0 0
Fusion Fusion 20 40 60 80 20 40 60 80
X - axis: number of threshold classifiers in AdaBoost ensemble
Sonar 53.66 59.76 75.61 53.66
Fig. 2. Variation of error rate for heart test data with ensemble size in
Heart 66.31 67.92 79.68 70.06 AdaBoost algorithm for linear threshold classifiers

Haberman vector autoscaling dimensional autoscaling

Breast- 75 71.7 76 75 1 1
Cancer
error rate

error rate

Pima- 0.5 0.5

Indian 66.7 77.7 76.7 78.4
Diabetes
0 0
20 40 60 80 20 40 60 80
The effect of present boosting strategy of the AdaBoost
data space fusion feature space fusion
algorithm can be seen from the error plots shown in Fig. 1 1 1
through Fig. 4 for the four sets of test data used in the analysis.
The error rate is defined as the fraction of misclassified test
error rate

error rate

samples, and the plots show the variation of error rate with the 0.5 0.5

number of base classifiers in the ensemble for AdaBoost. A

notable point from these plots is that the boosting strategy
0 0
yields good results only when the base classifier is weak. For 20 40 60 80 20 40 60 80

example, the selected threshold classifier results in the initial X - axis: number of threshold classifiers in AdaBoost ensemblle

error rates for the sonar and heart data close to 50%, and for
the breast-cancer and diabetes data close to 25%. The
AdaBoosting reduces error rate significantly in case of former,
Fig. 1 and Fig. 2, but not so much in case of latter, Fig. 3 and
Fig. 4. The similar impact on the classification rates is
Fig. 3. Variation of error rate for Haberman test data with ensemble size in
AdaBoost algorithm for linear threshold classifiers.
apparent from the results in Table II. Another notable point is
that the data space fusion facilitates better boosting.

COPO305-4
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
vector autoscaling
1 1
error rate
error rate
0.5 0.5
0 0
20 40 60 80
data space fusion
1 1
error rate
error rate
0.5 0.5
0 0
20 40 60 80
X - axis: number of threshold classifiers in AdaBoost ensemble
Fig. 4. Variation of error rate for pima-indian diabetes test data with
ensemble size in the AdaBoost algorithm for linear threshold classifiers.
V. DISCUSSION AND CONCLUSION
Sonar data is a complex distribution of energy of chirped
sonar signals in different frequency bands returned from two
types of targets (metallic cylinder and cylinder shaped rock).
All the attributes are therefore of the same kind. Besides, there
could be appreciable correlation between different attributes in
raw data. The best result for this data set is obtained by the
combination of data space fusion with AdaBoosting. The error
rate on the training data set drops to 0 after 20 rounds of base
learner creation. On the test data set however the error rate
continued to decrease up to 80 rounds, Fig. 1. The use of other
preprocessing methods did not produce much boosting for
ensemble sizes. The use of dimensional autoscaling and
feature space fusion reduced the error rate significantly after
only few rounds of iteration. Later, the error rate increased.
The vector autoscaling did not produce boosting effect in any
condition, Fig. 1.
Heart data is diagnostic cardiac data based on SPECT
(single proton emission computed tomography) images for
patients belonging to two categories: normal and abnormal.
The classification result in Table II and error result in Fig. 2
indicates a trend similar to that for sonar data. The
combination of data space fusion with AdaBoosting yields the
best result. The attributes of image data are also likely to be
correlated.
In contrast, the Haberman‟s survival data after breast cancer
surgery and the Pima Indian Diabetes data consist of patient‟s
history and physiological parameters like number of positive
axillary nodes and body glucose level. The variables in these
data sets are of different types and do not seem to have direct
correlation. The AdaBoosting in any combination of
preprocessing fusion strategy does not yield significantly
enhanced classification rate.
It appears therefore that the strategy of multiple
preprocessors based fusion of information enhances the
efficiency of the AdaBoost algorithm in those multivariate
situations where the variables are of the same type, and are
possibly correlated. The data preprocessing by the vector
dimensional autoscaling
autoscaling and the dimensional autoscaling followed by data
space fusion and feature extraction by principal component
analysis enhances the classification efficiency of the
AdaBoosted linear threshold classifier by 12% to 22% for the
sonar and heart data compared to the AdaBoosting without
fusion. We thus conclude that by bringing in diversity in the
preprocesssing methods for data representation to the feature
20 40 60 80
extractor yields more accurate feature set, which further
feature space fusion
enhances the efficiency of the AdaBoost algorithm.
ACKNOWLEDGMENT
This work is supported by the Government of India,
Defence Research & Development Organization Grant No.
20 40 60 80
ERIP-ER-0703643-01-1025. The authors are thankful to Mr.
Prashant Singh and Miss Prabha Verma for their support and
several useful discussions.
REFERENCES
.
[1] A. K. Jain, R. P. W. Duin, and J. Mao, “Statistical pattern recognition: a
review,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol.
22, no. 1, pp. 4-37, Jan. 2000.
[2] R. J. Schalkoff, Pattern Recognition – Statistical, Structural and
Neural Approaches, Wiley & Sons, 1992, ch. 1.
[3] T.G. Dietterich, “Ensemble learning,” in The Handbook of Brain
Theory and Neural Networks, 2nd ed. M.A. Arbib, Ed. Cambridge,
MA: The MIT Press, 2002, pp. 405-408.
[4] R. E. Schapire, “The strength of weak learnability,” Machine Learning,
vol. 5, no. 2, pp. 197-227, 1990.
[5] T. G. Dietterich, “Ensemble methods in machine learning,” in Lecture
Notes in Computer Science, J. Kittler, F. Roli, Eds. Berlin Heidelberg:
Springer-Verlag, vol. 1857, 2000. pp. 1–15.
[6] Y. Freund, Y. Mansour, R. Schapire, “Why averaging classifiers can
protect against overfitting,” in Artificial Intelligence and Statistics 2001
(Proc. of the Eighth International Workshop: January 4-7, 2001, Key
West, Florida), T. Jaakkola, T. Richardson, Eds. San Francisco CA:
Morgan Kaufmann Publishers, 2001.
[7] D. Chen, J. Liu, “Averaging weak classifiers,” in Lecture Notes in
Computer Science, J. Kittler and F. Roli, Eds. Berlin Heidelberg:
Springer-Verlag, vol. 2096, 2001, pp. 119-125.
[8] G. Levitin, “Threshold optimization for weighted voting classifiers,”
Naval Research Logistics, vol. 50, 2003, pp.322-344.
[9] Y. Freund, R. E. Schapire, “A decision-theoretic generalization of on-
line learning and an application to Boosting,” Journal of Computer and
System Sciences, vol. 55, no. 1, 1997, pp. 119–139.
[10] Cuneyt Mertayak (2007, may 25). AdaBoost, version 1.0. Available:
http://www.mathworks.com/matlabcentral/fileexchange/21317-
adaboost.
[11] R. G. Osuna, H. T. Nagle, “A method for evaluating data preprocesssing
techniques for odor classification with an array of gas sensors,” IEEE
Trans. Syst. Man Cybern. B, vol. 29, May 1999, pp. 626–632.
COPO305-5
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Real-time Hand Gesture Recognition Techniques for

Human-Computer Interaction
ANSHUL SHARMA
Department of Electronics and Communication Engineering, MNNIT Allahabad
Email ID: getanshulsharma@gmail.com

Abstract- This paper presents some real-time including the complicated nature of static and
video processing techniques for intelligent and dynamic hand gestures, complex backgrounds, and
efficient hand gesture recognition, with an aim occlusions. Attacking the problem in its generality
of establishing a virtual interfacing platform for requires elaborate algorithms requiring intensive
Human-Computer Interaction (HCI). The first computer resources. Due to real-time operational
step of the process is colour segmentation based requirements, we are interested in a
skin detection, followed by area-based noise computationally efficient algorithm.
filtering. If a gesture is detected, the next step is
to calculate a number of independent Previous approaches to the hand gesture
parameters of the available visual data and recognition techniques include the use of markers
assign a distinct range of values of each on various points on the hand, including fingertips.
parameter to a predefined set of different Calculation and observation of relative placement
gestures. The final step is the hierarchical and orientation of these markers specifies a
mapping of the obtained parameter values to particular gesture. The inconvenience of placing
recognise a particular gesture from the whole set. markers on the user’s hand makes this an infeasible
Deliberately, the mapping of gestures is not approach in practice. Another approach is to use
exhaustive, so as to prevent incorrect mapping sensor- fitted gloves to detect the orientation and
(misinterpretation) of any random gesture not other geometrical properties of the hand. The
belonging to the predefined set. The applications demerit of this approach is its cost ineffectiveness.
of the same are inclusive of, but not limited to,
Sign Language Recognition, robotics, computer The approach proposed in this text is quite user-
gaming etc. Also, the concept may be extended, friendly as it does not require any kind of markers
using the same parameters, to facial expression or special gloves for its operation. Also, the
recognition techniques. memory requirements are low because the
subsequent video frames are not stored in memory,
they are just processed and overwritten. Obviously,
1. INTRODUCTION it adds a new challenge to make the algorithm very
Gestures and gesture recognition are terms fast and efficient, fulfilment of which is ensured by
increasingly encountered in discussions of human- using low-complexity calculation techniques. For
computer interaction. The term includes character the ease of implementation, the proposed algorithm
recognition, the recognition of proof readers is based on three basic assumptions:
symbols, shorthand, etc. Every physical action
involves a gesture of some sort in order to be 1. The background should be dark.
articulated. Furthermore, the nature of that gesture 2. The hand should always be at a constant
is generally an important component in establishing distance from the camera.
the quality of feel to the action. The general 3. There should be a time gap of at least 200
problem is quite challenging due a number of issues ms between every two gestures.

COPO306-1
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
2. DATA EXTRACTION
The concept of video processing is based upon
data extraction from subsequent frames of the video.
The video is accessed frame by frame and the
following actions are performed on every
subsequent frame: (Some frames may be skipped,
without loss of information, because at an average
frame rate of 30 fps, any two subsequent frames
will have almost the same data.)
2.1 Hand Block Detection:
The first step of the gesture recognition process is
to detect if there is a gesture at all. It is done by skin
colour detection and area based detection of the
hand block. Also, the gesture should only be taken
into consideration if the whole hand is captured in
the video frame. This is achieved by calculating the
geometrical centroid of the hand block and
imposing a range restriction on the coordinates of
the centroid.
2.2 Elliptical Approximation:
The concept of elliptical approximation is the
heart of the gesture recognition process discussed in
this text. Once the hand block is detected, i.e. the
presence of a gesture (valid or invalid) is ensured, it
is approximated by an ellipse having the same
second-moments as the hand block detected.
Having obtained such a mathematical modelling,
we can interpret a number of independent
parameters from the geometrical properties of this
ellipse. This concept is depicted in Figure 1.
Figure 1: Elliptical approximation of hand block
The Orientation of the hand object is simply the
inclination of the major axis of the ellipse thus
obtained, from the positive horizontal axis.
Orientation is measured in a range of -90 to +90
degrees, as depicted in Figure 2.
Figure 2: Orientation of the hand block
The Eccentricity of the ellipse is simply the ratio
of semi-major axis (the distance between the two
foci) to the major axis. The value of eccentricity of
an ellipse lies between 0 and 1, both inclusive. An
ellipse with zero eccentricity is actually a circle.
The other extremity (eccentricity = 1) represents a
straight line. Thus, the eccentricity is a measure of
the narrowness of the ellipse. Higher be the
eccentricity, narrower is the ellipse (and vice versa).
2.3 Counting the number of fingers:
A very fast approach to count the number of
fingers in a gesture is discussed in this text, using
the data extracted by the means of elliptical
approximation as discussed already. With the vertex
of the ellipse as centre and a radius equal to 0.35
times the length of major axis, a circle is drawn.
The number of times the periphery of this circle
intersects any part of the hand block gives a direct
measure of the number of fingers in the gesture.
Figure 3 explains the concept. Clearly, if the circle
intersects the hand block at N places, then (N-1) is
the number of fingers in the gesture. As far as the
selection of the radius is concerned, 0.35 times the
major axis length was experimentally found to be
COPO306-2
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
the most optimum. A larger scaling factor (greater
than 0.4) entirely covers the smaller fingers and
hence, the presence of thumb and the little finger in
a gesture can’t be detected. On the other hand, a
small scaling factor (less than 0.3) is unable to
distinguish between two fingers separated by a
small gap, because the gap between the fingers is
lesser near the palm.
Figure 3: Counting the number of fingers
2.3 Perimeter Estimation:
Another distinguishing characteristic of a gesture
is the length of the periphery of the hand object.
Perimeter is calculated by simply counting the
number of pixels on the boundary of the hand
object.
Figure 4(a)
Figure 4(b)
Figure 4: Gestures having equal Areas but unequal Perimeters
Its significance can be appreciated by
acknowledging the fact that two gestures with same
area can have different perimeters. The hand
objects in Figure 4(a) and Figure 4(b) cover
approximately the same area, but the perimeter of
the hand object in Figure 4(a) is larger than that of
the hand object in Figure 4(b).
3. GESTURE RECOGNITION PROCESS
Having calculated the various parameters of the
hand block by mathematical modelling using
elliptical approximation, we now assign a unique
set of values of all these parameters to every gesture,
to be able to distinguish each one uniquely. Instead
of calculating all the parameters for every gesture, it
is efficient to categorize the gestures into different
hierarchical levels, so that only relevant properties
of every gesture are calculated, to make the
algorithm time-efficient. Statistically speaking, the
most optimum order is achieved by adopting the
following hierarchical recognition process:
1. Eccentricity. The first step is to mathematically
model the gesture by elliptical approximation and
calculate the eccentricity of the ellipse thus
obtained. If the eccentricity is too low (less than
0.5), it shows that the ellipse is wide enough
(approaching a circular shape), as is the case with
the gestures shown in Figure 5.
COPO306-3
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

three. This is obvious from Figure 4, which shows

one case in which perimeter estimation is required
as a distinguishing feature. Another similar case is
complimentary to these gestures, with negative
Orientation.

(a)

Figure 6: Gestures with Eccentricity > 0.95

4. Orientation. Orientation calculation is necessary

for every gesture, irrespective of the values of the
other parameters associated with it. This is because
the proposed gesture set is so made that every
gesture has a complimentary gesture whose all the
properties are the same as the former, with the only
difference being that its Orientation is the negative
(b) of that of former. Since Orientation calculation is a
Figure 5: Gestures with Eccentricity < 0.5 must for every gesture, this step may be executed at
any level, without any effect on the complexity of
On the other hand, if the eccentricity is too high the algorithm. Keeping it at the end of the process is
(greater than 0.95), it shows that the ellipse is very just one of the several equally efficient options.
narrow (almost a straight line), which is the case
with the gestures shown in Figure 6. In both these 4. THE ENTIRE GESTURE SET
cases, there is no need to count the number of
fingers, as it is obviously zero. The algorithm and the parameters already
discussed altogether can uniquely recognise a
2. Number of Fingers. If the value of Eccentricity gesture out of a closed set of 32 distinct gestures.
lies between 0.5 and 0.95 (both exclusive), the The complete gesture set is shown in Figure 7.
algorithm to calculate the number of fingers is
applied to the gesture frame. 5. APPLICATIONS
3. Perimeter Estimation. Perimeter estimation is Real-time gesture recognition is an emerging and
necessary as a distinguishing parameter only in a challenging field of Computer Vision. The main
few cases, when the number of fingers is less than applications of the same include, but are not limited
to:

COPO306-4
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
Figure 7: The Entire Gesture Set
Human- Computer Interaction (HCI).
Wireless robotic control mechanism.
Virtual Keyboard Interfacing Platform.
Security systems.
6. SCOPE FOR FUTURE WORK
The proposed gesture recognition techniques can
prove to be an integral basis for research work in
certain emerging fields, viz. Computer Vision, Real
Time Signal Processing, HCI, etc. Some major
fields that provide a challenging scope for the
future work of this text are discussed below:
Motion Gestures: So far we have discussed only
the recognition of static gestures. The above [3].Rafael C. Gonzalez, Richard Eugene Woods: “Digital
mentioned techniques, coupled with a concept Image Processing”, Third Edition, Prentice Hall Publications,
of motion recognition, can be used for
recognition of motion gestures.
No Background Limitation: The above
mentioned techniques are based on the
assumption that the background is always dark.
However, using dynamic colour segmentation
schemes, the same techniques can be applied
without this limitation also.
Double-Handed Gestures: All the same
concepts can be easily extended to gestures
involving both the hands at once. Doing so will
increase the size of the gesture set by about 5
times (with a substantial increase in complexity,
though).
Three- dimensional Gestures: The proposed
techniques deal with the recognition of only
two-dimensional gestures (hence the
assumption that the hand should always be at
constant distance from the camera, as discussed
earlier). However, if we use an additional
camera, orthogonal to the first one, we can
obtain the information about the distance of the
hand from the camera. This can be helpful for
efficient recognition of three-dimensional
gestures.
Facial Expression Recognition: The concept of
mathematical modelling of gestures using
Elliptical Approximation can be quite efficient
for the recognition of facial expressions.
Artificial Intelligence: Real-time Gesture
Recognition, coupled with a concept of Neural
Networks, may prove to be an integral
milestone in the field of Artificial Intelligence.
7. REFERENCES
[1]. Juan P. Wachs, Helman Stern, Yael Edan: “Cluster
Labeling and Parameter Estimation for the Automated Setup
of a Hand-Gesture Recognition System”, IEEE Transactions
on Systems, Man and Cybernetics—Part A: Systems and
Humans, Vol. 35, No. 6, November 2005.
[2]. Hyung-Ji Lee, Jae-Ho Chung: “Hand Gesture Recognition
Using Orientation Histogram”, 1999 IEEE Tencon.
2008.
COPO306-5
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Forensic Research by Face Reading Systems

Pramod Kumar Pal
Department of Computer Science & Engineering,
Radha Govind School of Engineering,
Radha Govind Group of Institutions
(U.P. Technical University)
Meerut-250004
Abstract: A contemporary definition for
The nature of computer vision causes the personality is offered by Carver and
fact that not only computer science Scheier: “Personality is a dynamic
researchers are interested in it, but organization, inside the person, of
neuroscientists and psychologists, too. psychophysical systems that create a
One of the main interests for psychology person’s characteristic patterns of
is identification of person’s behavior, thoughts, and feelings” [1]. In
psychological traits and personality other words, personality is a complex
types which can be accomplished by combination of traits and characteristics
different means of psychological testing: that determines our expectations, self-
questionnaires, interviews, direct perceptions, values and attitudes, and
observations, etc. Though that is a predicts our reactions to people, subjects
general tendency of people to read and events.
character into a person’s physical form, According to the Diagnostic and
especially face. In relation to Statistical Manual of the American
psychological characteristics Psychiatric Association [2], personality
recognition, face provides researchers traits are “enduring patterns of
and psychologists with instrument of perceiving, relating to, and thinking
obtaining information about personality about the environment and oneself that
and psychological traits that would be are exhibited in a wide range of social
much more objective than questionnaires and personal contexts.” Traits and
and neuropsychological tests and could characteristics are the same thing, as a
be obtained remotely using person’s trait is a distinguishing characteristic,
facial portrait, with no need for personal feature or quality. Theorists generally
involvement. The paper describes assume that a) traits are relatively stable
approaches to psychological over time, b) traits differ among
characteristics recognition from facial individuals, and c) traits influence
image such as physiognomy, phase behavior. Many psychologists have
facial portrait, ophthalmogeometry, and studied personality traits, and have
explains the need in automating it. attempted to identify or define all of the
Keywords: human traits. Some examples of
psychometrics, psychological personality traits identified and studied
characteristics, personality, Myers- by psychologists are: warmth, emotional
Briggs typology, ophthalmogeometry, stability, independence, dominance,
physiognomy, pattern recognition, face impulsivity, sensitivity, imagination,
recognition, facial expression introversion-extroversion,
recognition . suspiciousness, etc.
Introduction There is continuing debate about how
many different personality traits exist.

COPO307-1
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
The most popular model of
psychological traits is a five-dimension
personality model named as the “Big
Five” and proposed by Lewis Goldberg:
Extraversion, Neuroticism,
Agreeableness, Conscientiousness,
Openness to experience.
The difference between personality type
and personality trait should be
considered. Personality types are
distinguished from personality traits,
which come in different levels or
degrees. According to type theories, for
example, there are two types of people,
introverts and extraverts. According to
trait theories, introversion and
extraversion are part of a continuous
dimension, with many people in the
middle.
Identification of psychological
characteristics is the task widely used in
theoretical and practical psychological
research, education, coaching, career
guidance and hiring process, business
and political affairs, psychotherapeutic
diagnostics, self-exploration and
awareness, etc. Teachers evaluate
student’s personality traits and cognitive
abilities to know how to represent
material better and how to establish
communication and learning process in
more efficient way. Lawyers are
beginning to use personality testing for
criminal behavior analysis, litigation
pro- filing, witness examination and jury
selection. Medical stuff analyses
personality characteristics and observes
patient’s psychological state in regard of
its influence on medical treatment
process. Even websites design and
software interfaces are now developed
based on knowledge of user’s
personalities and their preferences.
It’s hard to underestimate the role of
identification of psychological
characteristics in modern society, as
everyone needs to know people’s
psychological traits to understand or
predict their reactions to various subjects
and situations.
2. Face analysis and other methods for
psychological characteristics
recognition
The question of methods and
instruments which may be used for
effective psychological characteristics
identification exists. To measure human
characteristics or identify personality
types psychological researchers apply
psychometrics which is the field of study
including the theory and technique of
psychological measurement primarily
concerned with the study of differences
between individuals. Psychometrics
involves such research tasks as the
construction of instruments and
procedures for measurement; and the
development of theoretical approaches to
measurement. For instance, the study
states that a useful psychological
measure must be both valid (i.e., actually
measures what it claims to measure) and
reliable (i.e., internally consistent or give
consistent results over time).
The most accepted means of measuring
personality using psychometrics study
is psychological questionnaires and
tasks. There are two major types of such
personality tests. Projective tests assume
that personality is primarily unconscious
and assess an individual by how he or
she responds to an ambiguous stimulus.
The idea is that unconscious needs will
come out in the person’s response.
Objective tests assume that personality is
consciously accessible and measure it by
self-report questionnaires. Research on
psychological assessment has generally
found that objective tests are more valid
and reliable than projective tests.
Examples of testing questionnaires and
tasks are Holland Codes, Rorschach test,
COPO307-2
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
Minnesota Multiphasic Personality
Inventory, Myers-Briggs Type Indicator,
Enneagram Type Indicator, NEO PI-R,
Thematic Apperception Test, Kelly’s
Repertory Grid, Online Depression
Screening Test, Online Screening for
Anxiety and so on [3].
Except psychological questionnaires and
tasks, interviews and direct observation
tests are applied. Interviews often
contain some observation and questions
about person’s past, abilities and
attitudes. Direct observation of people as
they complete activities may be with
clinical purpose, such as to observe the
nature of a parent-child interaction in
order to understand a relational disorder.
The Parent-Child Interaction
Assessment-II [4] is an example of a
direct observation procedure that is used
with school-age children and parents: the
parents and children are videotaped
playing at a make-believe zoo. Direct
observation procedures are also used in
research, for example to explore
sequences of behavioral interaction.
There is a range of approaches to
measuring personality and psychological
characteristics which are based on
human physical appearance. They are
described below:
Evaluation based on face
The face plays a crucial role in human
social cognition. Facial expressions
are important signals of internal states –
emotions and intentions. Humans also
see in the face signals of internal
qualities that are more stable over time,
like attractiveness as a mate [5,6,7] or
dominance [8,9,10], etc. According to
Liggett [11], “There can be little doubt
that the face plays a crucial part in our
everyday assessment of our fellows. Not
only does it enable us to identify
transient emotions – flashes of pleasure
and rage, disappointment and hatred – it
can also help us to make useful
judgments about more durable and
lasting qualities of personality and
character”. Judging personality from
face comes from ancient times [12,13] to
nowadays. The belief that the face
reveals information about underlying
character cuts across national, cultural
and geographical boundaries. Several
classical Greek and Roman scholars,
including Plato and Aristotle, argued that
faces contain clues about people’s
underlying personalities and
dispositions. For over 2500 years many
notable philosophers, historians and
physicians openly supported the idea
that a person’s disposition is reflected in
their face.Widespread interest in
physiognomy – the study of the face and
its relationship to human ability,
potential and character
– peaked at the end of the eighteenth
century when a physician and pastor
named Johann Kaspar Lavater produced
a formal classification system and set of
rules specifying the relationship between
the face and the mind [14]. Lavater
believed that character is assessed just as
well from the shape and size of the
forehead, for example, as from
observations of behavior. Indeed,
eighteenth and nineteenth century courts
commonly used physiognomy to assess
immoral tendencies in suspected
criminals.
Among contemporary applications of
face analysis not only psychological
interpretations are accomplished, but
also medical results are obtained. For
example, more than 700 genetic issues
influencing facial structure and facial
features are known, and special software
for disease identification from face had
been developed [15,16].
Body types
COPO307-3
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

William Sheldon classified personality type of mimics and voice may define
according to body type [17]. He called personality traits. For example, it’s used
this a person’s somatotype and identified in socionics (see Table 2) that is a
three main somatotypes shown in branch of psychology based on Carl
Table 1. Sheldon’s somatotypes and Jung’s work on Psychological Types.
character interpretations Moreover, many socionics experts use
the visual method of personality
Sheldon’s Character Shape
Somatotype
characteristics identification as a main
method for personality traits and types
Endomorph Relaxed, sociable, Plump, recognition.
[viscerotonic] tolerant, buxom,
comfort-loving, developed
Table 2. Example of some outer
peaceful visceral appearance characteristics and their
structure interpretation
Mesomorph Active, assertive, Muscular
[somatotonic] vigorous,
OUTER APPEARANCE
combative No Physical Sensoring Intuitive
characte
Ectomorph Quiet, fragile, Lean,
[cerebrotonic] restrained, delicate,
r
non-assertive, poor 01 The Short and Lengthy
sensitive muscles form of thick, and thin,
bones muscles are muscles
Person is rated on each of these three and pronounced aren’t
dimensions using a scale from 1 (low) muscles pronounce
to 7 (high) with a mean of 4 (average). d
Therefore, for example, a person who is 02 Form of Sen Sen Intu Intui
a pure mesomorph would have a score of the nose sori sori itive tive
1-7-1. ng ng + +
In Ayurvedic medicine (used in India + + Ethi Logi
since ˜3000 BC) there are three main Log Eth cal cal
metabolic body types (doshas) – Vata, ical ical
Pita, & Kapha – which in some way «triangle
correspond to Sheldon’s somatotypes. Horizonta with peak
Body types have been criticized for very l line in on the top»
weak empirical methodology and are not the nose «triangle
generally used in Western psychology bridge with peak
(they are used more often in alternative in the
therapies and Eastern psychology and bottom»
spirituality).
Complex physical appearance Neuropsychological tests
evaluation Around the 1990s, neuroscience entered
This is approach of evaluation of face the domain of personality psychology.
and body parts in complex, and it is It introduced powerful brain analysis
considered to be physiognomy too. tools like Electroencephalography
Physical appearance characteristics such (EEG), Positron Emission Tomography
as appearance of some facial features, of (PET), Functional Magnetic Resonance
the skull, shoulders, hands, fingers, legs, Imaging (fMRI) and structural MRI
including diffusion tensor imaging (DTI)

COPO307-4
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

to this study. One of the founders of this are often able to distort their responses.
area of brain research is Richard This is particularly problematic in
Davidson of the University of employment contexts and other contexts
Wisconsin-Madison [18]. Davidson’s where important decisions are being
research lab has focused on the role of made and there is an incentive to present
the prefrontal cortex and amygdala in oneself in a favorable manner. Social
manifesting human personality. In desirability is a tendency to portray self
particular, this research has looked at in a positive light, and faking bad also
hemispheric asymmetry of activity in happens, that is purposely saying ’no’ or
these regions. Neuropsychological looking bad if there’s a ’reward’ (e.g.
studies have illustrated how hemispheric attention, compensation, social welfare,
asymmetry can affect an individual’s etc.). Work in experimental settings
personality. [20,21] has shown that when student
In contemporary psychological research samples have been asked to deliberately
there should be an instrument which fake on a personality test, they
would provide a maximum amount and demonstrated that they are capable of
type of objective/unbiassed information doing this.
about personality in as short a time as Though several strategies have been
possible, preferably with no participation adopted for reducing respondent faking,
of person whose characteristics are this is still a problem for such traditional
identified. Comparison of approaches to psychological testing instruments like
identi- fication of psychological questionnaires, interviews, direct
characteristics described above is observations. Surprisingly,
represented in Table 3. neuropsychological tests are prone to
Some comparison of approaches to respondent faking, too [22,23]. Faking
identification of psychological response styles include faking bad
characteristics (malingering), faking good
Criterion Psy Inte Face, Neurops (defensiveness), attempts at invalidation,
chol rvie body ycholog
ogic w, evalua ical
mixed responding (faking good and
al dire tion tests bad), and a fluctuating, changing style
que ct that occurs within one evaluation
stio obs
nnai erva
session. These response styles lead to
res tion getting incorrect results.
Easy and not – – + – Concerning face and facial features,
time-consuming
for person who is
faking becomes much more complicated:
tested it’s impossible to change the shape of a
Person may not – – + – nose or cheekbones just when person
participate in
testing process
wants. Besides, it is often unknown to a
High validity and + – ? – holder what his/her face reveals exactly.
reliability [19] Theoretically people can “fake” facial
Practically no – – + – features intentionally changing their
possibility for
respondent faking shape, color, texture, for instance, using
No need in + + + – plastic surgery, and identifying personal
expensive hi–tech psychological characteristics becomes
hardware
much harder in this case, though it may
In psychological testing there is
be also accomplished.
considerable problem that respondents

COPO307-5
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
Face is the first subject that is unique for
people and used for people recognition.
Thus, face is the most available means of
evaluation among other instruments
based on questionnaires, interviews,
neuropsychological tests. People in
general may not participate in testing
process, identification of personality
characteristics may be done remotely,
even by exterior parties.
Summarizing, face provides researchers
and psychologists with instrument of
obtaining information about personality
and psychological traits that would be
much more objective than questionnaires
and neuropsychological tests (as we
can’s change facial features just when
such desire appears) and could be
obtained remotely using person’s facial
portrait, with no need for personal
involvement.
If such instrument is working
automatically (system gets facial
portrait, processes it and in result gives
out information about personality
characteristics) and has straight-forward
layout, then: 1) psychological testing
becomes more accurate, fast, objective
and available for different kinds of
research and applications; 2) deep
knowledge in interpretation of facial
features, which is rather rare in modern
society, isn’t needed to administer and
use the instrument. Methods and
algorithms originally developed for face
detection, face recognition and facial
expression recognition research fields as
well as contemporary trends (applying
standard face images, multimodality,
three-dimensionality) should be applied
and adjusted to so-called Automatic
Psychological Characteristics
Recognition from Face. From its side,
Automatic Recognition of Psychological
Characteristics from Face is believed to
bring scientific benefits to face
recognition, facial expression
recognition, face animation, face
retrieval, etc., and finally contribute to
development of human-computer
interaction on higher level. Thus, the
relations between such research areas as
face recognition, facial expression
recognition and psychological
characteristics recognition are mutually
beneficial.
3. Approaches to psychological
characteristics recognition from face
There are three main approaches to
psychological characteristics recognition
from face: physiognomy, phase facial
portrait and ophthalmogeometry, see
Fig.1. The first originally interprets
different facial features, the second
works with angles of facial features and
facial asymmetry, and the third extracts
and interprets eye region parameters.
Methods developed for these approaches
are described below. Physiognomy is a
theory based upon the idea that the
assessment of the person’s outer
appearance, primarily the face, facial
features, skin texture and quality, may
give insights into one’s character or
personality. Physiognomy has flourished
since the time of the Greeks
(Empedocles, Socrates, Hippocrates and
Aristotle), amongst the Chinese and
Indians, with the Romans (Polemon and
Adamantius), in the Arab world
(including Avicenna), and during the
European renaissance (Gerolamo
Cardano and Giovanni Battista della
Porta). It faded in
Figure 1. Approaches to psychological characteristics
recognition from facial portrait
COPO307-6
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
popularity during the 18th century, was
eclipsed by phrenology in the 19th and
has been refreshed by personologists in
the 20th century.
During 20th century attempts had been
made to perform scientific experiments
concerning validity of different facial
features interpretations and high
accuracy results had been claimed [24],
though they are mostly aren’t accepted
by official science [25]. At the same
time, science step by steps proves some
physiognomy beliefs. For instance,
correlations have been established
between IQ and cranial volume
[26,27,28,29]. Testosterone levels,
which are known to correlate with
aggressiveness, are also strongly
correlated with features such as finger-
length ratios and square jaws [30,31].
Interpretation of facial features
based on physiognomy has been
implemented into psychological
characteristics diagnosis tools such as
“Visage” Project [32] developed by Dr.
Paul Ekman and “Digital physiognomy“
software [33] developed by Uniphiz Lab.
“Visage” is a project for collecting
and organizing information about
relatively permanent facial features. It
includes methods for storing, retrieving,
and inspecting the data. Visage is a
unique database schema for representing
physiognomy and the interpretation of
physiognomic signs. The Visage
demonstration application illustrates
limited variations of some facial features
in the following categories: forehead and
eyebrows (see the Fig.2), eyes and
eyelids, nose, mouth and jaw, cheeks,
chin, ears. User should select features
that are distinctive about the face that is
going to be interpreted and then click the
“Get...” button. The application retrieves
information from the database relevant
to description of physiognomy,
including an estimation of the accuracy
of the sources of information.
“Digital physiognomy“ software
determines a person’s psychological
characteristics based on temperament
types, intellect, optimism – pessimism,
conformism – adventurism, egoism –
altruism, philanthropy – hostility,
laziness, honesty, etc.
Figure 2. Example of the table and interface of Visage
demonstration application: facial features in the
forehead and eyebrow area [34]
and then presents a detailed person’s
character analysis in a graphic format.
The tool works like a police sketch
(photo robot), so user has to select
different parts of the person’s face, and
doesn’t need to have a person’s
photograph, see Fig. 3. It’s claimed that
only the facial features that can be
interpreted with high accuracy were
used, and the confidence factor is
calculated for each interpretation by the
tool. It should be noted that “Digital
physiognomy“ tool also uses visual
systematic classification of 16
personality types based upon Myers-
Briggs typology, see Fig. 4.
“Visage” and “Digital Physiognomy”
projects are some of the first attempts
to develop physiognomic database and
use modern technology for
physiognomic interpretations. In spite of
having value for psychological diagnosis
based on physiognomy, both projects use
manual selection of facial features, and
thus, can’t be used extensively and
applied in scientific research.
COPO307-7
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Phase facial portrait approach to

psychological characteristics recognition
from facial portrait is primarily based on
calculating of angles of facial features
lines directions, as shown on Fig.5.
Video-computer psychological diagnosis
and correction method [35] had been
invented by professor Avtandil
Anuashvili. It’s remote method Figure 4. Socionics visual interpretation by “Digital
physiognomy“ based upon Myers-Briggs typology
for personal psycho-type identification,
method of one of the brain hemispheres
dominance identification, method of
psycho-diagnostics and psycho-
correction. The method is based on the Figure 5. The representation of phase facial portrait
concept [37]
thesis that face is an informational
these oscillatory processes among
background projecting in biological and
themselves [36]. Based on these
psychological terms the results of joint
meanings, the software classifies a given
brain hemispheres functioning. The
person to one of 49 psychological types
system of 49 types of people in
and gives out complete personal
coordinates “left-side thinker –
characteristic, professional characteristic
right-side thinker”, “psychologically
and recommendations on harmonization,
stable personality – instable personality”
effective interaction with other people
had been created and correlated to other
and environment.
personal typologies, see Fig. 6.
Video-computer psychological diagnosis
Software developed to apply video-
and correction method is one of the
computer psychological diagnosis and
primary instrumental psychological
correction method accepts facial image
methods concerning examination and
and determines: 1) dominance of one of
usage
the brain hemispheres on the basis of a
difference of amplitudes of oscillatory
processes occurring in right and left
hemispheres and 2) degrees of a
coordination of

Figure 6. Table of 49 psychological types system

invented used for video-computer psychological
diagnosis and correction; here L – logic (practical
mind), I – intuition (spirit, principles), S – stability, D
– destability [38]
Figure 3. Fragment of “Digital physiognomy“
software

COPO307-8
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
Figure 7. Translated picture from Muldashev’s book
[39]: here two parameters of facial eye region are used
for recognition of some basic psychological traits, e.g.
strong will and fearfulness, etc.
Figure 8. Ophthalmogeometrical pattern extraction 40]
of brain asymmetry phenomena and face
asymmetry. Although Anuashvili claims
that application developed for video-
computer psychological diagnosis and
correction method is entirely automated,
practically it may be considered to be
semi-automated as manual selection of
facial points on image is required. This
limits usage of such application for
extensive research and other purposes.
Concerning ophthalmogeometry
approach, it is based on idea that
person’s emotional, physical and
psychological states can be recognized
by 22 parameters of an eyes part of the
face [39], see Fig. 7. phthalmogeometry
phenomenon has been discovered by
prof. Ernst Muldashev. Apart from other
interesting facts, E. Muldashev has
found that in 4-5 years after birth the
only practically constant parameter of
human body is the diameter of the
transparent part of cornea which equals
10±0,56 mm. He also represented an
idea that ophthalmogeometrical pattern
is unique for people. The procedure of
this pattern identification and calculation
is described by Leonid Kompanets [40],
see Fig. 8.
Ophthalmogeometry is based on
interesting ideas and may be applied to
psychological, medical research as well
as to biometrics, though this is not very
deeply investigated area of facial
analysis which primarily needs
automation of ophthalmogeometric
pattern extraction and further
investigation.
4. Conclusion
The paper represents general idea that
face provides researchers and
psychologists with objective instrument
of obtaining information about
personality and psychological traits. An
up-to-date survey of approaches and
methods in psychological characteristics
recognition from facial image is
provided.
In perspective new research task of
automating procedures in applications of
psychological characteristics recognition
from face should be explored. Various
approaches and methods developed
within face recognition, facial
expression recognition, face retrieval,
face modeling and animation may be
applied and adjusted for recognition of
psychological characteristics from face.
Undeniably, such automated system of
psychological characteristics recognition
from face will get countless
psychological, educational, business
applications. It may be used also as part
of medical systems: 1) patient’s
psychological state and traits influence
the process of medical treatment, and it
should be taken into consideration and
researched; 2) patient’s psychological
characteristics should be taken into
account to reflect and construct the
psychosomatic model of disease in the
environment, which includes biological,
psychological, and social factors.
References
[1] Carver C. S., Scheier M. F. Perspectives on
personality (4th ed.) Boston: Allyn and
Bacon, 2000, page 5.
[2] DSM, Diagnostic and Statistical Manual of Mental
Disorders, can be found at
http://www.psych.org/ research/ .
[3] Hampson S. E. Advances in Personality
Psychology. Psychology Press, 2000.
[4] Holigrocki R. J., Kaminski P. L., Frieswyk S. H.
(2002). PCIA-II: Parent-Child Interaction
COPO307-9
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
Assessment Version II. Unpublished manuscript,
University of Indianapolis.
(Update of PCIA Tech. Rep. No. 99-1046. Topeka,
KS: Child and Family Center,
The Menninger Clinic.) (Available from Dr. Richard J.
Holigrocki or Dr. Patricia
L. Kaminski).
[5] Nigel Barber. The evolutionary psychology of
physical attractiveness: Sexual selection
and human morphology. Ethology and Sociobiology,
Volume 16, Issue 5, September
1995, pages 395-424.
[6] John P. Swaddle, Innes C. Cuthill. Asymmetry and
Human Facial Attractiveness:
Symmetry May not Always be Beautiful. Proceedings:
Biological Sciences, Vol. 261,
No. 1360 (Jul. 22, 1995), pages 111-116.
[7] Thomas R. Alley, Michael R. Cunningham.
Averaged faces are attractive, but very
attractive faces are not average. Psychological Science
2 (2), 1991, pages 123-125.
[8] Leslie A. Zebrowitz, Gillian Rhodes. Sensitivity to
”Bad Genes” and the Anomalous
Face Overgeneralization Effect: Cue Validity, Cue
Utilization, and Accuracy in Jud
72 Ekaterina Kamenskaya, Georgy Kukharev
ging Intelligence and Health. Journal of Nonverbal
Behavior Volume 28, Number 3
/ September, 2004, pages 167-185.
[9] Caroline F. Keating. Gender and the Physiognomy
of Dominance and Attractiveness,
Social Psychology Quarterly, Vol. 48, No. 1 (Mar.,
1985), pages 61-70.
[10] Ulrich Mueller, Allan Mazur. Facial Dominance
of West Point Cadets as a Predictor
of Later Military Rank. Social Forces, Vol. 74, No. 3
(Mar., 1996), pages 823-850.
[11] J. Liggett, The human face. New York: Stein and
Day, 1974, page 276.
[12] Phisiognomics, attributed to Aristotle. Cited in
J.Wechsler (1982), A human comedy:
Physiognomy and caricature in 19th century Paris
(p.15). Chicago: University
of Chicago Press.
[13] A. Brandt. Face reading: The persistence of
physiognomy. Journal Psychology Today,
1980, December, page 93.
[14] Sibylle Erle. Face to Face with Johann Caspar
Lavater, Literature Compass 2 (2005)
RO 131, pages 1 -4.
[15] Stefan Boehringer, Tobias Vollmar, Christiane
Tasse, Rolf P Wurtz, Gabriele
Gillessen-Kaesbach, Bernhard Horsthemke and
Dagmar Wieczorek. Syndrome identification
based on 2D analysis software. European Journal of
Human Genetics
(2006), pages 1-8.
[16] Hartmut S Loos, Dagmar Wieczorek, Rolf
PWürtz, Christoph von der Malsburg and
Bernhard Horsthemke. Computer-based recognition of
dysmorphic faces, European
Journal of Human Genetics (2003) 11, pages 555-560.
[17] Irvin L. Child, William H. Sheldon. The
correlation between components of physique
and scores on certain psychological tests. Journal of
Personality, Vol. 10, Issue 1,
September 1941, page 23.
[18] Richard Davidson, Ph.D. Vilas Professor of
Psychology and Psychiatry. Can be
found at https:// psychiatry.wisc.edu/
faculty/FacultyPages/Davidson.htm.
[19] The Validity of Graphology in Personnel
Assessment. Psychological Testing Centre.
Found at www.psychtesting.org.uk, November 1993
reviewed April 2002.
[20] Chockalingam Viswesvaran, Deniz S. Ones.
Meta-Analyses of Fakability Estimates:
Implications for Personality Measurement,
Educational and Psychological Measurement,
Vol. 59, No. 2, 1999, pages 197-210.
[21] Deniz S. Ones, Chockalingam Viswesvaran,
Angelika D. Reiss. Role of Social Desirability
in Personality Testing for Personnel Selection: The
Red Herring. Journal
of Applied Psychology, 1996. Vol. 81, No. 6, pages
660-679.
[22] Hall, Harold V.; Poirier, Joseph G.; Thompson,
Jane S. Detecting deception in
neuropsychological cases: toward an applied model.
From: The Forensic Examiner,
9/22/2007.
[23] Allyson G. Harrison, Melanie J. Edwards and
Kevin C.H. Parker. Identifying students
faking ADHD: Preliminary findings and strategies for
detection. Archives of
Clinical Neuropsychology. Volume 22, Issue 5, June
2007, pages 577-588
[24] Naomi Tickle. You Can Read a Face Like a
Book: How Reading Faces Helps You
Succeed in Business and Relationships, Daniels
Publishing, 2003.
[25] Robert Todd Carroll. The Skeptic’s Dictionary: A
Collection of Strange Beliefs,
Amusing Deceptions, and Dangerous
Delusions.Wiley; 1st edition (August 15, 2003)
[26] J. Philippe Rushton, C. Davison Ankney. Brain
size and cognitive ability: Correlations
with age, sex, social class, and race. Psychonomic
Bulletin & Review, 1996, 3
(1), pages 21-36.
[27] Michael A. McDaniel. Big-brained people are
smarter: A meta-analysis of the re
Recognition of Psychological Characteristics from
Face 73
lationship between in vivo brain volume and
intelligence. Intelligence, Volume 33,
Issue 4, July-August 2005, pages 337-346.
[28] J. Philippe Rushton. Cranial size and IQ in Asian
Americans from birth to age
seven. Intelligence, Volume 25, Issue 1, 1997, pages
7-20.
[29] John C. Wickett, Philip A. Vernon, Donald H.
Lee. Relationships between factors of
COPO307-10
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

intelligence and brain volume. Personality and

Individual Differences, Volume 29,
Issue 6, December 2000, pages 1095-1122.
[30] John T. Manning. Digit Ratio: A Pointer to
Fertility, Behavior, and Health. Rutgers
University Press, 2002.
[31] Bernhard Fink, Karl Grammer, Philipp
Mitteroecker, Philipp Gunz, Katrin Schaefer,
Fred L. Bookstein, John T. Manning. Second to fourth
digit ratio and face
shape. Proceedings of the royal society, Volume 272,
Number 1576 / October 07,
2005, pages 1995-2001.
[32] Visage Project for Physiognomy Data, can be
found at
http:// face-and-emotion.com/ dataface/ visage/
about\_visage.jsp. Date: 20
January 2008.
[33] Digital Physiognomy Software: Match person’s
face to his or her character. Can be
found at http://www.uniphiz.com/ physiognomy.htm.
Date: 20 January 2008.
[34] Facial Features in the Forehead & Eyebrows
Area, can be found at
http:// face-and-emotion.com/ dataface/ visage/
visage\_forehead.jsp. Date: 20 January
2008.
[35] Anuashvili Avtandil. Fundamentals of
Psychology. Scientific, Philosophic and Spiritual
Fundamentals of Psychology. The Institute for Control
Problems Press, Moscow,
2001 (In Russian).
[36] Video-computer system for psychodiagnosis and
psychocorrection is the
first objective tool for psychotherapist and
psychologist, can be found at
http://www.anuashvili.ru/Video-
computer\%20psychodiagnostics/ video-compute.
htm. Date: 20 January 2008.
[37] The audio-presentation for video-computer
psychological diagnosis and correction,
can be found at http://www.tsarvgolove.ru (in
Russian). Date: 20 January 2008.
[38] Avtandil Anuashvili’s website, can be found at
http://www.anuashvili.ru/ . Date:
20 January 2008.
[39] Muldashev Ernst R. Whom did we descend
from?, “OLMA-PRESS”, Moscow 2002
(In Russian)
[40] Leonid Kompanets. Biometrics of asymmetrical
face. From Biometric Authentication:
First International Conference, By David Zhang, Anil
K. Jain, ICBA 2004 ,
pages 67-73.

COPO307-11
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
An Automatic Machine Learning and Particle Filtering
Based Approach to Real Time Human Tracking in Videos
Chandra Mani Sharma1, Alok Kumar Singh Kushwaha1, Ashish Khare2 and Sanjay Tanwani1
1
School of Computer Science and IT, DAVV Indore. 2 Department of Electronics and Communication, University of Allahabad, Allahabad.
chandramani.sharma@yahoo.com, alokkumarsingh.jk@gmail.com, akhare@allduniv.ac.in,
sanjay_tanwani@hotmail.com
Abstract:
Object Tracking is an important task in video
processing because of its various applications like
visual surveillance, human activity monitoring and
recognition, traffic flow management etc. Multiple
object detection and tracking in outdoor environment
is a challenging task because of the problems raised
by poor lighting conditions, occlusion and clutter.
This paper proposes a noble technique for detecting
and tracking the multiple humans in a video. A
classifier is trained for object detection using haar-
like features from the training image set. The human
objects are detected with the help of this trained
detector and are tracked with the help of a particle
filter. The experimental results show that the propose
technique can detect and track the multiple humans in
a video adequately fast in the presence poor lighting
conditions, clutter and partial occlusion and the
technique can handle varying number of human
objects in the video at various points of time.
Keywords: Human detection, Automatic multiple
object tracking, Haar-like features, Machine learning,
Particle filter.
1. Introduction:
Detecting humans and analyzing their activities by
vision is key for a machine to interact intelligently
and effortlessly with a human inhabited
environment .The aim of visual surveillance is the
real-time observation of targets such as human
beings or vehicles in some environment which leads
to the description of objects’ activities with the
environment or within them. Visual surveillance has
been used for security monitoring, anomaly
detection, intruder detection, traffic flow measuring,
accident detection on highways, and routine
maintenance in nuclear facilities[1,2,3,4]. Hu et.
al.[1], provide a good survey on visual surveillance
and its various applications. Detecting and Tracking
humans in a video is a step to take in the process of
analyzing and predicting their behavior and
intention. This is a good practice to detect objects in
a video sequence before tracking them. The problem
of multiple object tracking is more complex and
challenging than the single object tracking because
of the issue of the management of multiple tracks
caused by newly appearing objects and the
disappearance of already existing targets. Viola et
al[5] proposed an object detection framework for
and used it for first time for detecting the human
faces in an image. The adaptive boosting can be
used to speed up a binary classifier [6] and this can
be used in machine learning for creating the real-
time detectors. After detecting an object or multiple
objects in a video sequence captured by the
surveillance camera, the very next step is to track
these objects (human, vehicle etc.) in the subsequent
frames of the video stream. The particle filter based
tracking techniques are gaining popularity because
of their ease of implementation and capability to
represent a non-linear object tracking system and
non-Gaussian nature of noise. Various object
tracking techniques based on particle filtering are
found in literature [7,8,9,10,11]. These approaches
fall under two main categories; single object
trackers and multiple object trackers. Lanvin et.
al.[7], propose an object detection and tracking
technique and solve the non-linear state equations
using particle filtering. Single object trackers suffer
from the problem of false positives when severe
COPO4O1-1
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
occlusions occur because of hidden first order
Markov hypotheses [8]. The problem of tracking
multiple objects using particle filters can be solved
in two ways. One is by creating multiple particle
filters for each track and another one is by having a
single particle filter for all tracks. The second
approach works fine as long as the objects under
tracking are not occluded but in case of occlusion
when objects come close by, these techniques fail to
track the objects. In [8] a multi-object tracking
technique using multiple particles has been
proposed. Algorithms which attempt to find the
target of interest without using segmentation have
been proposed for single target tracking based on
cues such as color, edges and textures [12].
Chen et. al.[9] propose a color based particle filter
for object tracking. Their technique is based on
Markov Chain Monte Carlo(MCMC) particle filter
and object color distribution. Sequential Monte Carlo
techniques also known as particle filtering and
condensation algorithms and their applications in the
specific context of visual tracking, have been
described in length in the literature[13,14,15]. In
[10], object tracking and classifications are
performed simultaneously. Many of the particle filter
based multiple object tracking schemes rely on
hybrid sequential state estimation.
The particle filter developed in[16] has multiple
models for the objects motion, and comprises an
additional discrete state component to denote which
of the motion models is active. The Bayesian
multiple-blob tracker[17] presents a multiple tracking
system based on statistical appearance model. The
multiple blob tracking is managed by incorporating
the number of objects present in the state vector and
the state vector is augmented as in [18] when a new
object enters the scene. The joint task of object
detection and tracking comes with a heavy
computational overhead but for visual surveillance
applications the speed is one of the most important
factors. Many previously proposed techniques are
slow in speed so cannot be good for visual
surveillance and some fail to cope with dynamic
outdoor environment conditions. This paper presents
a noble technique based on machine learning and
particle filtering to detect and track the humans in a
video. The humans are detected in video frames using
the object detection framework proposed by Viola
et.al.[5] using Haar-like features and then tracked
using a simple particle filter in the subsequent video
frames. Binary adaptive boosting [6] reduces the
training time of the human detector. The earlier
detection of the objects simplifies the process of
tracking and sheds the load of detection from tracker.
The exhaustive dataset used for training the detector
makes the system detect and track the multiple
objects in critical lighting conditions, in dynamic
background, partial occlusion and clutter.
The rest of the paper is organized as follows;
section(2) discusses the proposed technique for
human detection and tracking. In section (3) the
various experimental results using the proposed
technique are given which prove the validity and
novelty of the method. Section(4) comprises of the
conclusions and at last references are given.
2. Methodology
Here we solve two sub-problems. One is to detect
the humans in the video and another one is to track
them in the subsequent video frames. The sub
problem of object detection is solved using
machine learning approach for which we train our
human detector using binary adaptive boosting.
Algorithm for multi human detection and
tracking :-
 Let Z be the input video to the algorithm.
 In first frame  0
of Z, detect humans using
the human detector. Let  be the number
of detected humans.
 Initialize trajectories  j
, 1≤ j≤  with
initial positions  j,0
.
 Initialize the appearance model (color
histogram)  j
for each trajectory from
the region around  j,0
.

 For each subsequent frame of input
i
video,
(a) For each existing trajectory  j
,
COPO4O1-2
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

i. Use motion model to predict the
distribution  (
over locations for human j in
frame i, creating a set of
candidate particles
≤ K.
ii. Compute the color histogram
speed up the process. We collected positive and
 │
j ,i
),
j , i 1
negative image samples for training. The
positive images are those comprising the human
beings and negative images do not contain any
human being. We cropped the humans from the

(k )
j ,i
, 1 ≤k positive images and resized to the dimensions
of 40*40. Our positive dataset consisted of
2,000 images while the negative dataset


(k )
and likelihood consisted of 2,700 images. Fig.1 consists of
some cropped human images from positive
 ( │  , )
k (k )
for each samples used for human detector training. This
j ,i j human detector is nothing but a binary classifier
particle k using the appearance having two classes: human and non-human.
model. Next step after the sample collection is Haar-
iii. Resample the particles according to
their likelihood. Let k* be the like feature extraction from these samples. We
index of the most likely use here the rectangular Haar-like features and
particle. these features have their intuitive similarity to
iv. Perform confirmation by the Haar-wavelets. Integral Image
classification: run the Humans representation is used for fast feature evaluation
detector on the location (see fig. 3).


*
( k )
. If the location is
j ,i
classified as a human, reset
 j
 0; else increase

 j
  1.
j

v. If  j
is greater than threshold,
remove trajectory j.

(b) Search for new humans in frame

 i
and compute the Euclidean

distance  j ,k
between each
newly detected human k and each

Fig. 1: Sample positive human images used for detector
existing trajectory j
. When training.

 j ,k
 for all j, where  is a
threshold (in pixels) less than the
width of the tracking window,
initialize a new trajectory for
detection k.
2.1 Sample Creation and Human Detector
Training

We use a machine learning approach for object

detection based on Haar-like features, originally
proposed by Viola et. al.[5]. The training is
Fig2: Rectangle Haar-like features
performed using binary adaptive boosting to

COPO4O1-3
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

2.2.1 Predict: We predic  (  │ j ,i

) ,a
j , i 1

distribution over human object j's position in

frame i given our belief in its position in frame i-
1.

2.2.2 Measure: For each propagated particle

k, we measure the likelihood
 ( │ j , i , j ) using
k (k )
Fig3: Integral Image for fast feature evaluation a color histogram-

The Integral Image, ii ( x, y ) at location based appearance model. After computing the
likelihood of each particle we treat the
( x, y ) contains the sum of all pixels above and left likelihoods as weights, normalizing them to sum
of ( x, y ) and can be computed in single pass over to 1.
image using the following pair of equations.
2.2.3 Resample:- We resample the particles to
 ( x, y)  ( x, y 1)  i( x, y) (2.1.1) avoid degenerate weights. Without re-sampling,
over time, the highest-weight particle would tend
to a weight of one and the other weights would
ii( x, y)  ii( x  1, y)  ( x, y) (2.1.2) tend to zero. Re-sampling removes many of the
Where,  ( x, y) is the cumulative row sum and low weight particles and replicates the higher-
i ( x, y ) is the original image. We use adaptive weight particles. We thus obtain a new set of
equally-weighted particles. We use the re-
boosting technique[6] for training the human detector sampling technique described in [13].
system. The parameters and their values used in the
training set up are given in the table 1. 2.3 Motion and Appearance Models:-
Table 1: Human detector training parameters and their values.
Our motion model is based on a second-order auto-
regressive dynamical model. The autoregressive
Parameters Values Description
Npos 2,000 Number of positive images
consisting only human beings
model assumes the next state  t
of a system is a
Nneg 2,700 Number of negative images not function of some number of previous states and a

Nstages 20
consisting any human
Number of training stages
noise random variable  t
Minhitrate 0.991 Per stage minimum hit-rate

  f ( t 1,  t  2,.........,  t  p , i )
(99.10%)
maxfalsealarm 0.5 Maximum false alarm rate per t
(2.3.1)
stage (50%)
Mode All Use upright and tilted features
Width*height 40*40 Size of training images
We assume the simple second-order linear
Boosttypes DAB Discrete Adaptive Boosting autoregressive model

  2  j ,i 2  i (2.3.2)
j ,i j ,i 1

2.2 Particle Filter for Tracking:-

For tracking we use a particle filter. We use an Our appearance model is based on color histograms.
approach in which the uncertainty about an human's
state (position) is represented as a set of weighted
We compute a color histogram  j
in HSV space

particles, each particle representing one possible for each newly detected human and save it to
state. The filter propagates particles from frame i-1 to compute particle likelihoods in future frames. To
frame i using a motion model, computes a weight for compute a particle's likelihood we use the
each propagated particle using an appearance model, Bhattacharyya similarity coefficient between model

 
then re-samples the particles according to their (k )
histogram and observed histogram as
weights. The initial distribution for the filter is j
centered on the location of the object the first time it follows, assuming n bins in each histogram:
is detected. Here are the steps in more detail:

COPO4O1-4
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

 ( │ j , i , j )  e  
(k )
k (k ) d ( , )
j (2.3.4)

And ,

d ( ,  
n
) 1  
(k ) (k )
(2.3.5)
j j ,b b
b 1

 and  
(k )
and 
(k )
and b
denote bin b of ,
j ,b j Frame 100 Frame 125
respectively. A more sophisticated appearance model
based on local histograms along with other
information such as spatial or structural information
would most likely improve our tracking performance,
but we currently use a global histogram computed
over the entire detection window because of its
simplicity.

3. Experimental Results:-
Frame 150 Frame 175
We have experimented the proposed automatic
human detection and tracking technique on a number
of videos. The results with some of the representative
videos are given here. The human detection and
tracking starts automatically without providing any of
the initialization details unlike many other tracking
techniques in which operator intervention is required.
Our proposed detector detects the human beings
irrespective of their body poses and locations in
video (fig. 4). The results given in fig. 5 show that
the proposed technique performs fairly good in
outdoor environment in low lighting conditions and Frame 200 Frame 225
is quite suitable for visual surveillance applications.

Frame 1 Frame 25 Frame 250 Frame 275

Frame 300 Frame 325

Frame 50 Frame 75

COPO4O1-5
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
Frame 350 Frame 375
Frame 400 Frame 425
Fig. 4: Detection and tracking results with the proposed method on
a real-time video in outdoor environment. The video was shot in
evening in dim lighting conditions.
References
[1] Weiming Hu and Tieniu Tan , “A Survey on Visual
Surveillance of Object Motion and Behaviors ”, IEEE Trans.
Systems, Man, and Cybernetics, Vol. 34, No. 3, pp. 334-352,2006.
[ 2] E. Stringa and C.S. Regazzoni, ”Real-time Video Shot
Detection for Scene Surveillance Application,” IEEE Trans. Image
Processing, Vol. 9, pp. 69-79, 2009.
[3] I. Haritaoglu, D. Harwood, and L. Davis, “W4: Who? When?
Where? What? a Real-time System for Detecting and Tracking
People,” in Proc. IEEE Conference on Face Gesture Recognition,
pp. 222–227, 1998.
[4] D. Koller, J. Weber, T. Huang, J. Malik, B. Rao G. Ogasawara,
and S. Russell, “Toward Robust Automatic Traffic Scene Analysis
in Real-time,” In Proceedings of Int. Conference on Pattern
Recognition, Vol. 1, pp. 126–131, 1994,.
[5] Paul Viola and Michael Jones, “Rapid Object Detection using a
Boosted Cascade of Simple Features,” In the Proceedings of IEEE
International Conference on Computer vision and Pattern
Recognition, 2001.
[6] Freund, Y. and Schapire, R.” A decision theoretic
generalization of on-line learning and an application to boosting,”
Journal of Computer and System Sciences 55 pp.119–139, 2007
[7] P. Lanvin, J-C. Noyer, M. Benjelloun,” Object Detection and
Tracking using the Particle Filtering,” In Proc. of IEEE 37th
Asilomar Conference on Signals, Systems and Computers, Vol.
2,pp.1595-1599, 2003.
[8] Jingling Wang, Yan Ma, Chuanzhen Li, Hui Wang, Jianbo Liu,
“An Efficient Multi-Object Tracking Method using Multiple
Particle Filters” In the Proceedings of World Congress on
Computer Science and Information Engineering, pp. 568-572,
2009.
[9] Yuan Chen, Shengsheng Yu, Jun Fan, Wenxin Chen,
Hongxing Li, “An Improved Color-based Particle Filter for Object
Tracking,” In Proceedings of IEEE Second International
Conference on Genetic and Evolutionary Computing, pp. 360-363,
2008.
[10] Francois Bardet, Thierry Chateau, Datta Ramadasan, “
Illumination Aware MCMC Particle Filter for Long-term Outdoor
Multi-Object Simultaneous Tracking and Classification,” In
Proceedings of IEEE 12th International Conference on Computer
Vision, pp. 1623-1630, 2009.
[11] Cheng Luo, Xiongcai, Jian Zhang, “ Robust Object Tracking
using the Particle Filtering and Level Set Methods: A Comparative
Experiment,” In the Proceedings of IEEE 10th Workshop on
Multimedia Signal Processing, pp. 359-364, 2008.
[12] P. Brasnett, L. Mihaylova, N. Canagarajah, and D. Bull,
“Particle Filtering for Multiple Cues for Object Tracking in Video
Sequences,” in Proc. of the 17th SPIE Annual Symposium on
Electronic Imaging, Science and Technology, vol. 5685, pp. 430–
440,2005
[13] M. Isard and A. Blake, “Contour Tracking by Stochastic
Propagation of Conditional Density,” in European Conf. on Comp.
Vis. , pp. 343–356, Cambridge, UK,1996
[14] M. Isard and A. Blake, “Condensation–Conditional Density
Propagation for Visual Tracking,” Intl. Journal of Computer
Vision, vol. 28, no. 1, pp. 5–28, 1998.
[15] P. P´ erez, J. Vermaak, and A. Blake, “Data Fusion for
Tracking with Particles,” Proceedings of the IEEE, vol.92, no. 3,
pp. 495–513, March 2004.
[16] M. Isard and A. Blake, “A Mixed-State Condensation
Tracker with Autmatic Model-Switching,” in Proc. Int. Conf.
Computer Vision, pp. 107–112,1998.
[17] M. Isard and J. MacCormick, “Bramble: a Bayesian Multiple
Blob Tracker,” in Proc. Int. Conf. Computer Vision, pp. 34–
41,2001
[18] J. Czyz, B. Ristic, and B. Macq, “A Color-Based Particle
Filter for Joint Detection and Tracking of Multiple Objects,” in
Proc. of the ICASSP, 2005
COPO4O1-6
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

An Efficient Image Segmentation Method Based on Normalized Cut.

Vaibhav Sharma, Asstt. Professor RKGIT , Member, ICEIT

Abstract

An efficient method for segmentation of images has been proposed by incorporating the advantages of the
normalized cut partitioning method.The proposed method pre-processes an image by using the normalized cut
algorithm to form segmented regions and applied to form the weight matrix W. Since the number of the
segmented region nodes is much smaller than that of the image pixels. The proposed algorithm allows a varied
dimensional images with significant reduction of the computational complexity comparing to conventional
Ncut method based on direct image pixels. The experimental results also verify that the proposed algorithm
behaves an improved performance comparing to Ncut algorithm. This paper presents different ways to
approach image segmentation, explain an efficient implementation for each approach, and show sample
segmentations results. We show that an efficient computational technique based on a generalized eigen value
problem can be used to optimize this criterion. We have applied this approach to segmenting static images and
found the results to be very encouraging.

Index Terms---Normalized cut (Ncut);Grouping; image segmentation; graph partitioning.

I. Introduction algorithm for implementation method. For instance, a

The information we know directly about an image is
the intensity/color and location of each pixel, and
therefore these two variables are the central source of
data for any segmentation. Edges that look clearly
defined from far away are often blurry and hard to
distinguish at the pixel level. Similarly, an area that
looks like a solid color at a distance may be seen as
mix of many colors at the pixel level and may be
interpreted as a number of segments. The importance
and difficulties of perceptual grouping have been very
well studied in vision research and image processing
literature. How to partition image to subsets, when
there are many possible ways of doing it? The
intuitive answer would be a partitioning based on the
available prior knowledge. The difficulty is specifying
the knowledge in a compact and useful way i.e.
defining the descriptors of an image. Over the
decades, different scientific communities proposed
various approaches to generalized image segmentation
problem. Tree-based hierarchical clustering structures,
region-based merge and split image processing
algorithms and statistical approach to image
segmentation introduced additional constraints on a
possible solution. Before any algorithm
implementation we have to determine what criterion
we want to optimize and if there is an efficient
general image segmentation approach based on
normalized cut (Ncut) by solving an eigensystem,
which was proposed by Shi and Malik [4] .In recent
years, Ncut algorithm is widely used for image
processing and related fields, such as the motion
picture , medical imaging, vector field of the
segmentation. However, this approach is NP-hard
problem, the more pixels in the image, the larger the
generated graph of nodes is, which bring difficulties
to solve the algorithm.
Considering the above issues come from the two
algorithms, this paper designed a combination of both
methods, the specific algorithm can be find in the
following part.
II. Normalized Cuts
The normalized cut method was introduced by Shi
and Malik in their paper Normalized Cuts and Image
Segmentation [4]. The paper is the one of the most-
cited papers in computer vision according to Google
Scholar, which indicates that the method is used
frequently in practice.
COPO4O2-1
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
The normalized cuts algorithm defines the optimal
segmentation in terms of a segmentation that divides
the image into two parts. To achieve a full
segmentation, it finds an optimal binary segmentation
and then recursively finds an optimal segmentation
within each segment. The way the normalized cuts
algorithm defines an optimal segmentation is quite
intuitive. Given some measure of similarity between
pairs of pixels, we define the optimal segmentation to
be the one that maximizes the similarity between pixels
that belong to the same segment, and minimizes the
similarity between pixels that belong to different
segments. Formally, we can define the optimization
problem as follows:
Let G = (V,E) be a graph, where V is the set of pixels
P, and E is the set of all edges {p, q} between every
pair of pixels p and q. We set the weight of an edge,
denoted w{p,q}, equal to a similarity measure between
the pixels p and q, where a higher weight corresponds
to pixels that are more similar.
The similarity measure used by Shi and Malik [4] is the
following:
(1)
where X(i) is the spatial location of i and F(i) is the
vector based on intensity, color, or texture information
at i. This definition defines the similarity between two
pixels to be proportional to the probability of pixel i
being generated by a Gaussian distribution centered at
pixel j. Notice that with this definition we do not need a
fully connected graph, but instead only place edges
between two pixels that have a positive weight, i.e.
when ||X(i) − X(j)||2 < r2.To measure the similarity
between segments we use a normalized cut. Given a
graph as specified above, and sets A and B, a
normalized cut, denoted Ncut, is defined to be:
(2)
(3)
The associativity and cut both measure how similar
two sets are to each other. The only difference is that
a cut requires that the two sets are a partition of V ,
while associativity generalizes this to measure the
similarity between any two sets. A normalized cut
measures how similar A is to B compared to how
similar A is to the whole image and B is to the whole
image. Thus, when the value of the normalized cut is
at a minimum, we are finding the segmentation
where the difference between A and B is maximized.
It may seem a bit odd that we do not consider the cut
alone but rather the ratio of cut over associativity.
However, this is necessary to avoid a bias resulting
from a different amount of edges in the cut
depending on the relative size of the segments.
Unlike in graph cuts, where at least one terminal-
link from every node must be in the cut, here if a
segment is just one pixel, the value of the cut is just
the sum of at most |P| − 1 edges out of the pixel.
This sums to significantly fewer edges than a cut
that divides the image in half and cuts at most (P/2)
edges. If we consider just the cut alone then we will
falsely encourage one of the two parts to contain
very few pixels.
To measure the similarity between pixels within the
same segment we formalize something similar to the
normalized cut called the normalized associativity.
The normalized associativity between sets A and B,
denoted Nassoc(A,B), is the following:
COPO4O2-2
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
The eigenvector associated with the smallest

(D W ) y
eigenvalue will minimize  and is
Dy
equivalent to Ncut(A,B). But, if we assume it for
now, we can observe a number of properties of λ and
(4) the associated eigenvectors.

First, Ncut cannot be negative because all the

weights in the graph are positive, so

(D W ) y
Precisely, the optimization problem is equivalent to  = 0 . Also, we always have the
finding the eigenvector with the second smallest Dy
eigenvalue of the following generalized eigenvalue eigenvalue = 0 associated with the eigenvectors
       
equation: y 1 and y 0 because D1 W 1 and D0 W 0 .
But, in both of these cases, the vector only has
one value and all the pixels are assigned to the same
segment. The normalized cut of this segmentation is
undefined because we need non-empty sets A and B
so that we do not divide by zero in the normalized
(5) cut. Therefore, the smallest eigenvector is λ = 0, and
we have that the second smallest
where D is a |V |×|V | diagonal matrix with di,i eigenvalue is the one that minimizes the normalized
= W is a |V| X |V| matrix with wi,j = w{i,j}, cut.
W is a |V| X |V| matrix with wi,j =w(i,j), 
(D W ) y
Now, to see that  is the same as the value
   i A d i ,i Dy
y x b(1 x ) and b
i B d i ,i of Ncut(A,B) when (and ) are discrete we
  notice that:
The vector y is a linear combination of x and
 
therefore y will define the same segmentation as x .
Below we will give an explanation, but not a
complete proof, as to why this is the correct equation
to solve. The explanation expands and clarifies the
complete proof found in the paper by Shi and Malik
[4]. The paper also provides evidence that the Because can only have the values 1 and 0, we
eigenvalue problem is O(n(3/2)) where n is the have that = −b(1− ) can also only take on
number of pixels in the image, when using a sparse only two values. If xi =1( i A ) then yi = 1. If xi = 0
matrix where most of the edge weights are 0. 
( i B )then yi = −b. Also note that y TW multiplies
the ith column of W by yi. Because column i contains
To give some intuition for why the eigenvector the weights of all the edges going into a vertex i, we
associated with the second smallest eigenvalue of 
have that y TW multiplies problem by throwing out
minimizes Ncut(A,B), consider a such vectors and not dividing the segment further. It
rewriting the equation as follows: would be problematic if the first partition resulted in
such a vector because then we would not be able to
find any segmentation at all. The paper does not
address this issue aside from briefly presenting a
non-recursive solution for which they show no
results.
(6)
In summary, the normalized cut algorithm can be
written as follows:
COPO4O2-3
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
• Decide on a similarity measure and compute matrices memory is wasted. Each vector is converted into a
D and W. square 2D texture by the application program.
• Solve the generalized eigenvalue equation from Vectors are padded with zero entries if they do not
Equation (5) for the eigenvector entirely fill the texture.

associated with the second smallest eigenvector.
 Preparing the data
• Find a splitting point of such that the normalized
cut is minimized and use it to partition the graph. We represent a image of size nXm as a column
vector V R N where N =n˟ m
Recursively partition each segment until some We compute the weight matrix W R N N where W(i,j)
sufficient measure of similarity is reached inside the represent relationship between pixel i and j in V.
segment We can use Gaussian function that take into
III. Efficient Image Segementation Algorithm consideration of distance, intensity different, crossing
The multiplication of matrix-matrix and matrix- edge information. Since each pixel is connected to its
vectors is one of the most central operations applied limited number of neighbors, this matrix is very sparse.
in linear algebra. Parallelization of numerical solvers (,)Wij Now given a weight matrix W R N N and k
would greatly accelerate a variety of algorithms that classes. We compute the degree matrix , where and
have a primary bottleneck of waiting for these each element is the sum of the corresponding rows in
calculations and on the availability of computational W The proposed method consists of the following
resources. Speeding up matrix operations by itself Steps :-
presents a challenging problem. Many techniques 1 Find the optimal eigensolution Z* `by solving the
are theoretical or run across accuracy limitations on leading k eigenvectors V[ k ] [v1, v2 , , vk ] for
the GPU. Graphics hardware is specialized in a 1 1 1
manner that makes it easier for these types of D 2WD 2 v v.Z * D 2V[ k ]
problems than using a standard CPU. For the 1
purpose of this proposal, ~
2 Normalize Z* by X * Diag(diag 2 ( Z *Z *T ))Z *
The representation of a 2D matrix as a set of
diagonal vectors, and finally as a set of 2D textures 3 Initialize orthonormal matrix R* R N k for finding
is shown in the optimal discrete solution:
~ ~
R1* [ X * (i,1), , X * (i, k )]T where random i [N ]
C 0 N 1 , a , a zero vector
For j=2 to k do {
~
C C abs ( X * R*j 1 )
~
R*j [ X * (i,1), , X * (i, k )]T , where i=argmin C[i]
{
4 Initialize convergence monitoring parameter * =0.
5 Find the optimal discrete solution X * {0,1}N k
~ ~
X X * R*
1
X * (i, l ) if l=argmax
0

figure1. 6 Find the optimal orthonormal matrix R* RN k

Kuger and Westermann represent matrices as set of

diagonal vectors and we store vectors in 2D texture ~ ~
SVD( X *T X * ) U U T
maps (see Figure above). To every diagonal starting
at the i-th entry of the firrst row in the upper right =trace (∑)
*
part of the matrix, its counterpart starting at the (N- If | | machine precison,then stop and output
i)-th entry of the first column in the lower left part of X *

the matrix is appended. In this way, no texture

COPO4O2-4
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
*
= directly applying the Ncut algorithm to the image
~ pixels. With the partitioning parameter K=5 and
R* UU T
table I depicts the weight matrix W of the region
7. Go to step 5
nodes. Fig. 1(d) is the display results of the Ncut
During the whole segmentation process, we have
algorithm partitioning. With the partitioning
very heavy computation of matrix-vector
parameter K=5. Fig. 2(d) shows the results of the
multiplication, for example, in solving the k leading
1 1
proposed algorithm partitioning.The Fig. 1 image
2 2 includes the animal in the foreground and the
eigenvectors of D WD , finding the orthonormal
bulrush in the background. Therefore, an effection
matrix, etc. However, most of the time, the matrix in
segmentation algorithm should distinguish the
the matrix-vector multiplication stay the same.
animal from the bulrush background. In Fig. 2(b),
Therefore, we can take advantage of the computing
many separated regions are segmented by Ncut
speed of matrix-vector multiplication on GPU to
algorithm, it is not a good segment result. In Fig.
enhance the overall segmentation speed.
2(c), the Ncut algorithm is directly applied to the
IV. EXPERIMENTAL RESULTS AND ANALYSIS
image pixels, the image is partitioned into two parts.
In this paper, we demonstrate the superiority of the
from Fig. 2(c) and Fig. 2(d). the segment result
proposed method by comparing the performance of the
distinguish the animal from the bulrush background.
proposed approach to existing methods using two
So the proposed method provides an improved
images.
performance compared to the Ncut algorithm.
The Fig. 2(a) size is 100×100. Set h=(x,y)=(8,1) .
With the partitioning parameter K=5. the analysis of
the segment result is the same as the Fig. 1.
Finally, we consider the computational cost of the
proposed algorithm and compare it with that of the
Ncut method. A PC with a 2.8-GHz Pentium CPU
and 1G memory is applied to carry out the
computation of the proposed method which consists
of two sections. The first section is to segment the
where F(i)={ R(i) ,G(i) ,B(i)} is the color vector of original image using the Ncut method. The second
region i, if the image is the gray-level image, then section is to partition the region nodes by the Ncut.
F(i) is the gray value and X(i) is the spatial location The Ncut method [4] requires 8-9 s to process the
of region node i. in this paper, we defined the F(i) is image of 100x100 pixels. When use the proposed
the average gray of the region i, and the X(i) is the method, an image has to be decimated into a size of
center of the region i. in this case, the select of F(i) 120×120, the required time is only 8-10 s. Table II
and X(i) are more representative. In order to compares the computational cost between the
compare other method to the proposed algorithm, we proposed algorithm and the Ncut method for the
use the example of two image, as depicted in Fig. 1 images depicted in Fig. 1 and Fig. 2.
and Fig. 2. Fig. 1 image size is 100×100. Fig. 1(a)
shows the original image, and Fig. 1(b) depicts the
resultant image after applying the Ncut algorithm,
with the red contours depicting the boundaries
between the regions. We set h=(x,y)=(8,1) . As a
result, the Ncut algorithm produces 8 regions. This
eight regions of R(i) in F(i) is
34,51,54,75,78,82,84,and 193 respectively (from
small to large). The G(i) and B(i) are not list.
Therefore, the constructing the weight matrix W of
the graph based onregion nodes using the Ncut
algorithm is easier and smaller than that based on
pixels. Fig. 1(c) shows the partitioning results by
COPO4O2-5
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

(a)
(a)

(b)

(b)

(c)
(c)

(d)
Fig. 1. (a) Original image. (b) The resultant image after applying
the Ncut Algorithm for edge computation. (c)The segmented
results by directly applying the Ncut algorithm to the image pixels.
(d)The results of the Ncut algorithm partitioning .
(d)
Fig. 2. (a) Original image. (b) The resultant image after applying
the proposed Algorithm for edge computation. (c)The segmented
results by applying the proposed algorithm to the image pixels.
(d)The results of the proposed algorithm partitioning .

COPO4O2-6
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

1 2 3 4 5 6 7 8
1 0 0 0 0 0 0 0 0
2 0 0 0 0 0 0 0 0
3 0 63.96937 63.04625 0 0 0 0 0
4 0 0 0 61.45696 0 0 0 0
5 0 0 0 0 56.69138 0 0 0
6 0 0 0 0 0 52.0375 53.44258 54.79843
7 0 0 0 0 0 0 0 0
8 0 0 0 0 0 0 0 0

TABLE I. WEIGHT MATRIX W OF ALL REGION NODES

No. of Time of Time of

Image Segementation Ncut Proposed
Size location(x) location(y) gradient(gx) gradient(gy) Threshold desired Method(s) Method(s)
Fig.1 100x100 8 1 20 3 0.02 5 7.0159 4.8593
Fig.2 150x150 8 1 20 3 0.02 5 NIL 12.1754
Fig.3 200x200 8 1 20 3 0.02 5 NIL 14.61
Fig.4 250x250 8 1 20 3 0.02 5 Nil 46.4527

TABLE II. COMPARED COMPUTATIONAL COST BETWEEN

THE PROPOSED METHOD AND THE NCUTMETHOD

V CONCLUSION minimization in vision. IEEE Transactions on Pattern

In this paper, efficient image segmentation algorithm
has been proposed and designed based on the
conventional Ncut algorithm . The effectiveness and
robustness of the proposed algorithm have verified by
some experimental results to express an improved
performance compared to the Ncut algorithm. Also, the
significant reduction to the computational cost of the
proposed algorithm in the experiments is favorable for
practical applications.
VI Future Work
We plan to deliver a tool that will allow for a
computational fast and efficient image segmentation
that can be applied to a video sequence.
References
[1]. Boykov,Y. and Kolmogorov, V. Computing geodesics
and minimal surfaces via graph cuts. In International
Conference on Computer Vision, vol. I:26-33, 2003.
Analysis and Machine Intelligence, 26(9):1124-1137,
2004.
[3]. Shi, J. MATLAB Normalized Cuts Segmentation
Code, Online. Internet. University of Pennsylvania. 15
May 2007 hhttp://www.cis.upenn.edu/ jshi/software/i.
[4]. Shi, J. and Malik, J. Normalized cuts and image
segmentation. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 22(8):888-905,2000.
[5]. J. Shi and J. Malik, “Motion segmentation and tracking
using normalized cuts,” Proceedings of the Sixth
International Conference on Computer Vision (ICCV),
Bombay, India, pp. 1154-1160, 1998.
[6]. Georgescu, B. and Christoudias, C. Code Available to
Download, Online, Internet.Rutgers University:
Robust Image Understanding Laboratory. 15 May 2007

[2]. Boykov,Y. and Kolmogorov, V. An experimental

comparison of min-cut/maxflow algorithms for energy
COPO4O2-7
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

OPTIMIZATION OF WIRED NETWORKS USING LOAD

BALANCERS
Yaduvir Singh1, Vikas Gupta2, Rahul Malhotra3
1
Thapar University, Patiala, Punjab
2,3
Adesh Group of Institutions, Sri Muktsar Sahib
1
yad_pra@yahoo.com
3
blessurahul@gmail.com
Abstract: - Computer Network is an interconnection of computers
for sharing of data and network resources. With the evolution of
Internet, the networking technology has not been just confined to
resource sharing but has boomed in the arena of cellular and
mobile communication as well. Rapid development in the field of
very large scale integration of complex circuits on to a smaller chip
has led to the evolution of high speed computer networks. In this
paper, performance analysis of the wired computer networks
through simulation has been attempted using OPNET as network
simulating tool. For wired networks, the performance parameters
like delay and throughput have been investigated with varying
transmission links. The load-balancing has been analyzed through
traffic sent, traffic received and CPU utilization parameters.
Keywords: OPNET, Load Balancers, Ethernet.
I. INTRODUCTION
Networks are being upgraded or deployed from scratch
all over the world. The field of network planning is thus
becoming all the more significant. Networks have grown rapidly
over the past few decades, with a challenge of connecting the
relevant networks together, so that the users can share the
network’s wealth. The undying thrust of humans to excel and fly
beyond the limits of horizon has provided the impetus for the
up-gradation of different technologies and Wireless Technology
is the blessed outcome of this effort. Change is the law of
nature. Over a span of time the wireless Technology has
overshadowed the wired technology. Wireless Technology is a
rapidly growing segment of the communications industry, with a
potential to provide high-speed, high quality information
exchange between the portable devices located anywhere in the
world. A wireless communication is a flexible data
communication system implemented as an extension to or as an
alternative for a wired communication. Wireless Local Area
Networks (WLANs) have been developed to provide users in a
limited geographical area with high bandwidth and similar
services supported by the wired Local Area Network (LAN).The
IEEE (Institute of Electrical and Electronics Engineers) 802.11
standards offer performance nearly comparable to that of
Ethernet. In addition, they provide scalability, mobility,
flexibility and ease of wireless access. Unlike wired networks,
wireless LANs, based on IEEE 802.11 standards, transmit and
receive radio waves through the air medium. The
communication can be between a wireless client and an Access
Point (AP), as well as among two or more wireless clients
within a certain range of each other. A wireless LAN basically
consists of one or more wireless devices connected to each other
in a peer-to-peer manner or through APs, which in turn are
connected to the backbone network providing wireless
connectivity to the covered area. As a result of wireless
technology, the traditional ways of wired networks have become
inadequate in meeting the challenges of present arena posed by
our collective lifestyles. Now a day’s wireless technology has
become a part of our life with major of the present applications
being upgraded to support wireless technology and all upcoming
applications are being manufactured with integrated wireless
support. Wireless technology provided a boon to the users of
computers, mobiles and other wireless technology enabled
gadgets by providing cheap and flexible wireless
communication access, without using the cables. Wired LAN is
slowly getting replaced by wireless LAN due to the ease in
access, mobility and scalability provided by it.
Among a wide range of applications of wireless
technology, few prominent areas of application are
telecommunications, security systems, industrial automation,
wireless sensor networks, distance learning education, health-
care and retail, as it is also easy to install on campuses, airports,
in stock markets, offices, hospitals, and other places.
II. WIRED LOCAL AREA NETWORKS
Wired Local Area Networking includes several
technologies such as Ethernet, token ring, token bus, FDDI and
ATM LAN. Some of these technologies survived for a while,
but Ethernet is by far the dominant technology. Evolution from
a 10Mbps Standard Ethernet to bridged Ethernet and then to a
switched Ethernet paved a way for faster Ethernet. Ethernet is a
contention media access method. In its purest form, contention
means that the computers are contending for use of the
transmission medium (wire). Any computer in the network can
transmit at any time (first come, first serve).
IEEE 802.3 standard specifies CSMA/CD as the access
method for first-generation 10-Mbps Ethernet, a protocol that
help devices share the bandwidth evenly without having the two

COPO5O1-1
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

devices transmit at the same time on the network medium. This

CSMA/CD protocol was created to overcome the problem of
collisions that occur when the packets are transmitted
simultaneously from different nodes over the same medium.
OPNET has been used for the simulation of these
computer networks. It provides a rich library of models for
implementing wireless and wired network models. In this
chapter we begin with description of the performance metrics
used in the thesis work. The performance metrics focused upon
in wired computer networks are:
Ethernet- Delay (sec): This statistic represents the end to end
delay of all packets received by all the stations.
FTP- Traffic Received (bytes/sec): This statistic means
average bytes per second forwarded to all FTP applications by Figure 2 Wired LAN Network Model
the transport layers in the network.
Ethernet – Traffic Received (bits/sec): This statistic defines 3.1 Performance Analysis of Wired LAN Model
the throughput (bits/sec) of the data forwarded by the Ethernet The objects used to design the network model of wired LAN
layer to the higher layers in this node.
are:
HTTP – Page response time (sec): Specifies time required to
retrieve the entire page with all the contained inline objects.  One Server: Web Server
 Two Switches: ethernet16_switch node model
 Two Circle Annotations
 Thirty Users: 10 users per LAN.
 Link Types: 10 Base T, 100 Base T, 1000 Base X
 Application Configuration: To define the type of
network load.
 Profile Configuration: To create new profiles and
assign the configured applications

3.2 Performance Analysis using Link Type Variation

The network design shown in Figure 2has been
modeled. In this network design the links have been varied as
shown in Table 1.
Figure 1 Performance analysis metrics of Wired LAN Table 1 Performance Analysis scenarios for Link type variation

III. PERFORMANCE ANALYSIS Parameter Scenario 1 Scenario 2 Scenario 3

In this simulation scenario, comparison has been done Type of Link used 10 Base T 100 Base T 1000 Base X
by varying the types of media (Ethernet links) used in the
No. of 30 (10 Per 30 (10 Per 30 (10 Per
networks for communication between the server and the clients Workstations LAN) LAN) LAN)
in a wired LAN. The network design as shown in Figure 2has For performance analysis, following metrics have been
been modeled. A new project is created with a new scenario.
investigated:
This new scenario model created has logical network scale with
Ethernet topology. The object palette of the selected Model Ethernet- Delay (sec): This statistic represents the end
family is used to select the objects required to create the to end delay of all packets received by all the stations.
network model (using drag and drop method) based on the
required technology. The graphical user interface helps to
arrange and place the objects as required and also provides an
ease to connect these objects conveniently.

COPO5O1-2
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Figure 3 Ethernet Delay (sec) Figure 4 Ethernet Throughput (bits/sec) for link variation

As shown in Figure 3, the delay occurs due to the 3.3 Performance Analysis using Load-Balancer
heavy traffic that passes through the network. The cause of
heavy traffic is the large number of users (30) as shown in the In another simulation for performance analysis, the
Figure 1. The amount of traffic sent and traffic received number of stations in a network is varied as shown in the
increases/ decreases depending on the number of users and the figure5, in which there is only one server, a load-balancer, a
amount of data accessed. In first scenario according to Table firewall and other network objects as per the specifications of
4.1, the type of Ethernet link used is 10 Base T. The graphs in the network required.
Figures 3, show the delay statistics (which is maximum than the
other two types of links). The reason is the low data rate of the
10 Base T links as compared to the other two types of links
listed in Table 1 i.e. 100 Base T and 1000 Base X. Another
parameter used for performance analysis is:
Ethernet – Traffic Received (bits/sec): This statistic
defines the throughput (bits/sec) of the data forwarded by the
Ethernet layer to the higher layers in this node.
The Ethernet throughput depends on the amount of
delay that occurs in the transmission of packets. Figure 3
indicates that the delay in case of 100 Base T and 1000 Base X
is less as compared to a network with 10 Base T. Hence the
throughput achieved in the case of 100 Base T and 1000 Base X
is much higher as compared to throughput achieved in case of
Figure5 Wired LAN Network Model using Load Balancer
10 Base T as shown in following figure 4. Moreover, the links –
100 Base T and 1000 Base X provide more bandwidth as Then scenarios described in Table 2 have modelled and
compared to the 10 Base T. Hence the traffic received through simulated using the network design shown in figure 5. The table
the Ethernet increases with the increase in bandwidth. shows the variations in the number of nodes:

Table 2 Performance Analysis scenarios for varying number of nodes

Parameters Scenario Scenario 5 Scenario Scenario 7
4 6
No. of Workstations 10 20 30 30
No. of Servers 1 1 1 3
Load Balancing None None None Random
Policy policy
When the load balancer receives a packet from a client
machine, it must choose the appropriate server to handle the
request. The load balancer will use the load balancing policy to

COPO5O1-3
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

determine which server is most appropriate. Following load

balancing policies can be used:
Random: The load balancer chooses one of the candidate
servers at random.
Round-Robin: The load balancer cycles through the list of
candidate servers.
Server Load: The load balancer chooses the candidate server
with the lowest CPU load.
Number of Connections: The load balancer keeps track of the
number of connections it has assigned to each server. When a
new request is made, it chooses the server with the fewest
connections.
The Simulation study on load balancer has been done for the
following reasons: Figure 7 Response time (sec) for varying number of nodes
To determine, how much load on the individual servers can be
The response time associated to the highest load level
reduced, using a load balancer, in a network with heavy
under balancing is lower than any other single-server
Internet applications.
configuration (including the one with the smaller number of user
To determine the best load balancing for improving the
sessions) as is visible from figure 7.
performance of wired networks.

The following statistics are investigated for performance

analysis:
Traffic Sent (in Bytes/sec or Packet/sec): The amount of
traffic sent to a particular application server.
Traffic Received (in Bytes/sec or Packet/Sec): The amount
of traffic received from a particular application server.
As can be seen in Figure 6, as the number of user
increases, more traffic is generated. It is also observed that the
introduction of load balancing for a single server has no effect
on the level of generated traffic.

Figure 8 A wired LAN network model with load- balanced multiple servers

Using the network design modeled in Figure 8, four

Figure 6 Traffic sent (bytes/sec) for varying number of nodes
But as can be seen in figure 7, the use of load-balancing for
multiple servers (in figure 8) can significantly improve
performance.
scenarios (Table 3) have been designed. Firstly, the scenario
with random policy in which the load balancer randomly
chooses from one of the three application servers according to
the specified weight. Secondly, load balancer with the round-
robin policy in which the load balancer chooses each application
server in turn depending on the server weight. Thirdly in
scenario 9, the load balancer has server-load policy, in which the
load balancer chooses the server with the lowest load at the time
when a request is made. Then in scenario 10 where the load
balancer with number-of-connections policy tracks the number
of connections it has with each server. When it needs to
forward a request, the load balancer chooses the server with the
least connections.
Table 3 Performance Analysis scenarios with load-balancer

COPO5O1-4
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

For performance analysis, following statistics are collected at

Parameters Scenario Scenario Scenario Scenario 10
the load balancer and the server:- 7 8 9
The statistics at the load balancer are collected per No. of
30 30 30 30
application and per server. When load balancing is Workstations
enabled the statistics collected are: No. of Servers
3 3 3 3
Traffic Sent (in bytes/sec or packet/sec): The Load Random Round Server Number of
amount of traffic sent to a particular application server. Balancing policy Robin Load connections
Traffic Received (in bytes/sec or packet/sec): The Policy Policy Policy
amount of traffic received from a particular application
server.
The analysis from figure 9 shows averages for traffic
associated to each balanced server. In this section of
performance analysis, the simulation has been carried out for
10.0 hours.

Figure10 Traffic sent & Received (Bps) using No. of connections Policy

The investigations from Figure 10 showing average

traffic sent and Figure 4.17 showing average traffic received by
each balanced server indicates that the number of connections
policy provides faster load stability and the servers converge to
same average load over a period of 3 to 4 hours.
Figure 9 Traffic sent & received (bytes/sec) using server-load policy

The investigations from figure 9 for traffic sent

(bytes/sec) and traffic received (byte/sec) reveal that after a
period of more than four hours, all the balanced servers seem to
converge to similar average loads. This analysis has been done
using server-load balancing policy.
For further analysis of the network design presented in
Figure 8, a performance comparison of other load balancing
policies like random policy, round-robin policy and number of
connections policy, under the same user activity, has also been
done.
Figure 10 shows comparison using number of
connections load balancing policy on three load balanced
servers. Similarly Figure 4.17 and 4.18 show comparisons for Figure 11 Traffic sent & Received (bytes/sec) using Round-Robin policy
traffic sent and traffic received using round – robin policy and Figure 11 shows the average traffic sent and traffic
Figure 4.19 and Figure 4.20 show comparison of traffic sent and received (in bytes/sec) by each balanced server under the round-
received using random load balancing policy. robin policy. This policy assigns each incoming requests to the
next server in the sequence. Though the server 1 and server 2
with similar selection weights as assigned in the load balancer,
tend to have same average activity, the sever 3 with more
selection weight shows variance in the average server load.

COPO5O1-5
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
Figure 12 Traffic sent & Received (bytes/sec) using Random policy
The statistics of the average traffic sent and received in
bytes/sec under the random policy shows the same phenomenon
as shown by the round robin policy in Figure 12. This is due to
the selection weights assigned to the balanced server using the
load balancer. Another comparison of the load-balancing
policies has been done by simulating the effect of all these
policies on the same server.
Figure 13 Traffic sent Received (bytes/sec) using different load balancing policies
The investigations present that while average traffic
received or sent (in bytes/sec) with the number of connections
policy is more than others, the average traffic received or sent
(in bytes/sec) with the server-load policy is lesser than others
(Figure 13).
The following statistic is collected at each of the balanced
servers.
CPU load (%): The percentage load on server.
The following figures present the percentage load on
server. This statistic reports the utilization, in percentage (%),
of the 'simple CPU'. The 'simple CPU' is used to model the IP
packet forwarding delays and application processing delays.
This statistic also takes into account the background CPU
utilization specified in the "CPU background utilization"
attribute. This statistic measures the utilization of central CPU
only. The analysis of figures 14 (which defines the impact of
number of connections policy & random policy on all the
servers), Figure 15(which defines the impact of round-robin
policy & server-load policy on all the servers) reveal that the
load on all the balanced servers averages to be approximately
same using the number of connections policy and server-load
policy.
Figure 14 CPU Utilization using number of connections policy & Random policy
But when the random policy and round robin load balancing
policies are used, Figures 14 shows that the CPU utilization (%)
of server 3 is more while there is little or no variance of CPU
load for server 1 and server2. The cause of more utilization of
server 3 is the assigning of selection weights in the load server.
Figure 15 CPU Utilization using Round Robin policy& Server load policy
The selection weight is the attribute which is applicable
only when the load balancing technique is either set to "round-
robin" or "random". It is used when determining the server to
choose. The probability of choosing a given server is: (selection
weight) / (total weight). The value of selection weight assigned
to server 3 is 10, which is twice the value of selection weights
assigned to the server 1and server 2. The overall analysis
indicates that the CPU utilization is minimum using server load
balancing policy.
4. CONCLUSIONS
COPO5O1-6
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

In this paper work performance analysis of the wired network [11] Dr. Reinhard Kuch, “Studienbrief 2: Simulation of
configuration through simulation was started with the networks using OPNET ITGURU Academic Edition v
investigation of the network performance using various types of 9.1,” version: 1.0, 26-04-2009.
[12] “Opnet_Modeler_Manual,” available at
links. The impact of various network configurations on the
http://www.opnet.com
network performance was analyzed using the network [13] T.Velmurugan, Himanshu Chandra and S. Balaji,
simulator- OPNET. It has been investigated that performance of “Comparison of Queuing disciplines for Differentiated
the wired Networks is good if high speed Ethernet links are used Services using OPNET,” IEEE, ARTComm.2009, pp.
under heavy network loads. Moreover, the mechanism of load 744-746, 2009.
balancing also improves the performance by reducing and [14] Yang Dondkai and Liu Wenli, “The Wireless Channel
balancing the load equally among multiple servers. Modeling for RFID System with OPNET,” in the
Proceedings of the IEEE communications society
REFERENCES sponsored 5th International Conference on Wireless
communications, networking and mobile computing,
[1] Behrouz A Forouzan, Data Communications and
Beijing, China, pp. 3803-3805, September 2009.
Networking: Tata McGraw Hill, 4th Edition, 2007.
[15] Ikram Ud Din, Saeed Mahooz and Muhammad Adnan,
[2] Sameh H. Ghwanmeh, “Wireless network performance
“Performance evaluation of different Ethernet LANs
optimisation using Opnet Modeler,” Information
connected by Switches and Hubs,” European Journal of
Technology Journal, vol. 5, No 1, pp. 18-24, 2006.
Scientific Research, vol. 37, No 3, pp. 461- 470, 2009.
[3] Sarah Shaban, Dr. Hesham, M.El Badawy, Prof. Dr.
Attallah Hashad, “Performance Evaluation of the IEEE
802.11 Wireless LAN Standards,” in the Proceedings
of the World Congress on Engineering-2008 , vol. I,
July 2-4, 2008.
[4] A. Goldsmith, Wireless Communications: Cambridge
University Press, 1st Edition, August 2005.
[5] Mohammad Hussain Ali and Manal Kadhim Odah,
“Simulation Study of802.11b DCF using OPNET
Simulator,” Eng. & Tech. Journal, vol. 27, No. 6, pp.
1108-1117, 2009.
[6] Hafiz M. Asif, Md. Golam Kaosar, “Performance
Comparison of IP, ATM and MPLS Based Network
Cores Using OPNET,” in 1st IEEE International
Conference on Industrial & Information Systems
(ICIIS 2006), Sri Lanka, 8-11 August, 2006.
[7] Dibyendu Shekhar, Hua Qin, Shivkumar
Kalyanaraman, Kalyan Kidambi, “Performance
Optimization of TCP/IP over Asymmetric Wired and
Wireless Links” in the Proceeding of conference on
Next Generation Wireless Networks: Technologies,
Protocols, Services and Applications (EW-2002), –
Florence, Italy, February 25-28, 2002.
[8] Born Shilling, "Qualitative Comparison of Network
Simulation Tools,” Institute of Parallel and Distributed
Systems (IPVS), University of Stuttgart, 2005.
[9] “OPNET IT Guru Academic Edition” available at
http://www.opnet.com/university_program/itguru_acad
emic_edition
[10] Ranjan Kaparti, “OPNET IT Guru: A tool for
networking education,” REGIS University.

COPO5O1-7
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

AN INTELLIGENT ALGORITHM FOR TASK ALLOCATION IN DISTRIBUTED

NETWORK USING STATIC APPROACH
1
Kapil Govil, 2Dr. Avanish Kumar and 3Rakesh Kumar Dwivedi
1,3
Teerthanker Mahaveer University, Moradabad – 244 001, (UP), INDIA
2
Bundelkhand University, Jhansi – 284 126, (UP), INDIA
1
drkapilgovil@gmail.com, 2dravanishkumar@yahoo.com, 3r_dwivedi2000@yahoo.com

ABSTRACT suitable for arbitrary number of processor in

Distributed Network is one the appropriate
network for providing the optimal solution in
real-time applications problems. The
systematic allocation of tasks needs to plays
the key role to optimize the overall processing
cost of the distributed network. Distributed
network consists of multiple
autonomous computers that communicate
through a common communication channel.
The computers interact with each other in
order to achieve a common goal. A computer
program that runs in a distributed system is
called a distributed program, and distributed
programming is the process of writing such
programs. The allocation problems in any
computer system play the key role for
deciding the performance of the system. The
allocation put the direct impact of software
resources as well as hardware resources. In
distributed network, partitioning of the
application software in to module and the
proper allocation of these modules dissimilar
processors are important factors, which
determine the efficient utilization of resources.
The static model discussed in this paper
provide an optimal solution using algorithm of
Assignment problem for assigning a set of
“m” tasks of a task to a set of “n” processors
where m > n in a computer system for
evaluation for optimal processing cost of the
system. The several sets of input data are
considered to test the efficiency of the
algorithm and it is found that the algorithm is
all the cases.
Keywords: Distributed Network, Processor,
Task, Allocation, Processing Cost.
INTRODUCTION
Distributed network refers to the means by
which a single computer program runs in more
than one computer at the same time. In
particular, the different elements and objects
of a program are being run or processed using
different computer processors. In a distributed
networking, the program runs like it would in
a single computer even when it is, in fact,
using different computer processors. This
means that no single computer carries the
entire burden on system resources that running
a computer program usually implies.
One of the major research problems for
distributed networks is the allocation problem,
in which tasks are assigned to various
processors of the network, in such a way that
processing cost is to be minimized as per the
requirement. These problems may be
categorized as static (AVAN2001,
KUMA1995a, KUMA1995b, KWON2006)
and dynamic (JACQ2005, AVAN1999,
KUMA1996, KWON2006, PALM2005) types
of task allocation. Some of the other related
methods have been reported in the literature,
such as, Integer Programming (DESS1980,

COPO3O1-1
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

BRIA2003, MISR1991, MUHA2002), Load compute total processing cost by adding total
Balancing (JACQ2005, DANI2005, processing cost of task, which are assigned at
SAEE2005, SHU2009) and Modeling specified processor.
(RENE2002, JAVI2000, KENT2002). Tasks
are allocated to various processors of the ALGORITHM
distributed network in such a way that overall
processing cost of the network should be Start Algorithm
minimized. As it is well known that the tasks Read the number of task in m
are more than the number of processors of the Read the number of processor in n
network. For i = 1 to m
For j = 1 to n
OBJECTIVE
Read the value of
processing cost (c) in
In the Distributed Processing Environment
Processor Cost Matrix
(DPE), it is the common problem to allocate
tasks where the number of tasks is more than namely PCM(,)
j=j+1
the number of processors. The objective of the
Endfor
present research paper is to enhance the
i=i+1
performance of the distributed networks by
Endfor
using the proper utilization of its processors
Calculate the sum of each row and
and as well as proper allocation of tasks. In
column and store the results in
the present research paper the type of
Modified Processor Task Matrix
allocation of task to the processor is static in
MPTM(,)
nature. As in this paper the performance is
By arranging the MPTM(,) in
measured in terms of processing cost, so we
ascending order of their row_sum and
have to minimize the processing cost to obtain
column_sum, we get Arranged
the best performance of the processors. To
Processor Task Matrix APTM(,)
overcome from the problem we have designed
i=1
an intelligent algorithm for task allocation.
While all tasks != SELECTED
Select the biggest possible
TECHNIQUE
square matrix from left upper
corner and store it into SMi(,)
In order to evaluate the overall optimal
Apply algorithm of Assignment
processing cost of a distributed network, we
Problem [KANT2002] on SMi(,)
have chosen the problem where a set P = {p1,
i=i+1
p2, p3, …….pn} of „n‟ processors and a set T =
Endwhile
{t1, t2, t3, …….tm} of „m‟ tasks, where m>n.
Club processor-wise overall optimal
The processing cost of each task to each and
processing cost
every processor is known and it is mentioned
State the results
in the Processing Cost Matrix of order m x n.
End Algorithm
After making a matrix of same order taking in
ascending order of its sum of row and sum of
column, we apply the algorithm of assignment IMPLEMENTATION
problem on it. For each processor we evaluate
In the present research paper, the distributed
the overall allocation of each task; and
allocation of the task on the processor which network consist a set P of 4 processors {p1, p2,
p3, p4} and a set T of 10 tasks {t1, t2, t3, t4, t5,
has the minimum processing cost. Finally we

COPO3O1-2
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

t6, t7, t8, t9, t10}. It is shown in the figure1. The in the Processor Cost Matrix PCM(, ) of order
processing cost (c) of each task to each and 10 x 4.
every processor is known and it is mentioned
p1 p2 p3 p 4 Row_Sum
t1 t2 t3 t1 t2 t3 t1 600 200 900 300 2000
t4 t5 t6 t7 t4 t5 t6 t7
t2 500 300 200 100 1100
t8 t9 t 10 t8 t9 t 10

t3 800 700 300 400 2200

p1 p2
t4 100 400 600 300 1400
t5 400 500 600 200 1700
MP T M(,)
t6 200 100 800 900 2000
t7 200 800 600 700 2300
t8 600 700 800 900 3000
p3 p4

t9 400 500 600 200 1700

t1 t2 t3 t1 t2 t3 t10 300 700 400 200 1600
t4 t5 t6 t7 t4 t5 t6 t7 Column_Sum 4100 4900 5800 4200
t8 t9 t 10 t8 t9 t 10

Figure1: Processor Task Combination By arranging the MPTM(,) in ascending order

of their row_sum and column_sum, we get
p1 p2 p3 p4 Arranged Processor Task Matrix APTM(,) of
t1 600 200 900 300 order 10 x 4.
t 2 500 300 200 100 p1 p 4 p2 p3 Row_Sum
t 3 800 700 300 400 t2 500 100 300 200 1100
t 4 100 400 600 300 t4 100 300 400 600 1400
t10 300 200 700 400 1600
t 5 400 500 600 200
P CM(,) t5 400 200 500 600 1700
t 6 200 100 800 900 t9 400 200 500 600 1700
t 7 200 800 600 700 AP T M(,)
t1 600 300 200 900 2000
t 8 600 700 800 900 t6 200 900 100 800 2000
t 9 400 500 600 200 t3 800 400 700 300 2200
t7 200 700 800 600 2300
t10 300 700 400 200
t8 600 900 700 800 3000
Column_Sum 4100 4200 4900 5800
Now, calculate the sum of each row and
column and store the results in Modified
Processor Task Matrix MPTM(,) of order 10 x Now, selecting the biggest possible square
4. matrix from left upper corner, we get Selected
Matrix SMi(,) of order 4 x 4.
p1 p 4 p2 p3
t 2 500 100 300 200
t 100 300 400 600
SM 1 (, ) 4
t10 300 200 700 400
t 5 400 200 500 600

COPO3O1-3
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Now, by using algorithm of Assignment Processor Task

Problem [KANT2002] on SMi(,), we get task p1 t7
allocation as – p2 -
Processor Task p3 -
p1 t4 p4 t8
p2 t2
p3 t10 The overall processor-wise task allocation
p4 t5 along with optimal processing cost is
mentioned in table1.
Now, we eliminate allocated tasks (i.e. t2, t4,
t5, t10) from APTM(,); and again selecting the Table 1: Processor-wise Task Allocation
biggest possible square matrix from left upper Processor Assigned Task Processing Cost
corner, we get selected matrix SM2(,) of order p1 t4*t6*t7 500
4 x 4. p2 t1*t2 500
p1 p 4 p2 p3 p3 t3*t10 700
t 9 400 200 500 600 p4 t5*t8*t9 1300
t 600 300 200 900 Optimal Processing Cost 3000
SM 1 (, ) 1
t 6 200 900 100 800
The graphical representation of the optimal
t 3 800 400 700 300 allocation is shown in figure 2.
t7 t1
t6
t2
Now, on applying the algorithm of t4
Assignment Problem [KANT2002] on matrix p1 p2

SM2(,), we get task allocation as –

Processor Task
p1 t6
p2 t1
p3 t3
p4 t9
p3 p4
Again, we eliminate allocate tasks (i.e. t1, t3,
t3 t5
t6, t9) from APTM(,) and again selecting the
biggest possible square matrix from left upper t 10
t8
t9

corner, we get Selected Matrix SM3(,) of order

2 x 2. Figure 2: Optimal Allocation
p1 p 2
t 2 7 The processor-wise processing cost graph is
SM 3 (, ) 7 shown in the figure 3.
t8 6 9

Now, on applying the algorithm of

Assignment Problem [KANT2002] on matrix
SM3(,), we get task allocation as –

COPO3O1-4
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Processing Cost Processing Cost Table 2: Optimal Results

No. of No. of tasks Optimal
1400
1200 processors (n) (m) Results
1000 3 4 12
800
600
3 5 15
400 3 6 18
200
0
3 7 21
t4*t6*t7 t1*t2 t3*t10 t5*t8*t9 3 8 24
p1 p2 p3 p4 4 5 20
4 6 24
Figure 3: Processor-wise Processing Cost 4 7 28
Graph 4 8 32
4 9 36
CONCLUSION 5 6 30
5 7 35
In this research paper we have chosen the 5 8 40
problem, in which the number of the tasks is 5 9 45
more than the number of processors of the 5 10 50
distributed network. The model mentioned in
this paper is based on the consideration of The graphical representation of results of table
processing cost of the tasks to various 2 are shown by figure 4, 5 and 6 as mentioned
processors. The method is presented in pseudo below,
code and implemented on the several sets of m
No. of Processors = 3
input data to test the performance and Complexity

8 25
effectiveness of the pseudo code. It is the 7
common requirement for any allocation 6
20

problem that the tasks have to be processed

No. of tasks

5
15

with minimum processing cost. Here, 4

10
3
performance is measured in terms of 2
processing cost of the task that has been 1
5

processed by the processors of the network 0 0

1 2 3 4
Example
and also these tasks have been processed
optimally. As we know that, the analysis of an Figure 4: Complexity Graph
algorithm is mainly focuses on time
m
complexity. Time complexity is a function of No. of Processors = 4
Complexity

input size „n‟. It is referred to as the amount of 9 35

8
time required by an algorithm to run to 7
30

completion. The time complexity of the above 6

25
No. of tasks

20
mentioned algorithm is O(mn). By taking 5

4 15
several input examples, the above algorithm 3
10
returns following results as in table 2. 2
5
1

0 0
1 2 3 4
Example

Figure 5: Complexity Graph

COPO3O1-5
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

m
No. of Processors = 5 No. of Processors = 3 Algorithm SAGA1991
Complexity
Present algorithm
10 50
9 45 250
8 40
7 35
200
No. of tasks

6 30
5 25
150
4 20
3 15
2 10 100

1 5
0 0
50
1 2 3 4
Example

Figure 6: Complexity Graph 0

1 2 3 4 5 6

The performance of the algorithm is compared Figure 7: Comparison Graph

with the algorithm suggested by Sagar et. al. No. of Processors = 4 Algorithm SAGA1991
Present algorithm
[SAGA1991]. Table 3, shows the time 350

complexity comparison between algorithm 300

[SAGA1991] and present algorithm. 250

200
Table 3: Comparison Table 150
Time Time 100
Complexity Complexity 50
n m of algorithm of present
0
(SAGA1991) algorithm 1 2 3 4 5 6 7

O(m2n) O(mn) Figure 8: Comparison Graph

3 4 48 12 No. of Processors = 5 Algorithm SAGA1991

3 5 75 15 600
Present algorithm

3 6 108 18
500
3 7 147 21
400
3 8 192 24
4 5 100 20 300

4 6 144 24 200

4 7 196 28 100

4 8 256 32 0

4 9 324 36 1 2 3 4 5 6 7

5 6 180 30 Figure 9: Comparison Graph

5 7 245 35
5 8 320 40 REFERENCES
5 9 405 45
5 10 500 50 AVAN1999 Avanish Kumar,
Optimizing for the Dynamic
Task Allocation,
From the table 3, it is clear that present Proceedings of the „III
algorithm is much better for optimal allocation Conference of the
of tasks that upgrade the performance of International Academy of
distributed network. Figures 7, 8 and 9 show Physical Sciences,
the pictorial representation between algorithm Allahabad; (1999), 281-
SAGA1991 and present algorithm.

COPO3O1-6
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

294. Environments, Elsevier

AVAN2001 Avanish Kumar, An Inc., Vol. 60(2000) No. 6,
Algorithm for Optimal 741 – 763.
Index to Tasks Allocation KANT2002 Kant Swaroop, P. K. Gupta
Based on Reliability and and Man Mohan,
cost, Proceedings of Operations Research,
„International Conference Sultan Chand & Sons,
on Mathematical Modeling, ISBN No.: 8180540200,
Roorkee; (2001), 150-155. 2002.
BRIA2003 Brian Ensink, Joel Stanley, KENT2002 Kent Fitzgerald, Shahram
and Vikram Adve,. Latifi, and Pradip K.
Program Control Srimani, Reliability
Language: a programming Modeling and Assessment
language for adaptive of the Star-Graph
distributed applications; Networks, IEEE
Elsevier Inc., Vol. Transactions on Reliability,
63(2003), No. 12, 1082 – Vol. 51(2002), 49-57.
1104. KUMA1995a A. Kumar, M. P. Singh and
DANI2005 Daniel Grosu, and Anthony P. K. Yadav, An Efficient
T. Chronopoulos, Algorithm for Allocating
Noncooperative load Tasks to Processors in a
balancing in distributed Distributed System,
systems; Elsevier Inc., Vol. Proceedings of the „19th
65(2005), No. 9, 1022- National System
1034. Conference, SSI,
DESS1980 O.I. Dessoukiu-EI, and W. Coimbatore, (1995), 82-87.
H. Huna, Distributed KUMA1995b A. Kumar, M. P. Singh, and
Enumeration on Network P. K. Yadav, A Fast
Computers; IEEE Algorithm for Allocating
Transactions on Computer, Tasks in Distributed
Vol. 29(1980), 818-825. Processing System,
JACQ2005 Jacques Bahi, Raphael Proceedings of the „30th
Couturier, and Flavien Annual Convention of CSI,
Vernier, Synchronous Hyderabad, (1995), 347-
distributed load balancing 358.
on dynamic networks. KUMA1996 A. Kumar, M. P. Singh, and
Elsevier Inc., Vol. P. K. Yadav, An Efficient
65(2005), No. 11, 1397- Algorithm for Multi-
1405. processor Scheduling with
JAVI2000 Javier Contreras, Arturo Dynamic Reassignment,
Losi, Mario Russo, and Proceedings of the „6th
Felix F. Wu, DistOpt: A National seminar on
Software Framework for theoretical Computer
Modeling and Evaluating Science, Banasthally
Optimization Problem Vidyapeeth, (1996), 105-
Solutions in Distributed 118.

COPO3O1-7
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

KWON2006 Yu-Kwong Kwok, Anthony Vol. 51(2002), 133-140.

A. Maciejewski, Howard SAEE2005 Saeed Iqbal, and Graham F.
Jay Siegel, Ishfaq Ahmad, Carey, Performance
and Arif Ghafoor, A semi- analysis of dynamic load
static approach to mapping balancing algorithms with
dynamic iterative tasks onto variable number of
heterogeneous computing processors. Elsevier Inc.,
systems. Elsevier Inc., Vol. Vol. 65(2005), No. 8, 934-
66(2006) No. 1, 77-98. 948.
MISR1991 K. B. Misra, and U. SAGA1991 G. Sagar and A. K. Sarje,
Sharma, An Efficient Task Allocation Model for
Algorithm to solve Integer Distributed System,
Programming Problem International Journal of
arising in System System Science, Vol.
Reliability Design, IEEE 22(1991), 1671-1678.
Transactions on Reliability, SHU2009 Shu Wanneng, Wang
Vol. 40(1991), 81-91. Jiangqing, Min-Min
MUHA2002 Muhammad K. Dhodhi, Chromosome, Genetic
Imtiaz Ahmad, Anwar Algorithm for Load
Yatama and Ishfaq Ahmad. Balancing in Grid
An Integrated Technique Computing, International
for Task Matching and journal of Distributed
Scheduling onto sensor network, Vol.
Distributed Heterogeneous 5(2009). No. 1, 62-63.
Computing Systems
Elsevier Inc., Vol. 62(2002)
No. 9, 1338 – 1361.
PALM2005 J. Palmer, and I. Mitrani,
Optimal and heuristic
policies for dynamic server
allocation. Elsevier Inc.,
Vol. 65(2005) No. 10,
1204-1211.
RENE2002 Rene L. Bierbaum, Thomas
D. Brown, and Thomas J.
Kerschen, Model-Based
Reliability Analysis, IEEE
Transactions on Reliability,

COPO3O1-8
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Analysis of Bio-inspired Algorithms–

BFO, PSO and FFA
Vipul Singhal
Vipul.vanshi@gmail.com

Delhi College of Engineering, Delhi University, Delhi, India

ABSTRACT emergence. Advancements and research areas then

Engineering problems can be defined as search for
one near optimal description among many
possibilities, under real time constraints. Search
techniques such as Bacterial foraging, Particle swarm
and Firefly based optimization algorithms are now
among the latest research topics. This paper throws
light on these bio-inspired evolutionary algorithms
that are used to optimize the objective function, given
constraints. Optimization as an engineering problem
has been exposed as their one application along with
the several advancements made in this field. These
techniques are explained in detail along with work
and research that has been done in that area. It also
provides the insight of modification and
improvements, these algorithms continue to enjoy
performing better for specified problem.
1. INTRODUCTION
exposed.
2. BACTERIAL FORAGING OPTIMIZATION:
Bacterial foraging algorithm is inspired by the
pattern exhibited by bacterial foraging behaviours
[1]. They tend to maximise energy per unit time.
Bacteria have the tendency to gather to the nutrient-
rich areas by an activity called “chemotaxis”. It is
known that bacteria swim by rotating whip-like
flagella driven by a reversible motor embedded in
the cell wall. E. coli has 8-10 flagella placed
randomly on a cell body. When all flagella rotate
counter clockwise, the move is called Run. When
the flagella rotate clockwise, they all pull on the
bacterium in different directions, which causes the
bacteria to Tumble.

Optimization has been an active area of research

for several decades. As many real-world
optimization problems become increasingly
complex, better optimization algorithms are
needed. Over the past few decades, many
biologically inspired computational methodologies
have been invented, such as evolutionary
algorithms which include genetic algorithms,
bacterial foraging, particle swarm firefly algorithms
and many others. These tend to enjoy a regular
attention by scientist in several research areas of FIG. 1.1 SWIM AND TUMBLE OF A BACTERIUM [1]
.
soft computing and others as well.
The foraging algorithm is based on the behaviour
Here in this paper we have taken review of three
of the E. coli bacteria, which has tree fundamental
basic algorithms and have seen how they can be
steps: chemotaxis, reproduction, and elimination
applied to several problems. Section 1 contributes
and dispersion. Chemotaxis is used in order to
to explain basic and classic algorithm of BFA
move and search for food, this step has much
which is then extended to provide its further
resemblance with a biased random-walk model [2].
applications and research areas. It tells about how
In order to maximise energy in minimum time
the bacteria forage by maximising their energy
bacteria tend to follow different search strategies as
intake in minimum span of time. This concept then
described in fig. 1.2
is used to keep them moving and later reproduce to
give next generation. It also provides that those
with poor foraging strategy tend to eliminate.
Section 2 explains basics and classic particle
swarm optimization technique. It also throws light
on the work done in that area and its applications.
Section 3 then explains firefly algorithm and its

COPO1O2-1
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
FIGURE 1.2 SEARCH STRATEGIES FOR FORAGING
BACTERIA [1]
The chemotactic operator employed in BFOA is
supposed to guide the swarm to converge toward
optima. Reproduction is used to make an identical
copy; and elimination and dispersion is used to
avoid noxious elements that can kill or spread a
group of bacteria in the search space.
2.1. CLASSICAL ALGORITHM:
When bacteria get food in sufficient amount, they
increase in length, and in presence of suitable
temperature, they break in the middle to form an
exact replica of themselves. This phenomenon
inspired Passino[1] to introduce an event of
reproduction in BFOA. Due to the occurrence of
sudden environmental changes or attack, the
chemotactic progress may halt, and a group of
bacteria may move to some other places. This
constitutes the event of elimination–dispersal in the
real bacterial population, where all the bacteria in a
region are killed or a group is dispersed into a new
part of the environment. As said, bacteria try to
move to areas which are rich in nutrients and free
of noxious substances [4]. If θ is a position and J(θ)
is a function which is the more negative the more
the nutrients prevail over the noxious substances,
then we can imagine that bacteria naturally try to
find those, where J (θ) has its minima.
A bacterium position after a tumble can be
determined through (1), where the position at a
given instant is calculated in terms of the position
at the previous instant and the step size C(i) applied
in a random direction ϕ (j), generated by the
bacterium tumble
θi(j+1,k,l)= θi(j,k,l)+C(i)* ϕ(j,k,l) (1)
Where i indicate the bacterium in a total population
of S individuals, j the chemotactic step, the
reproductive step k, and the elimination and
dispersal step l.
The health function is calculated at these positions
of bacteria as (2)
J(i,j,k,l)=J(i,j,k,l)+Jcc(θ’(j,k,l),P(j,k,l)). (2)
Here second expression shows the effect of
attraction and repulsion among the bacteria. The
exact expression of Jcc(θ’(j,k,l),P(j,k,l)) is in [1].
And this then is compared with the previous health
function as
J (θi(j+1,k,l))<J(θi(j,k,l))
If it’s true then the bacterium will keep moving in
the same direction (“run” behaviour) for a given
maximum number of steps, after which a “tumble”
will occur anyway. The least healthy bacteria
eventually die while each of the healthier bacteria
(those yielding lower value of the objective
function) asexually split into two bacteria, which
are then placed in the same location. This keeps the
swarm size constant. This is done through
reproduction step. Some bacteria are liquidated at
random with a very small probability while the new
replacements are randomly initialized over the
search space.
2.2. PSEUDOCODE:
1. Initialize parameters n, N, Nc, Ns, Nre, Ned,
Ped,
C(i)(i=1,2…N), i =1,2,---
Where,
n: Dimension of the search space,
N: The number of bacteria in the population,
NC : Chemotactic steps,
Nre: The number of reproduction steps,
Ned : the number of elimination-dispersal events,
Ped: Elimination-dispersal with probability,
C (i): the size of the step taken in the random
direction specified by the tumble.
2. Elimination-dispersal loop: l=l+1
3. Reproduction loop: k=k+1
4. Chemotaxis loop: j=j+1
[a] For i =1,2…N, take a chemotactic step for
bacterium i as follows.
[b] Compute fitness function, J (i, j, k, l).
Let,J (i, j, k, l) =J (i, j, k, l)+Jcc(θi( j, k, l), P( j, k, l))
(i.e. add on the cell-to cell attractant–repellent
profile to simulate the swarming behaviour )
.Where, Jcc is defined in .
[c] Let Jlast=J (i, j, k, l) to save this value since we
may find a better cost via a run.
[d] Tumble: generate a random vector Δ(i)€ Rn
with each element ∆m(i),m =1,2,..., p, a random
number on
[-1, 1].
[e] Move:
Let
COPO1O2-2
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
This results in a step of size C(i) in the direction of
the tumble for bacterium i.
[f] Compute J (i, j +1, k, l) and
Let J(i, j,k,l)=J(i, j,k,l)+Jcc( θi( j,k,l),P( j,k,l))
[g] Swim
i) Let m=0 (counter for swim length).
ii) While m<Ns(if have not climbed down too long).
• Let m=m+1.
• If J (i, j +1, k, l)<Jlast( if doing better), let
Jlast= J (i, j+1, k, l) and
Let
And use this θi(i+1, j, k) to compute the new J (i,
j+1, k, l) as we did in [f]
Else, let m= Ns. This is the end of the while
statement.
[h] Go to next bacterium (i, 1) if i≠N (i.e., go to [b]
to process the next bacterium).
5. If j<Nc, go to Step 3. In this case, continue
chemotaxis, since the life of the bacteria is not
over.
6. Reproduction:
[a] For the given k and l, and for each i =1,2,...,N,
Let
be the health of the bacterium i (a measure of how
many nutrients it got over its lifetime and how
successful it was at avoiding noxious substances).
Sort bacteria and chemotactic parameters C(i) in
order of ascending cost health J (higher cost means
lower health).
[b] The Sr bacteria with the highest Jhealth values die
and the remaining Sr bacteria with the best values
split (this process is performed by the copies that
are made are placed at the same location as their
parent).
7. If k<Nre, go to Step 3. In this case, we have not
reached the number of specified reproduction steps,
so we start the next generation of the chemotactic
loop.
8. Elimination-dispersal: For i=1,2...,N, with
probability Ped, eliminate and disperse each
bacterium, and this result in keeping the number of
bacteria in the population constant. To do this, if a
bacterium is eliminated, simply disperse one to a
random location on the optimization domain. If
l<Ned , then go to Step. 2 ; otherwise end.
2.3. ADVANCEMENTS IN BFO AND ITS APPLICATION AND
RESEARCH AREAS:
Vast applications have been found where BFO has
shown remarkable results and has been modified
for different problems according to the objective
function. Initial applications of evolutionary
algorithm were meant for static optimization
problems but in recent years the emergence of
another member of the EA family[5]– bacterial
foraging algorithm (BFA), the self adaptability of
individuals in the group searching activities has
attracted a great deal of interests including dynamic
problems. W. J. Tang and Q. H. Wu have
contributed their work by proposing DBFA, which
is especially designed to deal with dynamic
optimization problems, combining the advantage of
both local search in BFA and a new selection
scheme for diversity generating. They used the
moving peaks benchmark (MPB) [6] as the test bed
for experiments. The performance of the DBFA is
evaluated in two ways. The first is concerned with
the convergence of the algorithm in random
periodical changes in an environment, which are
divided into three ranges from a low probability of
changes to a higher one. The second is testing a set
of combinations of the algorithm parameters which
are largely related to the accuracy and stability of
the algorithm. All results are compared with the
existing BFA [1], and show the effectiveness of
DBFA for solving dynamic optimization problems.
It is worth mentioning that the diversity of DBFA
changes after each chemotactic process rather than
the dispersion adopted by the BFA after several
generations. The DBFA utilizes not only the local
search but also applies a flexible selection scheme
to maintain a suitable diversity during the whole
evolutionary process. It outperforms BFA in almost
all dynamic environments. The results are shown in
[5]. They have further given solution for global
optimization given in [7].
The novel BSA has been proposed for global
optimization. In this algorithm, the adaptive tumble
and run operators have been developed and
incorporated, which are based on the understanding
of the details of bacterial chemotactic process. The
operators involve two parts: the first is concerned
with the selections of tumble and run actions, based
on their probabilities which are updated during the
searching process; the second is related to the
length of run steps, which is made adaptive and
independent of the knowledge of optimization
problems. These two parts are utilized to balance
the global and local searching capabilities of BSA.
Beyond the tumble and run operators, attraction
and mutation operations have also been developed.
A.ABRAHAM, A. BISWAS, S. DASGUPTA AND S.
DAS have shown [8] that the major driving forces
of Bacterial Foraging Optimization Algorithm
(BFOA) is the reproduction phenomenon of virtual
bacteria each of which models one trial solution of
the optimization problem.
BFO and PSO have been used in combination and
their combined performance has been utilised to
incorporate the merits [9] of two bio-inspired
algorithms to improve the convergence for high-
dimensional function optimization. It is assumed
COPO1O2-3
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
that the bacteria have the similar ability like birds
to follow the best bacterium (bacterium with the
best position in the previous chemotactic process)
in the optimization domain. The position of each
bacterium after every move (tumble or run) is
updated according to (3)
θi(j+1,k,l)=θi(j+1,k,l)+Ccc(θb(j,k,l)-θi(j,k,l), (3)
if Ji(j+1,k,l)>Jmin(j,k,l)
where θb(j,k,l) and Jmin(j,k,l) are the position and
fitness value of the best bacterium in the previous
chemotactic process respectively, Ccc is a new
parameter, called attraction factor, to adjust the
bacterial trajectory according to the location of the
best bacterium.
3. PARTICLE SWARM OPTIMIZATION:
Particle swarm optimization is a high-performance
optimizer that is very easy to understand and
implement. It is similar in some ways to genetic
algorithms or evolutionary algorithms, but requires
less computational bookkeeping and generally only
a few lines of code [10]. Particle swarm
optimization originated in studies of synchronous
bird flocking and fish schooling, when the
investigators realized that their simulation
algorithms possessed an optimizing
characteristic[11]-[13]. As the particles traverse the
problem hyperspace, each particle remembers its
own personal best position that it has ever found,
called its local best. Each particle also knows the
best position found by any particle in the swarm,
called the global best. Overshoot and undershoot
combined with stochastic adjustment explore
regions throughout the problem hyperspace,
eventually settling down near a good solution. This
process can be visualized as a dynamical system,
although the behaviour is extraordinarily complex
even when only a single particle is considered with
extremely simplified update rules. This new
optimization technique has much promise, and
electromagnetic researchers are just beginning to
explore its capabilities.
3.1. CLASSICAL ALGORITHM:
The particle swarm concept originated as a
simulation of a simplified social system. The
original intent was to graphically simulate the
graceful but unpredictable choreography of a bird
flock. Initial simulations were modified to
incorporate nearest-neighbour velocity matching,
eliminate ancillary variables, and incorporate
multidimensional search and acceleration by
distance (Kennedy and Eberhart 1995, Eberhart and
Kennedy 1995). At some point in the evolution of
the algorithm, it was realized that the conceptual
model was, in fact, an optimizer. Through a process
of trial and error, a number of parameters
extraneous to optimization were eliminated from
the algorithm, resulting in the very simple original
implementation (Eberhart, Simpson and Dobbins
1996).
PSO emulates the swarm behaviour and the
individuals represent points in the -dimensional
search space. A particle represents a potential
solution. The velocity Vid and position Xid of the dth
dimension of the ith particle are updated as follows
(4) & (5).
Vid ← Vid + C1* rand1id *(pbestid - Xid) + C2*
rand2id *(gbestd - Xid) (4)
Xid ← Xid + Vid (5)
Where Xi = (Xi1, Xi2, … , XiD) is the position of ith
particle Vi = (Vi1, Vi2, … , ViD) represents velocity
of the particle i. pbest =( pbesti1, pbesti2, … ,
pbestiD) is the best previous position yielding the
best fitness value for their ith particle; and gbest =
(gbesti1, gbesti2, … , gbestiD) is the best position
discovered by whole population[14]. C1 and C2 and
are the acceleration constants reflecting the
weighting of stochastic acceleration terms that pull
each particle toward pbest and gbest positions,
respectively. rand1id and rand2id are two random
numbers in the range [0, 1].
3.2. PSEUDOCODE:
1:Generate the initial swarm by randomly
generating the position and velocity for each
particle;
2: Evaluate the fitness of each particle;
3: repeat
4: for each particle i do
5: Update particle i according to (1) and (2);
6: if (xi) <f(xpbesti) then
7: xpbesti:=xi;
8: if f (xi) <f(xgbest) then
9: xgbest := xi
10: end if
11: end if
12: end for
13: until the stop criterion is satisfied
3.3. ADVANCEMENTS IN PSO AND ITS APPLICATION AND
RESEARCH AREAS:
APPSO (Agent based parallel PSO) is based on two
types of agents: one coordination agent and several
swarm agents. The swarm is composed of various
COPO1O2-4
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
sub-swarms, one for each swarm agent. The
coordination agent has administrative and
managing duties. All the calculations are done by
the swarm agents (see Figure 2).
In order to gain benefit from the large knowledge
and insights achieved in the research field of
sequential PSO it is important to modify the
swarm’s behavior as little as possible. The
inevitable changes to the algorithm due to the
parallelization should also lead to positive effects
FIG. 2 STRUTURE OF APPSO NETWORK SYSTEM [14]
with respect to solution quality. This is realized by
using strategically niching where according to the
requirements the sub-swarms either work together
searching for one global optimum or spread over
the search space to find various global/local
optima, e.g. on multimodal functions.
To address the problem of space locus searching, a
slowdown particle swarm optimization (SPSO) is
proposed to improve the convergence performance
of particle swarm from the position viewpoint. The
particle swarm in SPSO is divided into many
independent sub-swarms to guarantee that particles
converge to different position, since space locus
has multiple optimal solutions and requires the
convergence of both fitness and position of particle.
Furthermore, particle velocity is updated by half
according to fitness to achieve the position
convergence [14].
Several PSO algorithms have been recently
proposed to address DOPs of which using multi
swarms seems a good technique. The multi swarm
method can be used to enhance the diversity of the
swarm, with the aim of maintaining multiple
swarms on different peaks.
One of the reasons that particle swarm optimization
is attractive is that there are very few parameters to
adjust. One version, with very slight variations (or
none at all) works well in a wide variety of
applications. Particle swarm optimization has been
used for approaches that can be used across a wide
range of applications, as well as for specific
applications focused on a specific requirement. In
this brief section, we cannot describe all of particle
swarm’s applications, or describe any single
application in detail. Rather, we summarize a small
sample.
Generally speaking, particle swarm optimization,
like the other evolutionary computation algorithms,
can be applied to solve most optimization problems
and problems that can be converted to optimization
problems. Among the application areas with the
most potential are system design, multi-objective
optimization, classification, pattern recognition,
biological system modelling, scheduling
(planning), signal processing, games, robotic
applications, decision making, simulation and
identification. Examples include fuzzy controller
design, job shop scheduling, real time robot path
planning, image segmentation, EEG signal
simulation, speaker verification, time-frequency
analysis, modelling of the spread of antibiotic
resistance, burn diagnosing, gesture recognition
and automatic target detection, to name a few.
4. FIREFLY OPTIMIZATION
Fireflies, also called lighting bugs, are one of the
most special and fascinating creatures in nature.
These nocturnal luminous insects of the beetle
family Lampyridae (order Coleoptera), inhabit
mainly tropical and temperate regions, and their
population is estimated at around 1900 species.
They are capable of producing light thanks to
special photogenic organs situated very close to the
body surface behind a window of translucent
cuticle. Bioluminescent signals are known to serve
as elements of courtship rituals, methods of prey
attraction, social orientation or as a warning signal
to predators (in case of immature firefly forms
commonly referred to as glow worms). The
phenomenon of firefly glowing is an area of
continuous research considering both its
biochemical and social aspects.
Mechanisms of firefly communication via
luminescent flashes and their synchronization has
been imitated effectively in various techniques of
wireless networks design, dynamic market pricing
and mobile robotics.
4.1. CLASSICAL ALGORITHM
The flashing light of fireflies is an amazing sight in
summer sky in the tropical and temperate regions.
There are about two thousand firefly species, and
most fireflies produce short and rhythmic flashes.
The pattern of flashes is often unique for a
particular species. The flashing light is produced by
COPO1O2-5
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
a process called bioluminescence. However, two
fundamental functions of such flashes are to attract
mating partners and to attract potential prey. In
addition it may also serve as a protective warning
mechanism to remind potential predators of the
bitter taste of fireflies.
The rhythmic flash, the rate of flashing and the
amount of time form the part of the signal system
that bring both sexes together. We know that the
light intensity at a particular distance r from the
light source obeys the inverse square law. That is to
say, the light intensity I decreases as the distance r
increases in terms if I . Furthermore, the air
absorbs light which becomes weaker as the
distance increases. These two combined factors
make most fireflies visual to a limit distance,
usually several hundred meters at night, which is
good enough for fireflies to communicate.
The flashing light can be formulated in such a way
that it is associated with the objective function to
be optimized, which makes it possible to formulate
new optimization algorithms.
Assumptions:
1. All fireflies are unisex so that one firefly will be
attracted to the other fireflies regardless of their
sex.
2. The brighter one will move towards the less
bright one. The attractiveness is proportional to the
brightness and they both decrease as the distance
increase. And if there is no brighter one than a
particular firefly, it will move randomly;
3. The brightness of a firefly is affected or
determined by the landscape of the objective
function.
4.1.1 LIGHT INTENSITY AND THE ATTRACTIVENESS:
In the firefly algorithm, there are two important
issues: the variation of light intensity and
formulation of the attractiveness. For simplicity, we
can always assume that the attractiveness of a
firefly is determined by its brightness which in turn
is associated with the encoded objective function.
In the simplest case particular location x can be
chosen as I(x) . However attractiveness β is
relative.
In simplest form, the intensity I(r) varies according
to inverse square law
I(r) =
Where Is is the intensity at the source. For a given
medium with light absorption coefficient , the light
intensity I varies with the distance r. That is
,
where I0 is the original light intensity. In order to
avoid the singularity at r=0 in the expression Is/r2,
combined effect of both the inverse square law and
the absorption can be approximated as the
following Gaussian form.
I(r) =
Thus of the firefly can be defined as
Where β0 is attractiveness at r=0
The movement of a firefly i is attracted to another
more attractive firefly j is determined by
= + (xi - xj) +α ,
where second term is due to the attraction. The
third term is randomization with α being the
randomization parameter, and is a vector of
random numbers drawn from a Gaussian
distribution or uniform distribution. For eg. The
simplest form is and can be replaced by
rand(1/2).
The algorithm presented here makes use of a
synergic local search. Each member of the swarm
explores the problem space taking into account
results obtained by others, still applying its own
randomized moves as well. The influence of other
solutions is controlled by value of attractiveness. It
can be adjusted by modifying two parameters: its
maximum value β0 and an absorption coefficient γ.
The first parameter describes attractiveness at rj = 0
i.e. when two fireflies are found at the same point
of search space S. In general β0 [0; 1] should be
used.
SPECIAL CASES:
For γ→0, the attractiveness is constant β =β0, this is
equivalent to say that the light intensity does not
decrease in an idealized sky.
When γ→∞ leads to β(r)→δ(r), which means that
the attractiveness is almost zero in the sight of
other fireflies.
.4. 2. PSEUDOCODE:
1. Input:
f(z), z = [z1; z2; :::; zn]T
m, β0, γ, min ui, max ui (parameters)
2. Output:
Ximin
COPO1O2-6
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
Begin
for i=1 to m do
xi← Generate Initial Solution ()
end
3. repeat
imin← arg mini f(xi)
ximin← arg minxi f(xi)
4. for i=1 to m do
for j=1 to m do
if f(xj) < f(xi) then
rj← Calculate Distance (xi , xj)
β←β0e-γrj
ui← Generate Random Vector (min ui, max ui)
for k=1 to n do
xi,k ←(1- β)xi,k + βxi,k + ui,k
end
end
end
end
5. uimin← Generate Random Vector (min ui, max ui)
for k=1 to n do
ximin,k← ximin,k + uimin,k
end
until stop is condition true
6.end
4.3. ADVANCEMENTS IN FFA AND ITS APPLICATION AND
RESEARCH AREAS:
Many bio-inspired algorithm exist that use simple
rules and local information to create a global
consensus of a single parameter, such as clock time
in the flashing firefly algorithm, or location in the
slime algorithm. In this paper we identify a control
loop as the core structure of several bio-inspired
algorithms. Two ways of combining these loops are
identified, serially, and nested. These methods of
combination can be used to combine algorithms to
allow the control of multiple parameters (one per
algorithm) on a global level without the use of a
centralised point of control.
Thresholding is one of the most important
techniques for performing image segmentation. It is
generally simple and computationally efficient. The
MEFFT algorithm simulates the behaviour of
fireflies and the phenomenon of bioluminescent
communication to develop the algorithm to select
the adequate thresholds for image segmentation.
The segmentation results of MEFFT algorithm are
promising and it encourages further researches for
applying this algorithm to complex and real-time
image analysis problems such as target recognition,
complex document analysis and biomedical image
application [16]. The other application include
synchronisation of wireless network which be seen
in [17]. An application in pulse-coupled oscillator’s
model addresses the issue of synchronization of the
oscillators with different frequencies. As a
continuation of this research, it is important to
study how the topology and communication delay
of the distributed oscillators system affect the
synchronization process.
5. CONCLUSION
While studying and working on these algorithms
we have seen that performance of any technique is
objective dependent. There are various parameters
that define the aptness of the algorithm. So it
depends on the requirements of the objective
function and on the constraints imposed. For
example if we take several benchmark problems
and compare the results of three algorithms then it
has been seen each would perform differently. Not
only their individual performance vary even they
tend to become less or more stable in terms of the
parameters defining the quality of performance
such as elapsed time, convergence rate, optimum
value (maximised/minimized), mean and standard
deviation etc. So depending upon our requirement
we may define our objective function and the
performance of algorithm for that particular
function is based on how aptly the required
objective is received, which could be in terms of
final value, elapsed time, mean etc.
Thus having witnessed these variations one can
generalise the performance on overall terms and
state any of the algorithm as best among the three
taken.
6. REFERENCES
[1] K. M. PASSINO, “Biomimicry of bacterial foraging for
distributed optimization and control,” IEEE Control Syst., vol.
22, no. 3, pp. 52–67, Jun. 2002.
[2]. S. DAS, S. DASGUPTA, A. BISWAS and A. ABRAHAM,
“On Stability of the Chemotactic Dynamics in Bacterial-
Foraging Optimization Algorithm.” IEEE Trans. on Systems,
Man, and cybernetics, part: A: Systems and Humans, vol. 39,
no. 3, pp. 670-679, May 2009.
[3]. B. D. Hughes, Random Walks and Random Environments.
London, U.K.: Oxford Univ. Press, 1996.
[4]. L. DOS SANTOS COELHO, C. DA COSTA SILVEIRA,
C.A. SIERAKOWSKI and P. ALOTTO, “Improved Bacterial
Foraging Strategy Applied to TEAM Workshop Benchmark
Problem.” IEEE Trans. On Magnetics, Vol. 46, no. 8, pp. 2903-
2906, Aug. 2008.
[5] W. J. TANG and Q. H. WU, “Bacterial Foraging Algorithm
For Dynamic Environments.” Proc. of IEEE Congress on
Evolutionary Computation Sheraton Vancouver Wall Centre
Hotel, Vancouver, BC, Canada, pp. 1324-1330, July 16-21,
2006.
[6] R. W. MORRISON and K. A. DE JONG, “A test problem
generator for non stationary environments”, Proc. of IEEE
Congress on Evolutionary Computation, IEEE Press, pp. 2047–
2053, 1999.
[7]. W. J. TAN and Q. H. WU, “A Bacterial Swarming
Algorithm For Global Optimization.” Proc. of IEEE Congress
on Evolutionary Computation, pp. 1207-1212, 2007.
COPO1O2-7
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
[8]. A. ABRAHAM, A. BISWAS, S. DASGUPTA AND
SWAGATAM DAS, “Analysis of Reproduction Operator in
Bacterial Foraging Optimization Algorithm.” Proc. of IEEE
Congress on Evolutionary Computation, pp. 1476-1483, 2008.
[9]. YING CHU, HUA MI, HUILIAN LIAO, ZHEN JI, AND
Q. H. WU, “A Fast Bacterial Swarming Algorithm For High-
dimensional Function Optimization.” Proc of IEEE Congress on
evolutionary computation, pp. 3135-3140, 2008.
[10]. DANIEL W. BOERINGER AND DOUGLAS H.
WERNER, “Particle Swarm Optimization Versus Genetic
Algorithms for Phased Array Synthesis.” Trans. Of IEEE on
Antennas and Propagation, vol. 52, no. 3, pp. 771-779,March
2004.
[11] R. EBERHART AND J. KENNEDY, “A new optimizer
using particle swarm theory,” in Proc. 6th Int. Symp. Micro
Machine and Human Science (MHS ’95), pp. 39–43, 1995.
[12] J. KENNEDY AND R. EBERHART, “Particle swarm
optimization,” in Proc. IEEE Int. Conf. Neural Networks, vol. 4,
pp. 1942–1948, 1995.
[13]. J. J. LIANG AND A. K. QIN, “Comprehensive Learning
Particle Swarm Optimizer for Global Optimization of
Multimodal Functions.” Trans of IEEE on Evolutionary
Computation, vol. 10, no. 3, June 2006.
[14]. LI ZHI-JIE, LIU XIANG-DONG, DUAN XIAO-DONG,
WANG CUN-RUI, “An Improved Particle Swarm Algorithm
for Search Optimization.” Proc. of IEEE Global Congress on
Intelligent System, pp.154-158, 2009.
[15]. MICHAEL BREZA, JULIE MCCANN, “Can Fireflies
Gossip and Flock?: The possibility of combining well-knowbio-
inspired algorithms to manage multiple global parameters in
wireless sensor networks without centralised control.
[16]. MING-HUWI HORNG AND TING-WEI JIANG,
“Multilevel Image Thresholding Selection based on the Firefly
Algorithm.” Symposia and Workshops on Ubiquitous,
Autonomic and Trusted Computing, pp. 58-63, 2010.
[17]. LIN CUI, HONGPENG WANG, “Reachback Firefly
Synchronicity with Late Sensitivity Window in Wireless Sensor
Networks.” Proc. of IEEE Ninth International Conference on
Hybrid Intelligent Systems, pp. 451-456, 2009.
COPO1O2-8
 BIOMETRIC AUTHENTICATION USING IMAGE PROCESSING TOOLS
Neha Mittal Ruchi Gupta
Department of Electronics and communication
J.P Institute of Engineering & Technology, Meerut
Email: nehamittal2008@gmail.com
Abstract
This paper presents the study of biometric
authentication using various digital image
processing techniques and some future
possibilities in this field. The position of
biometrics in the current field of security has
been depicted in this work. For biometrics
recognition the Region of interest (ROI) is
extracted from the biometric modalities.
Various different features are extracted from
ROI. These features can be fuzzy features,
wavelet features, Energy and entropy
features. These sets of features are used for
the authentication of users using Euclidean
Distance, Chi square measure and fuzzy
Support Vector Machines as classifiers.
Recognition score of 82% is obtained with
wavelet feature on palmprint database of
PolyU. We have also outlined the
comparative study and usability of biometric
authentication systems in this paper.
Keywords: wavelet features, multimodal
biometric, ROI, Euclidean Distance, chi-
square,
1. INTRODUCTION
Computer aided personal authentication is
becoming more and more reliable in this
information era. As in order to deal with
security, Authentication plays an important
role. Biometrics is one of the most reliable in
this field. A wide variety of systems require
reliable personal recognition schemes to
either confirm or determine the identity of an
individual requesting their services. The
purpose of such schemes is to ensure that the
rendered services are accessed only by a
genuine user, and not anyone else. By using
Pratibha Singh
biometrics it is possible to confirm or
establish an individual’s identity.
The most widely used biometric feature is the
fingerprint while most reliable one is the
iris[1].Palmprint, as a biometric feature
gaining acceptance in the field of biometric
recognition compared with other modalities:
low cost capture device, easy self positioning,
cannot be easily stolen, large area to get large
information.
In this paper, an authentication of biometrics
is studied. Selection of region of interest,
feature extraction, and classification of these
features is explored.
The organization of the paper is as follows:
Section 2 presents the extraction of ROI. The
extraction of features is described in Section
3. Results are described in Section 4 and
finally multimodal biometrics systems and
conclusions are drawn in Section 5 and 6.
2. ROI EXTRACTION
The useful information content of an image of
any biometric modality can be abstracted by
different image parameters; this is nothing but a
region of interest. ROI is simply a subimage of
interest. An algorithm for abstracting ROI of
Palmprint is given in [8].
COPO1O1-1
 BIOMETRIC AUTHENTICATION USING IMAGE PROCESSING TOOLS
Fig. 1: A typical Palmprint Fig. 2: The extracted ROI of
palmprint
3. FEATURE EXTRACTION
The issue of choosing the features to be
extracted should be guided by some concerns
like the features should carry enough
information about the image and should not
require any domain-specific knowledge for
their extraction. They should be easy to
compute in order for the approach to be
feasible for a large image collection and rapid
retrieval.
Finger print matching techniques can be
placed into two categories. One of them is
Minutiae based and the other one is
Correlation based. Minutiae based techniques
find the minutiae points first and then map
their relation placement on the finger.
Correlation based techniques require the
precise location of a registration point and are
affected by image translation and rotation [6,
11,12].
The iris of the eye is colored area surrounded
by the pupil. Iris patterns are unique and are
obtained through video based image
acquisition system. Each iris structure is
featuring a complex pattern. This can be a
combination of specific characteristics known
as corona, crypts, filaments, freckles, pits,
furrows, striations and rings [4].
The palmprints possess random textures. A
palmprint is a rich source of texture. It
consists of zigzag and crisscrossing lines in
addition to three principal lines. Palmprint
also contains creases, wrinkles and ridges.
Texture features can be extracted using
statistical measures (entropy, homogeneity),
wavelet, and fractals.
In [8] the ROI is divided in to non-
overlapping windows. First the membership
function is calculated and from this
membership function different features (i.e.
energy feature, fuzzy feature, Entropy feature)
can be calculated. Some features are as
follows:
3.1 Wavelet Features
Wavelets are immensely suitable for texture
analysis. In [1] and [8] Haar wavelets are
effective on the palmprints. Simplicity of
these wavelets is an overriding factor in the
choice of these wavelets though other
celebrated wavelets such as Daubechies
wavelets, Symlets are also tried but the
improvement accrued from the latter ones is
not proportionate with the effort.
A two-dimensional discrete wavelet transform
is applied on the ROI of a palmprint resulting
in four subbands: HL1, LH1 and HH1, which
correspond to fine-level coefficients and LL1,
which corresponds to coarse-level
coefficients. The HL1, LH1 and HH1together
contribute to a detail image. Each detail
image represents a high frequency structure,
i.e. lines, creases, wrinkles, ridges on a
palmprint. Fig. 3 shows one level of
decomposition. LL1 can be further
decomposed into four sub bands and the
process is repeated till the information from
LLn at the nth decomposition is insignificant.
In any decomposition, the detail images
include horizontal, vertical and diagonal parts
COPO1O1-2
 BIOMETRIC AUTHENTICATION USING IMAGE PROCESSING TOOLS
of the decomposed image. These three detail
images are superimposed in [8] to yield
composite mage. It may be noted that as the
decomposition level increases the size of the
detail images also decreases.
The results of authentication correspond to
the composite image.
From each detail images, energy feature is
calculated as [1]:
M N
(1)
Eid ( Si ( x, y )) 2 , i 1, 2,..., 5
x 1 y 1
Fig. 3: Two dimensional one level DWT
decomposition
Fig. 4: Three-level DWT decomposition
where i is the decomposition level. Hi,Vi and
Di are the detail images in horizontal, vertical
and diagonal directions respectively. Other
features are under investigation.In [8] all the
detail images are super imposed and then
energy is calculated.
4. Results and implementation
In [8] a 100% recognition score is obtained
using fuzzy feature. A simple Euclidean
distance measure is used to find out the
recognition rate. In [8] the database used, is
created in the biometrics lab of IIT Delhi.
ROI is divided in to non-overlapping
windows. Features are calculated from these
windows. The size of the window is varied
and the recognition score is obtained.
Given two data sets of features corresponding
to the training and testing samples, a
matching algorithm determines the degree of
similarity between them. A Euclidean
distance is adopted as a measure of
dissimilarity for the palmprint matching using
both wavelet and fuzzy features.
The wavelet feature in [8] is applied on polyU
database [15] on 50 users and 5 images of
each user (total 250 images).4 images are
taken for training data and 1 image for testing
data. A recognition score of 82% is obtained.
Euclidean distance measure is taken as a
classifier.ROC plot is shown in fig.5
COPO1O1-3
 BIOMETRIC AUTHENTICATION USING IMAGE PROCESSING TOOLS
Fig 5: ROC plot for PolyU database
5. MULTIMODAL BIOMETRIC
SYSTEM
Multimodal biometric systems are capable of
utilizing more than one physiological or
behavioral characteristic for authentication. In
order for the biometrics to be ultra-secure and
to provide more-than-average accuracy, more
than one form of biometric identification is
required. Hence the need arises for the use of
multimodal biometrics. This uses a
combination of different biometric
recognition technologies.
In certain situations, the user might find one
form of biometric identification is not exact
enough for identification. This can be the case
with fingerprints, where at least 10% of the
population have worn, cut or unrecognizable
prints.
Multimodal biometric technology uses more
than one biometric identifier to compare the
identity of the person. Therefore in the case of
a system using say three technologies i.e. face
mimic and voice. If one of the technologies is
unable to identify, the system can still use the
other two to accurately identify against.
By using more than one means of biometric
identification, the multimodal biometric
identifier can retain high threshold
recognition settings. The system administrator
can then decide the level of security he/she
requires. For a high security site, they might
require all three biometric identifiers to
recognize the person or for a lower security
site, only one or two of the three. With this
methodology, the probability of accepting an
impostor is greatly reduced.
The unimodal biometric systems can have the
variety of limitations which are as follows
[13]:
(a) Noise in sensed data: A fingerprint
image with a scar or a voice sample
altered by cold, are examples of noisy
data. Noisy data could also result from
defective or improperly maintained
sensors (e.g., accumulation of dirt on a
fingerprint sensor) or unfavorable ambient
conditions (e.g., poor illumination of a
user’s face in a face recognition system).
(b) Intra-class variations: These variations
are typically caused by a user who is
incorrectly interacting with the sensor
(e.g., incorrect facial pose), or when the
characteristics of a sensor are modified
during authentication (e.g., optical versus
solid-state fingerprint sensors).
(c) Inter-class similarities: In a biometric
system comprising of a large number of
users, there may be inter-class similarities
(overlap) in the feature space of multiple
users.
(d) Non-universality: The biometric system
may not be able to acquire meaningful
biometric data from a subset of users. A
fingerprint biometric system, for example,
may extract incorrect minutiae features
COPO1O1-4
 BIOMETRIC AUTHENTICATION USING IMAGE PROCESSING TOOLS

from the fingerprints of certain Multimodal biometric system also prevent

individuals, due to the poor quality of the from spoofing as it is difficult for an impostor
ridges. to spoof multiple biometric modalities of a
(e) Spoof attacks: This type of attack is genuine user simultaneously.
especially relevant when behavioral traits
such as signature or voice are used. References
However, physical traits such as
fingerprints are also susceptible to spoof [1] Xiang-Qian Wu, Kuan-Quan Wang, David
attacks. These systems are able to meet Zhang, ”Wavelet Based Palmprint
the rigorous performance requirements Recognition”, First international
imposed by various applications. They conference on Machine learning
address the problem of non-universality, andcybernetics, Beijing, pp-1253-1257, 4-
since multiple traits ensure sufficient 5 Nov 2002.
population coverage. They also deter [2] R. C. Gonzalez, R. E. Woods, Digital
spoofing since it would be difficult for an Image Processing, Addison Wesley
impostor to spoof multiple biometric traits publishers, 1993.
of a genuine user simultaneously. [3] L. Hong and A. K. Jain, “Integrating Faces
and Fingerprints for Personal
Identification”, IEEE Trans. on Pattern
Analysis and Machine
Intelligence, Vol. 20, No. 12, pp. 1295-
1307, Dec1998.
[4] Sanjay R. Ganorkar, Ashok A. Ghatol,
“Iris Recognition: An Emerging Biometric
Technology”, In Proc. of the 6th WSEAS
International Conference on Signal
Fig 6: A multimodal biometric system Processing, Robotics and Automation,
Greece, Feb. 2007, pp. 91 – 96.
Limitations of unimodal biometrics are [5] N. Duta, A.K. Jain, and K.V. Mardia,
overcome by the multimodal biometric “Matching of Palmprint,” Pattern
system. Recognition Letters, Vol.23, No. 4, pp.
477–485, 2001.
6. CONCLUSIONS [6] David Zhang, Wai-Kin Kong, Jane You,
and Michael Wong, “Online Palmprint
Biometric recognition can offer a high degree Identification”, IEEE Trans. Pattern
of security, they are far from perfect solution. Analysis &Machine Intelligence, Vol. 25,
As the fingerprint can give the high accuracy No. 9, pp. 1041 – 1050,September 2003.
of recognition but it have draw backs as [7] Ying-Han Pang, Andrew Teoh Beng Jin,
discussed. So one solution to this problem is David Ngo Chek Ling, "Palmprint based
to integrate two or more modalities, to Cancelable Biometric Authentication
provide high security.
COPO1O1-5
 BIOMETRIC AUTHENTICATION USING IMAGE PROCESSING TOOLS

System", Int. J. of Signal Processing, http://www4.comp.polyu.edu.hk/~biometri

Vol.1, No.2, pp.93-99, 2004. cs/.
[8] M. Hanmandlu, H M Gupta, Neha Mittal,
S. Vasikarla,”An authentication system
based on palmprint”,ITNG-2009,pp-399-
404
[9] Lei Zhang, ZhenhuaGuo, Zhou Wang,
DavidZhang, "Palmprint Verification
Using Complex Wavelet Transform", IEEE
Intl. Conf. on ImageProcessing, San
Antonio, TX, Sept.16-19,2007.
[10] Li Fang, Maylor K.H. Leung , Tejas
Shikhare, Victor Chan, Kean Fatt Choon ,
“Palmprint Classification”, IEEE Int. Conf.
on Systems, Manand Cybernetics, Vol. 4,
8-11 October 2006, pp.2965-2969.
[11] L. Hong and A. K. Jain, “Integrating
faces and fingerprints for personal
identification”, IEEE Trans. Pattern
Anal. Mach. Intell., Volume 20, No. 12,
Dec. 1998, pp. 1295–1307.
[12] R. Cappelli, D. Maio, D. Maltoni, J. L.
Wayman, and A. K. Jain, “Performance
evaluation of fingerprint verification
systems”, IEEE Trans. Pattern Anal. Mach.
Intell., Volume 28, issue 1, Jan. 2006, pp.
3–18.
[13] M. Indovina, U. Uludag, R. Snelick,
A. Mink, and A. K. Jain, “Multimodal
biometric authentication methods: A
COTS approach,” in Proc. of Workshop on
Multimodal User Authentication, (Santa
Barbara, CA), pp. 99–106, Dec 2003.
[14] Wen-Shiung Chen, Yao-Shan Chiang
and Yen-Hsun Chiu, ”Biometric
Verification by Fusing Hand Geometry and
Palmprint”, IIH-MSP 2007
[15] H. K. Polytechnic University,
“Palmprint database”, Biometric Research
Center Website. 2005.

COPO1O1-6
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Medical Image Enhancement In Spatial Domain: Statistical Analysis

Sanjeev Kumar Gupta1, Sonika Singh2& Rahul Singh3
1
Assistant Professor , 2,3 B.Tech Scholar, ECE Department, Faculty of Engineering and
Technology, R.B.S. College, Bichpuri Campus, Agra

ABSTRACT

Image quality and utility become crucial issues for engineers, scientist, doctors, patients,
insurance companies and lawyer whenever there are changes in the technology by which
medical images are achieved. Examples of such changes include analog to digital conversion,
lossy compression for transmission and storage, image enhancement, and computer aided
methodology for diagnosis of disease in medical images. Edit an image so that it is more suitable
for a specific application than the original image is termed as image enhancement technique.
Image is defined as a two dimensional function f(x, y), where x and y are spatial coordinates,
that bears information, which can be generated in any form such as visual, x-ray and so on. X-
rays are the oldest source of electromagnetic radiation used for medical imaging. Medical image
enhancement methods are used like all other methods and algorithm in image processing, as
chain subsequent edits aimed at achieving a suitable result. Improving one function in the chain
is only useful if the end result is really unproved, and that does not solely depend on that
particular function; it also depends on the quality of the first image. In this paper we have
compared different types of image enhancement technique of medical image in spatial domain
and also presented a statistical analysis.

Medical Image Enhancement domain methods, which operate directly on

Digital image processing is a technique
using which the data from an image is
digitized. With the help of computer
algorithms, various mathematical operations
are applied to the data to create an enhanced
image. In medical image enhancement we
use the input, which is the form of X-ray
film. In the input image we perform some
mathematical operation and the gray values
of the input image are mapped to another set
of gray values, so that the image is made
suitable for specific application. it is called
Image Enhancement method.
Image enhancement technique can be
divided in two broad categories spatial
pixel and frequency domain methods, which
operate directly on the Fourier, transform of
an image. In this present paper we focus our
attention on spatial domain. There are
several enhancements in spatial domain such
as-
i. Gray Scale manipulation.
ii. Histogram equalization.
iii. Image smoothing-Neighborhood
averaging and Edge preserving
smoothing.
The term spatial domain refers to the
aggregate of pixels composing an image.
These methods are procedures that operate

COPO1O5-1
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
directly on these pixels. Spatial domain
process will be denoted by the expression
g(x>y) =T[f(x, y)] (1)
Where f(x, y) is the input image, g(x, y) is
the processed image, and T is an operator on
f, defined over some neighborhood about a
(x, y). The principle approach in defining a
neighborhood about a point (x, y) is to use a
square or rectangular sub image area
centered at (x, y), as shown in fig.1.
Fig.1.
The center of the sub image is moved from
pixel to pixel starting from the top left
corner to the bottom most right corner. The
operator T is applied at each location (x, y)
to yield the output at that location. The
process utilizes only the pixel hi the area of
the image spanned by the neighborhood.
Although other neighborhood shapes, such
as approximation to a circle, square and
rectangular arrays are by far the most
predominant because of their ease of
implementation. The simplest form of T is a
single pixel when the neighborhood is of
size 1X1.In this case, g depends only on the
value of fat(x, y), and t becomes a gray-level
(also called intensity or mapping)
transformation function of the form:
s=T(r) (2)
Where, r and s are variables denoting, the
gray level of f(x, y) and g(x, y) at any
point(x, y) respectively.
The effect of thresholding would be to
produce an image of higher contrast than
original by darkening the levels below m
(threshold) and brightening the level above
in the original image, this technique is
known as contrast stretching.
The gray scale modification method
includes Log transformation, Power law
transformation and Piecewise transformation
and Histogram equalization method include
local histogram equalization and global
histogram equalization. The third method
and many more methods are possible but we
have not discussed in this paper. Gray-level
transformation is very simple and powerful
processing method of radiographic image
enhancement. The log transformation is used
to compress dynamic range. In this
transformation function large numbers of
gray level values are mapped in to narrow
range. The general form of log
transformation is:
S=c log(1+r) (3)
Where c is a constant and it is assumed that
r>=o
Power-law transformation is used to expand
a certain range of intensity values. In these
methods the exponent of the power-law
equation is called gamma f. By varying
gamma, a family of power-law
transformation can be obtained. Power law
transformations have the basic form:
S=crγ (4)
Where c and γ are positive constants.
The simplest form of Piecewise linear
function is a contrast stretching
COPO1O5-2
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
transformation. Low-contrast image can
result from poor illumination, lack of
dynamic range in the imaging sensor, or
even wrong setting of lens aperture during
image acquisition. The idea behind contrast
stretching is to increase the dynamic range
of the gray levels in the image being
processed.
Global histogram equalization method is a
simple and computationally inexpensive
implementation, but it treats the entire
content image uniformly, regardless of the
local variation of gray levels. This may lead
to loss of visual quality of images for certain
areas and increased visibility of noise.
Discrete domain
Sk= T(rk)= = (5)
Analog Domain
s= T(r) = (6)
Local histogram equalization enhances the
small region in the image features. A square
or rectangular neighborhood is defined and
the centre of this area from pixel to pixel is
moved. At each location, the histogram of
die points in the neighborhood is computed
and either a histogram equalization or
histogram specification transformation
function is obtained. Fig.3 shows various
outputs.
Medical X-ray images have narrow latitude
or dynamic range. The film’s ability to
record and view all the anatomy within the
x-ray field is limited by this narrow dynamic
range. Image enhancement is the process of
applying these techniques to facilitate the
development of a solution to a computer
imaging problem. Consequently, the
enhancement methods usually are
application specific and are often developed
empirically. Actuality of image
enhancement problem conditioned by
necessity of it’s solving in medical
diagnostics, pattern recognition,
enhancement amateur photography and etc.
Since radiation oncologists believe that final
images have a higher contrast, and from the
statistical assessment it can be concluded
that an increase of average numbers and
deviation of images value will increase the
contrast. As a result processed images via
this program will lead to have a higher
diagnostic ability.
Result and discussion
Here we have presented a statistical analysis
in table form. The variation of the mean
value and standard deviation value of input
radiographic images and enhanced
radiographic images are calculated in this
present paper. We have worked on many
radiographic images but only three images
chest, hand and leg are shown in this paper.
In global histogram equalization method
enhances the whole area of image features.
Therefore for achieving uniform distribution
of gray scale values for histogram
equalization the intensity are increased. This
raises the mean value of enhanced image.
With the increase in mean value, the average
distribution of intensity values also
increases. Consequently the standard
deviation increases.
In local histogram equalization method,
main focused area is selected regions
determined by window. The entire image
need not be enhanced. So, mean values and
standard deviation values of the enhanced
COPO1O5-3
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
images are not affected as much as that of
global histogram method. If window size
increases then proportionately larger region
is chosen out of entire image region causing
mean values to decrease. If value of constant
k changes or increases, more uniform
distribution is seen in the histogram it
indicates that more equalization is achieved.
It is observed that in log transformation
method for different value of c, if gray value
increases then statistical values (mean,
standard deviation) of enhanced image also
increases which in turn increases the image
brightness. Similarly in power law
transformation using fixed values of gamma
and different value of c, if the gray value is
increased, the statistical values are also
increased proportionately. In power law
transformation with gamma correction
approach, if the value of gamma increases,
the value of mean of enhanced image
decreases while the value of standard
deviation increases, as result the brightness
of enhanced image is decreased.
Similar observation have been found in
piecewise contrast stretching method i.e.
enhanced image is dark for gamma=0.1 and
contrast increase as the value of gamma
increases from gamma-0.2 onwards.
Table 1, 2, 3 shows the comparison of
different method of Radiographic Image
Enhancement.
Because of flexibility of changing window
size and equalization factor k an auxiliary
region can be enhanced to the extent
required in the local histogram equalization
method. Therefore this method is also
sometimes referred to as adaptive method.
In the present enhancement approaches,
power law transformation using gamma
correction performs much better as
compared to all other methods addressed.
This conclusion is on the basis of the change
in standard deviation and mean value on the
histogram of the enhanced image.
Figure 4, 5, 6 shows plot of mean value and
standard deviation value of four methods of
Leg1, Hand1, and chest 1 image. In bar
graph the global histogram method is not
shown.
Conclusion
In this present paper we have compared
different techniques of image enhancement
in spatial domain. They have been evaluated
in terms of standard deviation, mean values
as statistical measures. As medical image
has only a finite number of gray scales, an
ideal equalization is not possible. Image
enhancement technique can be improved if
the enhancement criteria are application
dependant and can be illustrated precisely.
References
Millener AE, AubinM, Pouliot J,
ShinoharaK, Roach HIM (2004)
“Daily electronic portal imaging
from morbidly obes men undergoing
radiotherapy for localized prostate
cancer”. Int J Radiat Oncol
Bioiphys, 6.10.
Van Denberg DL, De Ridder M,
Storm GA (2004) “Imaging in
radiotherapy” European Journal of
radiology, 41-48.
Shi X.Q., Sallstrom P, and Welander
U. “A Color Coding Method for
Radiographic Image and vision
COPO1O5-4
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

computing” Vol.20, pp:761- Advances in Image Guided

767.20Q2. Radiotherapy”. Euspico 2007,
Pznon, Poland, Sep2007.
Gonzalez RC and Woods R.E.
“Digital Image Processing”, S.S.Y. Lau, „Global image
Pearson Education Singapore, enhancement using local
Fourth Indian Reprint, 2nd edition information”. Electronics letters. Pp.
2003. 122-123, January 1994.

Jain A. “ fundamentals of Digital Jha Vandana, Thakur Kavita and

Image Sinha G.R. “Image Enhancement
Processing”, Englewood Cli.s Techniques in spatial domain”. A
Printice Hall, 1989. case study, J. of Acta Ciecia Indica,
Vol.XXXHI p, No.2, pp:201, 2007.
Pratt W. “Digital Image
Processing”, New York, Wiley JohnB. Zimmerman. et al., “An
Interscience, 1991. evaluation of efectiveness of adaptive
histogram equalization for cantrast
Andrews H.C., Monochrome enhancement”. IEEE Trans. On
“Digital Image Enhancement”, Medical Imaging, pp.304-312.
Appl., p.495, 1976. Deember 1988.
Gonzalez RC and Woods R.E., and www.wikipedia.com.
Eddins SL. “Digital Image
Processing using MATLAB”. http://www.freepatentsonline.com/E
Pearson Edition. P 1397780.html.

R.S.Gaborskl Kodak Health Imaging

Syst. Inc. Rochester, NY, USA,
…BK_ Jang, Kodak Health Imaging
Syst. Inc. Rochester, NY, USA,
“enhanceme… for Computed
Radiographic Images”.

Joung-youn Kirn; Lee-Sup Kim;

Seung-Ho Hwang. “An advanced
contrast enhancement using partially
overlapped subblock histogram
equalization” Circuits and
Systems.2000. Proceedings. ISCAS
2000. Geneva.

Matuszewski Bogdan J.. Moore

Christopher J.. “Deformable Models
in Medical Image Processing,
COPO1O5-5
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Figure 3: Original Image and Enhanced

images using different Methods

Log Transformed

Original Image

Power Law Transform γ=1

Global Histogram Equalized

Power Law Transform γ=0.9

Adaptive Histogram Equalized

Power Law Transform γ=0.8

COPO1O5-6
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Power Law Transform γ=0.7

Figure-4: Mean values and Standard

Deviation of Medical Image Enhancement
for Leg

Figure-6: Mean values and Standard

Deviation of Medical Image Enhancement
for Chest

Figure-5: Mean values and Standard

Deviation of Medical Image Enhancement
for Hand

COPO1O5-7
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Table 1: Comparison of different method of medical image enhancement for chest

Chest

Method Input Image Enhanced Image

Mean Std Mean Std

1.Global Histogram 116.0488 93.1951 127.3704 74.9504

Equalization

2.Adaptive histogram 116.0488 93.1951 114.8210 84.0122

equalization

3.Log transformation 116.0488 93.1951 180.8969 77.4127

4.Power law 116.0488 93.1951 116.0488 93.1951

transformation γ=1

5. γ =0.9 116.0488 93.1951 120.6711 93.2783

6. γ =0.8 116.0488 93.1951 125.9139 93.0551

7. γ =0.7 116.0488 93.1951 132.0142 92.2453

Table 2: Comparison of different method of medical image enhancement for Hand

Hand

Method Input Image Enhanced Image

Mean Std Mean Std

1.Global Histogram 68.9736 80.6995 127.6161 74.2843

Equalization

2.Adaptive histogram 68.9736 80.6995 72.7664 72.7077

equalization

COPO1O5-8
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

3.Log transformation 68.9736 80.6995 157.7583 60.5003

4.Power law 68.9736 80.6995 68.9736 80.6995

transformation γ =1

5. γ =0.9 68.9736 80.6995 73.9827 81.1636

6. γ =0.8 68.9736 80.6995 80.1207 81.0990

7. γ =0.7 68.9736 80.6995 87.7081 80.3351

Table 3: Comparison of different method of medical image enhancement for Leg

Leg

Method Input Image Enhanced Image

Mean Std Mean Std

1.Global Histogram 60.4228 70.2646 149.6279 50.3502

Equalization

2.Adaptive histogram 60.4228 70.2646 72.8500 73.9090

equalization

3.Log transformation 60.4228 70.2646 170.7315 108.0530

4.Power law 60.4228 70.2646 60.4228 70.2646

transformation γ =1

5. γ =0.9 60.4228 70.2646 64.7082 73.3188

6. γ =0.8 60.4228 70.2646 69.4427 76.5475

7. γ =0.7 60.4228 70.2646 74.7634 79.9895

COPO1O5-9
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
A Critical Analysis on Prosthetic Arm
Anil kumar
B.Tech (EI 3 yr)
Galgotia‟s college of Eng. & Tech.
Greater Noida
anil00699@gmail.com
Abstract – Prosthetic limbs are a real
time challenging and demanding
biomechatronics field due to more and
more day-to-day basis work objectives for
amputees. This paper investigates the
design of various prosthetic arms to
mimic its human counterpart. The
comparative analysis of available
technology and research work in field of
bionic arms are being done. In this
analysis, the merits and demerits of
various prosthetic hand for real time
application is being highlighted. In-depth
analysis shows that further enhancement
could be done in order to augment DOF,
grasp control, available strength of the
artificial arm and signal processing
techniques.
Keywords- Prosthetic Arm, Robotic Arm,
Artificial hand
INTRODUCTION
Prosthesis is an artificial extension that
replaces a missing body part. Prostheses are
typically used to replace parts lost by injury
or missing from birth or to supplement
defective body parts. In addition to the
standard artificial limb for every-day use,
many amputees have special limbs and
devices to aid in the participation of sports
and recreational activities. Myoelectric
Dr. Monika Jain
Professor-Electronics & Instrumentation
Galgotia‟s college of Eng. & Tech.
Greater Noida
monika_jain24@rediffmail.com
prosthetic hand is used as natural hand by
amputees (who lost their natural hand) by
amputation. Initially, hooks are being used
for this purpose. In recent time, with the
advancement in technology, artificial smart
arm are being used. In myoelectric hands,
electromyography (EMG) signals are used
as control input. EMG signals are
bioelectrical signal released by muscles.
They are normally of 10-500 Hz frequency
and up to 10 mV. The main part of signal is
near 450Hz. conventionally, surface
myoelectrodes are being used in myoelectric
prosthetic arms, however with help of SRI
EMG technique, signal strength and
accuracy of signal would increase. This can
be use to control individual fingers. If
prosthetic hand can be made of same
mechanical properties as a neuromuscular
control system of hand then amputee can use
it more effectively. A lot of work is done in
this field to improve prosthetic arms but
feedback is not used effectively. By using
feedback connected at artificial fingers like
pressure sensor, accelerometer, temperature
sensor, hand‟s working can be improved.
Also DOF and grasping capacity of artificial
arm can be improved. There is no doubt that
microelectronics offers numerous
opportunities for improvement in the
engineering of substitutes for the human
limb. The possibilities become exciting as
soon as we move away from the all-
mechanical prostheses which have been the
COPO1O1-1
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
conventional prescription until today. We
have described a relatively simple
application of hybrid microelectronics in
this novel field.
The purpose of this project is to develop a
myoelectric controlled individual finger,
multi-DOF adaptive grasping technique
equipped commercial hand capable to fast
response. Also amputee can feel touch,
temperature and force applied by prosthetic
hand on the joint of prosthetic and natural
hand.
LITERATURE SURVEY
In the field of prosthetic arm, lots of
research work is in progress. Maximum
prosthetic hands are simple grippers with
one or two degrees of freedom (DOF). They
are using smart hooks (passive fingers and
thumb) like Otto Bock [1] they have 2 or 3
point of contact so more force is required for
grasping. Also they are capable of gripping
as in myohand variplus speed [2] but not
capable of some more works that include
finger, like door opening, rotating car key
etc. Kenzo Akazawa [3] developed a hand
using dynamic behaviour of antagonist
muscles, flexor muscles and extensor
muscles but is similar as in [1]. Also its
response is slow and a large training session
is required to train the amputee. For fast
processing, Isamu [4] has presented
evolvable hardware chips with genetic
algorithm. It takes less time to train
amputee. But it restricts the DOF of hand.
Also dedicated chip is required for
prosthetic hand which reduces versatility to
use high speed microcontrollers.
Microprocessor and high torque motor based
hand is being made by Ryuhei [5], however,
this uses only two surface EMG signals, due
to which less DOF is obtained. Also, more
training required for amputee. Another main
drawback in above is that they uses EMG
signal as stationary signal but for multi DOF
it is required to process it as non stationary
signal. Real time EMG pattern recognition
[6] is done to improve the hand functioning.
But still signal is not sufficient so fingers
can‟t be controlled separately. For this
purpose SRI EMG (surface recorded
intramuscular EMG [7] signals can be used
they are more accurate than surface EMG
signals. Prosthetic hand with individual
finger movement will be more similar to
natural hand. Prosthetic hand approximately
similar to natural hand is made by Bryan
Christie [8] and Dean Kamen [9] with the
help of DARPA. But it is connected by
nerve system so not easy to use also it‟s still
in experimental state. Chappell [10] has
presented an approach having an artificial
hand with sensors allowing for the inclusion
of automatic control loops, freeing the user
from the cognitive burden of object holding
which is similar to the natural low level
spinal loops that automatically compensate
for object movement. Force, object slip and
finger positions are variables that need to be
measured in a hand designed for the
physically impaired person. It shows that a
high specification sensor is required for
designing an arm. Also it must be designed
separately for each amputee. But latest
technology provides adaptive signal
processing helping to adapt the amputees‟
requirements. An electrically driven locking
mechanism has been built by Law and
Hewson [11], which is controlled by the
electromyogram (EMG) of the surviving
muscles in the upper arm. Hybrid
technology is used for the construction of
the associated electronic circuitry. Many
similar applications are now being
considered in an attempt to improve the
performance of upper-limb prostheses using
latest researches. Development, testing and
experimentation work of a device for the
hand rehabilitation was done by Mulas et al
COPO1O1-2
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
[12]. The system designed is intended for
people who have partially lost the ability to
control correctly the hand musculature, for
example after a stroke or a spinal cord
injury. Based on EMG signals the system
can "understand" the subject volition to
move the hand and actuators can help the
fingers movement in order to perform the
task. This paper describes the device and
discusses the first results conducted on a
healthy volunteer. It requires a number of
actuators to increase the DOF of prosthetic
arm. Also EMG processing done is not
sufficient to provide significant
performance. Massa et al [13] designed a
hand to augment the dexterity of traditional
prosthetic hands while maintaining
approximately the same dimension and
weight. This approach is aimed at providing
enhanced grasping capabilities and natural
sensory-motor coordination to the amputee,
by integrating miniature mechanisms,
sensors, actuators, and embedded control. A
biomechatronics hand prototype with three
fingers and a total of six independent DOFs
has been designed and fabricated. This
research work is focused on the actuators
system, which is based on miniature
electromagnetic motors. However, still not
using better EMG processing technologies
which can dramatically increase
performance of the hand also the grasping
force is low because of the limited torque
generated by the miniature actuators used
for the hand (which are among the best
available on the market in that range of
size). Embedded control architecture for the
action and perception of an
anthropomorphic 16 degree of freedom, 4
degree of actuation prosthetic hand for use
by transradial amputees has also been
reported. The prosthetic hand is provided
with 40 structurally integrated sensors useful
both for automatic grasp control and for
biofeedback delivery to the user through an
appropriate interface (either neural or non-
invasive). Cipriani et al [14] briefly
describes the mechatronic design of the
prosthesis, the set of sensors embedded in
the hand and finally focuses on the design of
the control architecture that allows action
and perception for such a sophisticated
device. It consists of 8-bit architecture
microcontrollers, however, not using the
available signal processing techniques.
Herrera et al [15] designed and constructed
a prosthesis that will be strong and reliable,
while still offering control on the force
exerted by the artificial hand. The design
had to account for mechanical and electrical
design reliability and size. These goals were
targeted by using EMG in the electrical
control system and a linear motion approach
in the mechanical system. The prosthetic
gripper uses EMG to detect the amputee's
intended movement. The control system
requires an adaptation mechanism for each
amputee's characteristics. Gordon et al [16]
used proportional myoelectric control of a
one-dimensional virtual object to investigate
differences in efferent control between the
proximal and distal muscles of the upper
limbs. Restricted movement was allowed
while recording EMG signals from elbow or
wrist flexors/extensors during isometric
contractions. Subjects used this proportional
EMG control to move the virtual object
through two tracking tasks, one with a static
target and one with a moving target (i.e., a
sine wave). Eriksson et al [17] studied the
neural network feasibility for categorizing
patterns of EMG signals. The signals
recorded by the surface electrodes are
sufficient to control the movements of a
virtual prosthesis. The presented method
offers great potential for the development of
future hand prostheses. A signal processing
system based on RAM as a look-up table
(LUT) has been presented by Torresen et al
[18]. This provides a fast response besides
being compact in size. Several algorithms
for programming it have been proposed For
COPO1O1-3
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
the given data set used in the following
experiments, the time used for programming
the RAM was approximately equal to the
time used for training a feed-forward neural
network – solving the same problem in the
best possible way. However, the main
advantage of this scheme is the fast runtime
speed. Ferguson [19] described the
development of a system that will allow
complex grasp shapes to be identified based
on natural muscle movement. The
application of this system can be extended to
a general device controller where input is
obtained from forearm muscle, measured
using surface electrodes. This system
provides the advantage of being less
fatiguing than traditional input devices. V.
Tawiwat et al. [20] applied a mouse„s roller
with a gripper to increase the efficiency of a
gripper could lead to material handling
without slipping. To apply a gripper, the
optimization principle is used to develop
material handling by use of a signal for
checking a roller mouse that rotates. In case
the roller rotates, meaning that the material
slips. A gripper will slide to material
handling until the roller does not rotate. In
an attempt to improve the functionality of a
prosthetic hand device, a new fingertip has
been developed, that incorporates sensors to
measure temperature and grip force and to
detect the onset of object-slip from the hand.
The sensors have been implemented using
thick-film printing technology and exploit
the piezoresistive characteristics of
commercially available screen printing
resistor pastes and the piezoelectric
properties of proprietary lead-zirconate-
titanate (PZT) formulated pastes. The force
sensor exhibits a highly linear response to
forces. The force sensor response is also
extremely stable with temperature. The
ability of the piezoelectric PZT vibration
sensor to detect small vibrations of the
cantilever, indicative of object slip, has also
been demonstrated [21]. Externally powered
upper extremity prosthesis has been
considered as a system in which the
necessary components to design a better
prosthetic arm are viewed and divided into
four subsystems: input, effecter, feedback
and support. Current research is reviewed in
terms of these subsystems. Each subsystem
performs its own task, but they are related to
each other and together they function to
make up a prosthetic upper extremity, which
provides the movement to the amputee [22].
Hands such as the All Electric Prosthetic
Hand utilize a series of gears to transmit the
motion of motors housed in the forearm to
the relevant fingers [23]. Other designs have
the actuators transmitting the power directly
to the joint. An example of this is the
Anthroform Arm, which uses pneumatic
„muscles‟ mimicking the muscles of the
human arm connected directly to the „bones‟
it moves [24]. Shape Memory Alloy wires
are also used, to both provide the force and
transmit the motion. SMA wires contract
when heated and return to their initial shape
when cooled [25]. This method of actuation
is utilized in the Shape Memory Alloy
Activated Hand constructed by DeLaurentis
et al [26].
CONCLUSIONS
A lot of work is done for developing the
prosthetic arm but still more precision work
could be done. Grasping technique can be
enhanced so that amputees can work more
effectively, response time of available arms
is not sufficient. Still the available prosthetic
arms are not comparable with natural hand
in terms of DOF. The main limitations are
less space available for motors and their low
size to torque ratio. Power consumed is also
more if number of motors increases. The
lack of enhanced signal processing
techniques in artificial hands and associated
controllers has limited their functionality
COPO1O1-4
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
and technological progress. Still a lot of
work can be done to improve the available
prosthetic arms. Use of hydraulics in place
of motors can provide great amount of
controlled force in lesser space. Due to this
available control of force applied by the
hand on the object, amputees can handle soft
or brittle items easily. By effective use of
hydraulics, DOF can also be increased
without any increased power consumption.
Also use of adaptive signal processing
techniques can improve overall performance
of artificial hand.
REFERENCES
[1]Available on:
http://www.ottobock.com.au/cps/rde/xchg/o
b_au_en/hs.xsl/384.html
[2Available on:
http://www.ottobock.com.au/cps/rde/xchg/o
b_au_en/hs.xsl/19932.html
[3] Kenzo Akazawa, Ryuhei Okuno and
Masaki Yoshida “Biomimetic EMG-
prosthetic-hand” 18th Annual International
Conference of the IEEE Engineering in
Medicine and Biology Society Amsterda,
pp. 535-536.
[4] Isamu Kajitani and Tsukuba Masahiro
Murakawa “An Evolvable Hardware Chip
for Prosthetic Hand Controller”
Microelectronics for Neural, Fuzzy and Bio-
Inspired Systems, 1999. MicroNeuro' 99.
Proceedings of the Seventh International
Conference on pp. 179-186.
[5] Ryuhei Okunol, Masahiro Fujikawa',
Masaki Yoshida2, Keno Akazawal
“Biomimetic hand prosthesis with easily
programmable microprocessor and high
torque motor” Engineering in Medicine and
Biology Society, 2003. Proceedings of the
25th Annual International Conference of the
IEEE Vol.2 pp.1674-1677.
[6] Jun-Uk Chu, Inhyuk Moon, Member,
IEEE, and Mu-Seong Mun “A Supervised
Feature Projection for Real-Time
Multifunction Myoelectric Hand Control”
International Conference of the IEEE
Engineering in Medicine and Biology
Society pp.282-290.
[7] Nikolay S. Stoykov, Madeleine M.
Lowery, Charles J. Heckman, Allen Taflove,
and Todd A. Kuiken “Recording
Intramuscular EMG Signals Using Surface
Electrodes” Proceedings of the 2005 IEEE
9th International Conference on
Rehabilitation Robotics June 28 - July 1,
2005, Chicago, IL, USA pp. 291-294.
[8] Available on:
http://spectrum.ieee.org/robotics/medical-
robots/winner-the-revolution-will-be-
prosthetized/2
[9] Available on:
http://spectrum.ieee.org/biomedical/bionics/
dean-kamens-luke-arm-prosthesis-readies-
for-clinical-trials
[10] P H Chappell, “A fist full of sensors”
Journal of Physics: Conference Series, vol.
15, 2005, pp. 7-12.
[11] H.T. Law and J. J. Hewson, “An
Electromyographically Controlled Elbow
Locking Mechanism for an Upper Limb
Prosthesis”, Electro component Science and
Technology, vol. 10, 1983, pp. 87-93.
[12] Marcello Mulas, Michele Folgheraiter
and Giuseppina Gini, “An EMG-controlled
Exoskeleton for Hand Rehabilitation”, IEEE
9th International Conference on
Rehabilitation Robotics, Chicago, IL, USA,
June 28 - July 1, 2005, pp 371-374.
[13] M. C. Carrozza, S. Micera, R.
Lazzarini, M. Zecca and P. Dario,
COPO1O1-5
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
“Development of a Novel Prosthetic Hand”
Ongoing Research and Preliminary
Results‖ IEEE/ASME Transactions on
Mechatronics, vol. 7, 2002, pp. 108-114.
[14] Christian Cipriani, Marco Controzzi,
Fabrizio Vecchi and M. Chiara Carrozza,
“Embedded Hardware Architecture Based
on Microcontrollers for the Action and
Perception of a Transradial Prosthesis”,
International Conference on Biomedical
Robotics and Biomechatronics Scottsdale,
October 19-22, 2008, AZ, USA, pp. 848-
853.
[15] Andres Herrera, Andres Bernal, David
Isaza and Malek Adjouadi, “Design of an
Electrical Prosthetic Gripper Using EMG
and Linear Motion Approach”, National
Science Foundation Grants EIA-9906600
and HRD-0317692.
[16] Keith E. Gordon, Daniel P. Ferris,
“Proportional Myoelectric Control of a
Virtual Object to Investigate Human
Efferent Control”, Exp Brain Res, vol. 159,
2004, pp 478–486.
[17] Lars Eriksson, Fredrik Sebelius,
Christian Balkenius, “Neural Control of a
Virtual Prosthesis”, IEEE 8th International
Conference on Artificial Neural Networks,
vol. 2, 1998, pp. 905-910.
[18] Jim Torresen, Vidar Engh Skaugen, “A
Signal Processing Architecture Based on
RAM Technology”, 16th European
Simulation Multiconference on Modelling
and Simulation, ISBN:90-77039-07-4, pp.
317-319.
[19] Simon Ferguson, G Reg Dunlop,”Grasp
Recognition from Myoelectric Signal”,
Australasian Conference on Robotics and
Automation, Nov 27-29, 2002, Auckland,
pp.83-87.
[20] V. Tawiwat, and S. Sarawut, “Optimal
Design for Grip Force of Material
Handling”, Proceedings of World Academy
of Science, Engineering and Technology,
July, 2008, pp 864-868.
[21] A. Cranny , D.P.J. Cotton, P.H.
Chappell, S.P. Beeby, N.M. White, “Thick-
film Force and Slip Sensors for a Prosthetic
Hand”, Sensor and Actuator 2005, pp. 162-
171.
[22] HaeOck Lee, Dawnlee J. Roberson, “A
Systemic View of an Upper Extremity
Prosthesis”, IEEE International Conference
on System of Systems Engineering, April
2007, pp. 1-6.
[23] I. Kato and K. Sadamoto (ed.),
“Mechanical Hands Illustrated”, Survey
Japan (1982).
[24] G. K. Klute, J. M. Czerniecki and B.
Hannaford, “McKibben Artificial Muscles:
Pneumatic Actuators with Biomechanical
Intelligence”, [Online, accessed 26 March
2001]
URL:
http://brl.ee.washington.edu/BRL/devices/m
ckibben/aim99v6.pdf
[25] J. L. Banks, “Anthropomorphic Robotic
Finger Platform Based on Shape Memory
Alloy”, Master‟s Thesis, MIT Artificial
Intelligence Laboratory, [Online, accessed 4
May 2001]
URL:
http://www.ai.mit.edu/projects/humanoid-
robotics-group/Abstracts2000/banks.pdf
[26] K. J. DeLaurentis, C. Mavroidis and C.
Pfeiffer, “Development of a Shape Memory
Alloy Activated Robotic Hand”, [Online,
accessed 19 March 2001] URL:
COPO1O1-6
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

http://www.cronos.rutgers.edu/~mavro/pape
rs/act2000.pdf

COPO1O1-7
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Improvement in Dijkstra algorithm

Ashutosh Singh1, Sandeep Rathour2
Department of Electronics Engineering, H.B.T.I., Kanpur, India
1
singhash@iitk.ac.in, 2ersandeep.hbti@gmail.com

Abstract 2. Storage structure[1]

To improve the efficiency of a
network, many people have conducted many
studies. For finding the shortest path between
two points, Dijkstra Algorithm is a research
hotspot. The Dijkstra Algorithm has its own
merits and demerits when finding the shortest
path between two points. The two main
disadvantage of Dijkstra Algorithm can be
summarized as (i) Storage structure (ii)
Searching area. One more process by which
Dijkstra Algorithm can be improved. It can be
improved by reducing the no. of repeated
operation and can find the shortest path and
path length from the start point to all the
other remaining nodes by a concept of
Feature matrix.
Keywords: Dijkstra Algorithm , Storage
structure, Searching area, Shortest Path.
The storage structure of Dijkstra
Algorithm is large i.e. it takes large space for
storing it.
Let take a example of 10 nodes. For the
classical Dijkstra Algorithm it requires
10x10[100 values] for its storage. The figure
1 shows the example for 10 nodes. Since it
require 10x10 matrix for storage, the matrix
shown in figure 2 have following properties.
(a) if the start point and the end point are
same,
i.e. i=j, then aij=0
(b) aij=aji
(c) if there is no any direct path between
two nodes , then it has ∞(large value) value.
Example
3
4 3

2 3
2 9 4
1. Introduction 2
5

In the graph theory, Dijkstra 5

2 5
1
Algorithm is the most classical algorithm for 3
10
4
finding the shortest path in the graph. But the 3
1 6
8
time complexity of the Dijkstra Algorithm is
3 3
higher and also it takes larger space for 7

storage. To overcome these shortcomings of

Dijkstra Algorithm many scholars have done
many researches and their result have its own
merits and demerits. In this paper there are
some improvement in classical Dijkstra
Algorithm to reduce the time complexity of
the Dijkstra Algorithm and also reduce the
searching area of the algorithm. And also for
improvement in Dijkstra Algorithm, if the no.
of repeated operation for searching the
shortest path is reduced, then the efficiency of
the Dijkstra Algorithm is also get improved.
Figure.1 nodes arrangement(10 nodes)
For reducing the storage structure of the
above network shown in figure.1, in the
matrix shown in figure.2, the places where the
values are 0(zero) and ∞(infinity) can be
removed, only non-∞ elements are stored in
the simple linked list format. The improved
structure is shown in figure.3. in this form
each node is having only two elements, first
one is the column value and second one is its
weight.

COPO3O2-1
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

4. Searching Area
0 2 ∞ ∞ ∞ ∞ ∞ 3 ∞ ∞ Since the searching area of a classical
2 0 4 ∞ ∞ ∞ ∞ 3 2 ∞
Dijkstra Algorithm is large, there are some
∞ 4 0 3 ∞ ∞ ∞ ∞ ∞ 5
process by which the searching area of the
∞ ∞ 3 0 5 2 ∞ ∞ ∞ ∞
∞ ∞ ∞ 5 0 4 ∞ ∞ ∞ ∞
classical Dijkstra algorithm can be reduced.
∞ ∞ ∞ 2 4 0 3 ∞ ∞ ∞ Since the shortest path between two points is
∞ ∞ ∞ ∞ ∞ 3 0 3 ∞ 1 a straight line. The direction from the start
3 3 ∞ ∞ ∞ ∞ 3 0 ∞ ∞ point to the is generally strike of the shortest
∞ 2 ∞ 3 ∞ ∞ ∞ ∞ 0 5
∞ ∞ ∞ ∞
path when plan the route of the real road
∞ 5 ∞ 1 5 0
network. The shortest path between two
points is generally the both side of connection
line of the start point to the destination point
Figure.2 adjancy matrix(10x10) and usually it is near to the connection line. If
there are only one edge between the start
The storage format of the above matrix get point and the destination point, the edge itself
reduced and given as- is the shortest path.

G1 2 2 8 3 ^ 4.1 Searching Area is rectangle

G2 1 2 3 4 8 3 9 2 ^ The author[1] shows in figure 4 to
G3 2 4 4 3 ^
find the shortest path from the source S to the
destination D, the theoretical searching area
G4 3 3 5 5 6 4 9 3 ^
of classical Dijkstra algorithm is a circle
G5 4 5 6 4 ^ whose center is source S and radius is R. the
G6 4 2 5 4 ^ restricted searching area algorithm, the
G7 6 3 8 4 10 1 ^ theortical searching area is small rectangle. It
G8 1 3 2 3 7 4 ^
may be considered that there may exist a
reserve path, each side of the small rectangle
G9 2 2 4 3 10 5 ^
can be increased to threshold T2 , and it make
G10 7 1 9 5 ^ a large rectangle. So the restricted searching
area is formed by the two lines L1 and L2
figure.3 improved storage strucuture which cuts the larger rectangle, so the final
searching area is shown by dark line in figure
From the above figure it can be told that this 4.
structure takes less space in comparison to the
classical Dijkstra algorithm.

3. Complexity Analysis
Let for the list array, the space
complexity is given as O(T). where T is the
edgentain one or more tables to no. of the
directed graph.
In the worst case T=n2 , then the space
complexity will be O(n2).

COPO3O2-2
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

the distance between the neighbours, the time

L1
complexity difference is more significant. So
this improved algorithm restricting the
T1 L2 searching area reasonably reduce the scale of
D
searching, so it minimizes the complexity.
T2
And also it improves the efficiency of the
system. The improved algorithm itself restrict
the searching area, unfortunately it may be
S
the algo. Sometimes can not search the actual
R2 shortest path and it can only be the an
approximate shortest path, it may be a
defective shortest path algo and also it reduce
the accuracy.
Figure.4 rectangular searching area

4.2 Searching area is an ellipse

In the figure, S is the start point, When searching area is an ellipse
D is the destination point, shown in figure 5, the following parameter
T1 & T2 are threshold. are used for finding the searching area. In
R is the distance between source and case of ellipse,
destination. SD=R=2a [major axis of ellipse],
(i) for the classical Dijkstra Algorithm the Let the eccentricity e=0.7 [<1],
searching area is a circle of radius R , then the b [semi minor axis] can be obtained.
will be In this case the searching area is shown in
A1=π*R2 figure 5 by a thick line, and can be obtained
(ii) for the restricted searching area algorithm as,
the searching area is given by A2 = T1(4a+πT2)
A2=2T1[R+ T2] Now the time complexity ratio can be given
Where T1/T2= as
(iii) since the searching node is the directly Time complexity ratio=
proportional to the searching area, so the time
complexity can be given as O(A12) and = [ ]2
O(A22).
Commonly two threshold T1 & T2 are
relatively small constant. SD=2a(major axis)
Eccentricity(e)=.7
L1
T2
So the ratio of the time complexities can be D L2
T2
expressed as –
T1
Time complexity ratio= T1

S
T2
= [ ]2

From the above equation it can be compared

that the searching area algorithm is smaller
than the classical Dijkstra algorithm. If there
is increase in the scale of road network and figure 5 when searching area is an ellipse

COPO3O2-3
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

from the above equation, it can be compared The abutment matrix for the example is
that the ratio of time complexities of, when given as A=
searching area is an ellipse is less in
comparison to the ratio of time complexities
0 2 1 ∞ 7 12
of , when searching area is rectangular. 2 0 ∞ 2 ∞ ∞
Elliptical searching area gives better result. 1 ∞ 0 1 2 ∞
∞ 2 1 0 3 ∞
5. Feature matrix[2] 7 ∞ 2 3 0 4
Another way for the improvement of 12 ∞ ∞ ∞ 4 0
classical dijkstra algorithm. Since if we to
find the shortest path between two points, for
this purpose there are a no. of operations by There are following steps which from which
which it can be done. But if in any way it can feature matrix can be obtained.
reduce these no of operations, then the 1. source S={v1}, D={0,2,1,∞,7,12}
efficiency of the classical dijkstra algorithm 2. firstly find the shortest distance from the
will get increases. For this purpose there is a source i.e. D[3]=1, so S={v1,v3}
concept of feature matrix. From this feature 3. now the nodes connected via 3 to the 1 ,
matrix we can draw the shortest path tree and distance can be obtain by
can get the shortest path and shortest path D[3]+A[3][4]=2<D[4]=∞
length from source node to all other & D[3]+A[3][5]=3<D[5]=7
destination node. So that modification will D[4]=2 ,D[5]=3 and
For understand the concept of feature matrix the D matrix will become D={0,2,1,2,3,12}
let’s take a example of 6 node which are 4. iterate operation as follows,
connected as shown in figure.6. The second time we get that,
S={v1,v3,v2,v4}
& D={0,2,1,2,3,12}
The third time we get S={v1,v3,v2,v4,v5}
1
12 & D={0,2,1,2,3,7}
6

2 7
The fourth time S={v1,v3,v2,v4,v5,v6}
1 4 & D={0,2,1,2,3,7}
3
2
2 5 5. compare the following D and A matrix we
2
1 will get the following matrix called feature
4
3
matrix(F).

0 0 0 0 0 0
Figure.6 nodes arrangement
2 0 0 0 0 0
1 0 0 0 0 0
0 0 1 0 0 0
0 0 2 0 0 0

0 0 0 0 4 0

COPO3O2-4
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
6.From the feature matrix it can be seen that
v1 is the precursor node of the node v2 and
v3. v3 is the precursor node of the node v4
and v5, and v5 is the precursor node of the
node v6.
7. now shortest path tree can be obtained like
shown in figure.7.
1
1
2 to any other node by reducing the no. of
3 2
1 2
4 5
4
6
Figure.7 shortest path tree
From the shortest path tree we can find the
shortest path from the source to the all other
nodes and also can find the path length.
6. Conclusion
Finally after studying the above
algorithm, it can concluded that the restricted
searching area algorithm has lower time
complexity than classical dijkstra algorithm
and also when searching area is an ellipse has
lower time complexity when searching area is
a rectangular.
By using the concept of feature matrix, it can
be found the shortest path from source node
searching operation which increases its
efficiency.
7. References:
[1]. Dong Kai fan, Ping Shi, “Improvement of dijkstra
algorithm and its application in route planning”,
Shandong University of technology, China, IEEE,
International conference on FSKD, 2010, pp 1901-
1904.
[2]. Ji-Xian Xiao, Fang-Ling Lu, “An improvement of
the shortest Path algorithm based on Dijkstra
algorithm”, IEEE, ICCAE, 2010, pp. 383–385.
COPO3O2-5
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
EXECUTED WRIST MOVEMENT CLASSIFICATION FOR
BRAIN COMPUTER INTERFACE
Mariam Siddiqui, Ayman Nidal Khalidi, Omar Farooq,Yusuf U Khan
Department of Electronics Engineering, Department of Electrical Engineering
Zakir Hussain College of Engineering & Technology, Aligarh Muslim University, Aligarh (India)
Abstract—The aim of our study is to
investigate best features for the
classification of pronation and extension
of wrist movements and further use the
best possible classifier to achieve accurate
classification of movements. 10-20
International Montage was used for the
recording of EEG data on 16 channels.
Data was processed, filtered and
windowed. Histogram plots were used for
selection of features. Standard statistical
features were employed in the
investigation and the result suggests that
they provide discriminatory conditions.
While the scope and depth of the study
was limited, the results do suggest that
variance, mean, skewness and kurtosis
can be used for designing the classifier for
both the movements. Different Standard
classifiers are finally designed and trained
to distinguish between different motor
movements.
Keywords:BCI, EEG, Pronation, Extension, Variance,
Skewness, Kurtosis, Mean
I. INTRODUCTION
A brain-computer interface (BCI) is a
system which translates a subject’s
intentions into a control signal for a device,
e.g., a computer application, a wheelchair or
a neuroprosthesis[1]. BCI has given us hope
where an alternate medium can be created
for people with severe motor disability so
that they can somehow work over their
disability to communicate with the external
world in a non-biological and easier way.
Encouraged by this understanding , a
growing interest has developed in recording,
detecting and analyzing brain (neural)
signals to investigate, explore and
understand human motor control systems
with the aim of attempting to build
interfaces that use real– time brain signals to
generate commands to control and/or
communicate with the environment [1].
We mainly focus on noninvasive,
electroencephalogram- (EEG) based BCI
systems which can be used as tools of
communication for the disabled or for
healthy subjects who might be interested in
exploring a new path of human-machine
interfacing. When measured non-invasively,
brain activity is acquired by scalp-recorded
EEG from a subject that tries to convey its
intention by reacting to well-defined tasks or
instructions [1]. Research in developing
successful brain–computer interfaces (BCIs)
endeavors to improve the speed and
accuracy of communication of these
interfaces. This is achieved by improving the
EEG signal processing techniques and
selection of features that are translated into
commands to control and communicate with
augmentative and assistive devices [3].
In this paper we propose and experimentally
verify different features that can be used for
training the classifier for wrist movement in
two directions. A histogram-based analysis
was applied to the EEG signals from
COPO303-1
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27
2011
different channels, to determine combination
of features and channels that yielded the best
class separability. Standard classifiers were
then designed and trained to achieve the best
possible classification accuracy
II. BACKGROUND RESEARCH
This domain is still quite fresh and
acqiuring different shapes ,since it emerged
40 years back in 1970.At that time, the
research subjects were initially animals(rats
and monkeys)but then it took a dramatic
shift in 1990s when the first successful
working implant in humans occured .Now
not only damaged eyesight, hearing,
impaired movements’ etc. can be restored
but researchers could now conceivably
attempt to produce BCI that augment human
functions rather than simply restoring
them.[1]
In order to meet the requirements of the
growing technology expansion,some sort of
standardization was required not only for the
guidance of future researchers but for the
validation and checking of new
developments with other systems as
well,thus a general purpose system was
developed called BCI2000 which made
analysis of brain siganl recording easy by
defining the output formats and operating
protocols to facilitate the researchers in
developing any sort of application.This
made it easier to extract specific features of
brain activity and translate them into device
control signals. [4]
.
III. METHODOLOGY
Fig1: Block diagram for feature extraction of wrist movement
A.Experimental Setup and Data Acquisition
For this study, data was recorded from three
healthy subjects, who signed consent before
their participation in the study. EEG was
recorded at the biomedical signal processing
lab, department of electronic engineering,
AMU.16 channels of EEG data were
recorded using a Brain Tech clarity TM
system with the electrodes positioned
according to the standard 10-20 system.
Sampling rate was 256 Hz. The subject was
seated on an armchair with eyes closed. The
subject was instructed to avoid speaking and
to ensure least body movement. The
subjected was instructed to move the wrist
from neutral position on an audio prompt
from the experimenter. The series of trials
recorded consisted of following sequences:
First of all the subject was asked to relax for
some time then followed by pronation, after
reaching the extreme position the subject was
asked to hold for some time which was
followed by extension.
Fig2 Sequence followed during
experimental recording
Times of recording of different movements
were different.
B.Data Processing
EEG data was notch filtered to remove 50Hz
frequency from each channel present in the
data due ac power supply of the EEG
machine. Harmonics of 50Hz frequency
were also removed from the data. IIR second
order notch filter with the quality factor (or
Q factor) of 3.55 was used to remove the
desired frequency.
COPO303-2
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27
2011

Magnitude Response (dB) then averaged over all possible channels

5
.These features could further be used to
classify between the given movements and
0
between movements and neutral state.
-5
Magnitude (dB)

-10

-15
D. CLASSIFIER DESIGN:
Filter #1
-20 Motor movements are finally differentiated
-25 by using different classifiers on the recorded
-30
data
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Normalized Frequency ( rad/sample) Table1: percentage values of classification for motor
Fig3: Magnitude response of a notch filter designed at 50Hz (or positions
.39∏ radian per sample) frequency

Power spectral density of filtered and

Classifier Extension Pronation
unfiltered EEG signal of all the 16 channels
stacked over one another is shown below SVM 77 80
1200 Quad 90 88.8
1000 Linear 80 77.7
800 Diaglinear 80 66.6
600
amplitude

400
From the observations in Table1 we see that
Quad classifier gives us the best possible
200 results with high classification percentage
0
accuracy for both extension and pronation
movements.
-200
0 20 40 60 80 100 120 140
frequency

Fig4: PSD plot of 16 channel filtered and unfiltered EEG signal

IV. RESULT AND DISCUSSION

EEG by nature is non stationary signal. So it
was fragmented into frames so that it can be Histogram plots for 16 channels for different
assumed stationary for small segment. EEG movements pave way for demarcating
data is divided into frames of 1s duration i.e. different movements. From Fig 5 which
frame size of 256 samples /s. shows the histogram plot for electrode F3 for
extension shows that the data has long tail,
hence kurtosis could be one of the features.
C.Feature extraction Further, we can see in the same figure that
The recorded data was performed for both the data is asymmetric about the mean
pronation and extension wrist movement position which shows that it skewed towards
related activity and neutral state of right right. Moreover, if we compare Fig 5 & 6,
hand. Histogram of each channel was plotted which is plot for pronation shows that both
and compared for each of the three different dynamic range and have different
movements and the best possible plots which mean values. Likewise, compare Fig 5, 6,
showed maximum demarcation were used to and 7 they differ in their dynamic ranges. Fig
extract possible features. Further,from16 7 which shows the histogram plot of neutral
channel filtered and segmented data for each state doesn’t have long tail and it doesn’t
of motor activity and non-motor activity’s have much variation about the mean position
basic statistical features [2] mean, variance, thus statistical features can be used to easily
skewness and kurtosis were extracted and

COPO303-3
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27
2011

differentiate between motor and non-motor 4000

F3-neutral state

activity.
3500

Thus from the above discussion we conclude 3000

that mean, variance, kurtosis and skewness

frequency of occurance
2500
can prove to be good features for
classification of wrist movement. 2000

1500

1000

500

F3-Extension 0
2500 -40 -30 -20 -10 0 10 20 30 40
Range of values

Fig7: Histogram plot of frontal electrode (F3) for Neutral

2000 state.
frequency of occurance

1500

features Neutral Extension Pronation

1000 position
Variance 120.16 586.85 299.06
500
Mean -0.07 0.33 0.24
kurtosis 4.77 6.44 5.75
0 skewness 0.13 0.42 0.28
-120 -100 -80 -60 -40 -20 0 20 40
range of values

Fig5: Histogram plot of frontal electrode (F3) for Table2: Statistical features values for different
extension.
positions

F3-Pronation
2500 From the result tabulated in Table2
Extension has larger variance, mean, kurtosis
2000 and skewness values while pronation has
smaller values. Also for both the motor
frequency of occurance

1500 activity values of features are far away from

the non-motor activity. Thus both the
1000 movements can be easily demarcated from
each other.
500

0
From Table1, we justify our findings by
-60 -40 -20 0 20 40 60
range of values designing different classifiers and training
Fig6: Histogram plot of frontal electrode (F3) for
them for the two class movements to get the
pronation. best possible accuracy from the 16 channel
recorded sample values

COPO303-4
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27
2011

V. CONCLUSION
EEG data class separation was investigated
for two wrist movement classes, pronation
and extension using a 16 channel setup. Four
features were extracted from the histogram
plots of signal for both the movement and
neutral state. Variance, mean, skewness and
kurtosis are the features which easily
demarcates between the two movements. It
was a three class problem since three data
sets were used. Different classifiers were
designed and trained based on the recorded
data sets on the basis of which we classified
between motor movement

VI. ACKNOWLEDGEMENT
The authors are indebted to the UGC.This
work is a part of the funded major research
project CF.No 32-14/2006(SR)

REFERENCES

1. Wolpaw, J.R., Birbaumer, N.,

McFarland, D.J., Pfurtscheller, G., Vaughan,
T.M.: Brain–computer interfaces for
communication and control. Clinical
Neurophys. 113 (2002) 767–791
2. Vaughan, T.M.: Brain-computer
interface technology: a review of the second
international meeting IEEE Trans. on Neural
Systems and Rehab. Eng. 11(2) (2003) 94–
109
3. Syed M.Siddique, laraib Hassan
Siddique: EEG based Brain computer
Interface:Journal of software.4(6)(2009)550-
555
4. F. Sepulveda, M. MeckesCluster
:.Separation Index Suggests Usefulness of
Non-Motor EEG Channels in Detecting
Wrist Movement Direction Intention:
Proceedings of the 2004 IEEE Conference
on Cybernetics and Intelligent Systems

COPO303-5
 Color Images Enhancement Based on Piecewise
Linear Mapping
Anubhav kumar 1, Awanish Kr Kaushik 1 , R.L.Yadava 1, Divya Saxena 2
1
Department of Electronics & Communication Engineering
Galgotia’s College of Engineering & Technology, Gr.Noida, India
2
Department of Mathematics
Vishveshwarya Institute of Engineering and Technology , G.B.Nagar,India
rajput.anubhav@gmail.com

illumination and to enhance contrasts in the image. It’s a

Abstract- Image enhancement is a basic process for all vision and
image processing function in various areas like computer vision,
biomedical image analysis, forensic video/image analysis, remote
sensing and fault detection image enhancement technology plays
a very important role in image processing. By enhancing some
information and restraining other information selectively, it can
improve image visual effect. For a digital image enhancement
process, the digital image is represented in the YCbCr color
space. In this paper, an enhancement arithmetic is used for color
images enhancement based on piecewise linear mapping is
discussed. The method can improve the image sharpness and
make the image characteristics more evident. The enhancement
result is useful for further analysis. Experimental results tested
from a data set of images have demonstrated that the proposed
method is effective and practical. Some parameters like as
RFSIM (Image Quality Assessment) are analyzed for both
existing and proposed method to determine the success and
limitation of our method.
frequency filtering, preferred to others techniques because it
corrects non uniform lighting and sharpens the edges at the
same time [2]. Another method for color image enhancement
is Bright stretching, also used auto scaling method which is a
common linear mapping function to enhance the brightness
and contrast level of an image [10].
Color image processing and enhancement is a more
complicated process than its counterpart of black-and-white
images [4] due to the involvement of multiple color channels
and the need to preserve the color information content [5]
while enhancing the contrast. Novel techniques that address
these problems are in great demand. For instance, in [6], Chen
et al. proposed to enhance the image quality by making use of
local contrast information and fusing the morphologically
enhanced image with the original.

Keywords - Color image enhancement, RGB color space, YCbCr
color space, Piecewise linear mapping, RFSIM.
I.INTRODUCTION
Image enhancement processes consist of a collection of
techniques that seek to improve the visual appearance of an
image or to convert the image to a form better suited for
analysis by a human or machine. Nowadays there is a rapid
increase in the application of color video media. This has
resulted in a growing interest in color image enhancement
techniques.
Some other common technique to enhance the contrast of
images is to perform histogram equalization and
Homomorphic methods. The advantage of histogram
equalization technique is that it works very well for grayscale
images, however, when histogram equalization is used to
enhance color images, it may cause shift in the color scale
resulting in artifacts and an imbalance in image color. The
homomorphic filtering is used to correct non uniform
Similarly Fairweather [8] have used techniques such as
contrast stretching and Markov Random field. They applied
bimodal histogram model to the images in order to enhance
the underwater image. Yoav [9] have used a Physics-based
model. They developed scene recovery algorithm in order to
clear underwater images/scenes through polarizing filter. This
approach addresses the issue of backscatter rather than blur.
In this paper color images enhancement based on piecewise
linear mapping has been proposed .In Section II, an illustration
of theoretical foundations of piecewise linear mapping
function are presented. The proposed enhancement algorithm
is developed in Section III. Experiments conducted using a
variety of color images are described in Section IV and results
are discussed. A conclusion is drawn in Section V.
II. THEORETICAL FOUNDATIONS
Piecewise linear Mapping Function –
The enhancement is done by first transforming the intensity
values using a piecewise linear mapping function for the
intensity component. The piecewise linear mapping function
consists of three line segments as indicated in Figure 1, where
vmin and vmax denote the minimum and maximum intensity
levels in the original image, respectively. This type of
mapping function permits proper allocation of the dynamic
range to different ranges of intensity levels. The actual values
of vlower and vupper determine the dynamic range allocation
for lower, intermediate and higher ranges of intensity levels.
Fig. 1 The piecewise linear mapping function for intensity
Component.
The first moment of the intensity histogram serves as an
indicator about the exposure of the image. The first moment is
closer to vmin for an under exposed image, and closer to vmax
for an over exposed image. Thus based on the first moment of
the histogram, two intensity levels, vlower and vupper are
determined such that the closer the first moment is to vmzn,
,  < 1
the larger the vupper is, and the smaller the dower is, and vice
original image are transformed to create the optimally
enhanced color image in NTSC YCbCr space directly.
Fig. 2 Flow chart of Proposed Algorithm
Step 4 - In the last contrast increases by adjusting image
intensity values. It suppresses eventually outliers’ pixels to
improve contrast stretching. It then stretches contrast to use
the whole range of intensity channel and if necessary it
saturates some low or high values. Contrast stretching
attempts to improve the contrast in an image by stretching’ the
range of intensity values it contains to a desired range of
values, e.g. the full range of pixel values that the image
concerned allows.
 0 <

,
 

 
g,  = 0  0 >
,  
versa.

1  1 <
, 
(1)

III. PROPOSED ALGORTHEM

The color image enhancement based on piecewise linear
mapping function is specified as follows.
Step 1 – In the first step, firstly read the input image and
reshape the input color image (m*n size) in to (m*n)*1
column vector using array. After this change the Sort array
elements in ascending or descending order in the image has
been done.
Step 2 – In the second step, Employ ceil and array function
and find out Vmin and Vmax.Let the full dynamic range for the
display device be [VMIN, VMAX], where VMIN and VMAX
correspond to the minimum and maximum displayable
intensities, respectively.
Step 3 - Using the luminance and chrominance mapping
functions, the luminance and chrominance values of the
where, g( x,y ) represents the images, on the left side it
represents the output image, while f(x,y) represents input
image. In this equation the min and max are the minimum
intensity value and the minimum intensity value in the current
image.
Enhanced Image = g( x,y ) * No. of intensity levels (2)
Here No. of intensity levels shows the total number of
intensity values that can be assigned to a pixel. For example,
normally in the gray-level images, the lowest possible
intensity is 0, and the highest intensity value is 255. Thus "no.
of intensity levels" is equal to 255.
IV. RESULT AND ANALYSIS
On our personal laptop of Pentium 4 the experiment has been
carried out at 2.4 GHz. The test set for this evaluation
experiment randomly selected from the internet with 256*256
and 512*512 sizes. Figures(3a,4a,5a,6a ) show the original
color images, figure (3b,4b,5b,6b) show the enhanced image
by using Murtaza method [1], and the figure (3c,4c,5c,6c)
show the enhanced images with the help of our proposed .

(a ) (b)

(a ) (b)

(c)
Figure-4 –Watch image (a) Original image (b) Enhanced from
Murtaza method [1] (c) Enhanced from Proposed Method

(c)
Figure-3 -Lena image (a) Original image (b) Enhanced from Murtaza
method [1] (c) Enhanced from Proposed Method

(a ) (b)

(a ) (b)

(c)
Figure-4 –Play image (a) Original image (b) Enhanced from Murtaza
method [1] (c) Enhanced from Proposed Method

Table 1 show the comparison of performance of our proposed

(c)
Figure-4 –Flower image (a) Original image (b) Enhanced from
Murtaza method [1] (c) Enhanced from Proposed Method
method for enhancement with existing method, where our
proposed method has a better performance in contrast
enhancement and Image Quality Assessment (RFSIM). A
novel IQA model, namely Riesz-transform based Feature
Similarity index (RFSIM) method used to evaluate the existing
and proposed experiment result.
This naturally leads to the following formula to define the transforms", in: Proc. IEEE International Conference on Image
similarity between the feature maps fi and gi by considering Processing, IEEE 2010.
only the key locations marked by mask M, [4] N. M. Kwok, Q. P. Ha, D. Liu, and G. Fang, “Contrast

∑ ∑  ,.  ,
enhancement and intensity preservation for gray-level images

∑ ∑  ,
using multi objective particle swarm optimization,” IEEE
Di = (3) Trans. on Automation Science and Engineering, vol. 6, no. 1,
pp. 145–155, 2009.
The similarity between two feature maps fi (i = 1~5) and gi at the [6] S. K. Naik and C. A. Murhty, “Hue-preserving color image
corresponding location (x, y) is defined as enhancement without gamut problem,” IEEE Trans. on Image

 ,.!,"#
Processing, vol. 12, no. 12, pp. 1591–1598, 2003.

 $ ,.! $ ,"#
[7] Q. Chen, X. Xu, Q. Sun, and D. Xia, “A solution to the
di (x,y) = (4) deficiencies of image enhancement,” Signal Processing, vol.
90, pp. 44–56, 2010.
[8] A J R Fairweather, M A Hodgetts, A R Greig, “Robust

= ∏'() &
Then, we compute the RFSIM index between f and g as scene interpretation of underwater image sequences”, In 6th
International Conference on Image Processing and its
RFSIM [3] (5) Applications, 1997, pp. 660 -664, ISBN: 0 85296 692 X
[9] Schechner, Y and Karpel, N., “Clear Underwater Vision”.
Proceedings of the IEEE CVPR, Vol. 1, 2004, pp. 536-543.
Table-I [10] M.Isa,M.Y.Mashor,N.H.Othman,"Contrast Enhancement
Image Processing on Segmented Pap smear Cytology Images",
RFSIM (Image Quality Assessment) Prof. of Int. Conf. on Robotic, Vision, Information and Signal
Images Murtaza et al. [1] Proposed Method Processing.pp. 118 – 125, 2003.
Lena Image 0.3445 0.7261
Flower Image 0.4183 0.8321
Watch Image 0.4604 0.8953
Play image 0.4577 0.7614

The proposed model was implemented using five images. The

implementation results were compared with other results and
the performance of our proposed method is better than that of
existing color image enhancement algorithms.

V .CONCLUSION

In this paper color images enhancement based on piecewise

linear mapping has been presented. From the above results and
the Table -I, it is clear that the RFSIM (feature based image
quality assessment) of the proposed algorithm is high.
Experiments show that our proposed method obtains better
enhancement result, and it is more effective for color image
enhancement in normal and noisy environment

V. REFERENCE

[1]. Syed Mashhood Murtaza, Jahanzeb Ahmad, Usman Ali

‘Efficient Generalized Colored image Enhancement’ in 2006
IEEE conference.
[2]. C.M. Sujatha, K. Navinkumar, K. S. Arunlal, S. A. Hari,
Performance evaluation of Homomorphic filtering,
Anisotrophic filtering and Autocontrast algorithm,
International Conference on Advances in Computing, Control,
and Telecommunication Technologies, IEEE 2009.
[3] Lin Zhang, Lei Zhang, and Xuanqin Mou, "RFSIM: a
feature based image quality assessment metric using Riesz
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
Information Fusion Based on Multiplicity of
Data Preprocessing Improves AdaBoost
Classification Efficiency
Divya Somvanshi and R. D. S. Yadava
Abstract—The paper considers the possibility of boosting
efficiency of the AdaBoost algorithm by introducing diversity in
the information extraction procedures and subsequent
information fusion. A given raw data set is transformed by two
well known methods known as vector autoscaling and
dimensional autoscaling. These methods transform data by mean-
centering and variance-normalization with respect to the
measurement variables and the measured samples, respectively.
The feature (information) extraction is implemented by the
principal component analysis. Final data representation is
obtained by two alternate fusion strategies. In one, feature
extraction is done immediately after the scaling transformations
and the two feature sets are fused by a simple procedure of
concatenation. In the other, data vectors obtained by the two
scaling transformations are first fused by simple concatenation
and then the feature extraction is done. The data representation is
thus by fused feature vectors of dimensionality twice that of the
original data. The AdaBoost algorithm by Freund and Schapire
is implemented by using a simple threshold base classifier. The
classification efficiency of these procedures is compared with that
obtained by using single preprocessor based AdaBoosting. Four
benchmark data sets are used for validation. The analysis
demonstrates that the performance of the AdaBoost algorithm is
enhanced further by multiple preprocessor based fusion of data
space, particularly if the original variables are of same kind.
Index Terms—Boosting the AdaBoost algorithm, data
preprocessing, information fusion for pattern recognition,
machine learning.
I. INTRODUCTION
T he success in any pattern recognition problem depends on
three factors: quality of raw data, efficiency of information
extraction procedures, and classifier algorithm in
succession. An improvement in each boosts the performance
of its successor. The overall performance of the pattern
recognition algorithm depends on the optimization of all steps
Manuscript received March 5, 2011. This work was supported by the
Government of India, Defence Research and Development Organization
Grant No. ERIP-ER-0703643-01-1025.
Divya Somvanshi is with the Department of Electronics Engineering,
Institute of Technology, Banaras Hindu University, Varanasi 221005, India
(e-mail: divya.jrf@gmail.com).
R. D. S. Yadava is with the Department of Physics, Faculty of Science,
Banaras Hindu University, Varanasi 221005, India (phone: 91- 9450246572
mobile; fax: 91-542-2368390; e-mail: ardius@ gmail.com,
ardius@bhu.ac.in).
doing their operations in tandem. Given a raw data, the
objective of information extraction procedures is to represent
the data in a form that optimally helps the classifier. The
information extraction consists of cleaning up the data for
noise, removal of outliers, shifting and scaling, and
transformation to generate a new representation of data which
is potentially more helpful for pattern classification. The data
transformation involves combining raw data variables (often
linearly) to create new set of variables by focusing on certain
statistical attributes such as variance, statistical independence,
or separability. This is referred to as feature extraction. The
procedures of data preparation for feature extraction are
referred to as preprocessors. Some commonly used feature
extraction methods are principal component analysis,
independent components analysis and linear discriminant
analysis. The new variables (feature components) define the
feature space, and the representation of a sample vector in this
space is called the feature vector or pattern. The feature
vectors are representation of raw data vectors, and are input to
the classifier for discrimination of patterns of different classes.
The pattern classifier algorithms map feature space (in effect
input data space) to different class labels or object identities. A
number of pattern classification algorithms are developed that
use various strategies for creating input data space to output
class space mapping functions. For example, neural network
methods produce nonlinear mapping by optimizing feature
weights and node activation functions, support vector
machines create large dimensionality hyperspace using a
kernel function, and discriminant analysis seeks separation
boundaries in the feature space. An overview of these methods
can be found in [1], [2].
A pattern classification algorithm is basically a decision
function which computes decision values by taking feature
vectors as input. The construction of decision functions has
been inspired by a variety of paradigms from statistical and
biological domains. For example, the clustering and
discriminant function algorithms implement statistical models,
while the neural network and genetic algorithms are bio-
inspired [2]. Given a pattern classifier, weak or strong, its
performance can be further enhanced by methods of so called
„ensemble learning‟ [3]. The ensemble learning methods use a
number of classifiers and combine their decision functions in
some way to produce a new classifier with enhanced
capabilities. It was proved by Schapire [4] that a number of
weak learners trained on the same set of examples can be
COPO305-1
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

combined to yield a strong learner with enhanced predictive The class identities were given numeric representation of 1
capability. The ensemble learning methods produce a fused and 1 . That is, yi Y 1, 1 for all x i X for all
decision fusion. In principle the performance of any classifier i 1, 2,..., N .
can be enhanced. However, usually the risk of overfitting Step 1. Define input. It consists of N training examples, a
prompts to use an ensemble of weak learners [5]-[7]. There is base learning algorithm, and number of training runs T .
no limit on the number of learners in the ensemble. The Step 2. Initialize weight distribution over training examples
performance of an ensemble learning method can be optimized
according to w1 (i) 1 / N for i 1,2,..., N where i stands for
by proper selection of ensemble size and fusion method to suit
any specific application [5], [8]. the training examples and subscript 1 denotes the first round
Most popular among many ensemble learning algorithms is of T runs to determine weights wt (i ) . In the first round all
the AdaBoost (nickname for adaptive boosting) algorithm training examples are assigned equal weights. This assignment
described by Freund and Schapire [9]. A summary of the ensures that weight distribution is normalized, that is,
AdaBoost algorithm is given in section II. The objective of N
present study is to show that the predictive ability of the w1(i) 1.
i 1
AdaBoost algorithm can be further boosted by varied
Step 2. Start a loop „for t 1 to T ‟ that will create
representation of the same raw data by employing several
sequential base classifiers according to the following substeps
preprocessing and feature extraction methods. Different
in succession:
preprocessing and feature extraction procedures reveal hidden
Train the base classifiers ht based on the example
data structure from different perspectives, and yield alternate
sets of features to represent the same example. By combining weight distribution wt (i ) .
these sets in some way can, in principle, provide a more Determine the training error defined as t wt i
reliable and accurate representation. Motivated by this idea, i:ht xi yi
we report here a study on enhancing performance of the which is the sum of weights for all misclassified
AdaBoost algorithm by using a simple model of feature fusion examples by ht .
based on the two common preprocessors and one feature Assign weight to the t-th classifier ht according to:
extractor combination. The procedure is described in section
III. Using a linear threshold classifier for the AdaBoost 1 1 t
t ln .
ensemble generation, section IV presents validation results 2 t
based on some benchmark data sets available from open Update the example weights according to:
sources. The paper concludes with a discussion in section V.
wt i e t for ht xi yi
wt 1 i
zt e t for ht xi yi
II. THE ADABOOST ALGORITHM where z t is the normalization factor for making
The AdaBoost [9] is a supervised learning boosting wt 1 (i) to be a probability distribution. That is, for
algorithm that produces a strong classifier by combining N
several weak classifiers from a family. It needs a set of training making wt 1 (i) 1.
examples and a base learning algorithm as input. Let i 1
X x1 , x 2 ,..., x N be the set of N training vectors drawn End loop.
Step 3. Output the boosted classifier as weighted sum of
from various target classes, and let Y y1, y2 ,..., y N denote
the T base classifiers
their class labels. The base learner is chosen such that it T
produces at least more than 50% correct classification on the H x sign t ht x .
training set. Let base learner be denoted by ht x . The basic t 1

idea of ensemble generation in the AdaBoost algorithm is to

assign a distribution of weights on the training examples
according to the error rate in such a manner that, in
sequentially repeated application of the base learner, the
misclassified examples get more weight in the current
application than the preceding one. The base classifier in each
application is assigned a weight value t according to its error
rate. The process is iteratively repeated until the error rate is
minimized. The ensemble consists of all base classifiers
generated thus in successive runs until the error rate minimum.
The boosted classifier, say H x , is obtained as weighted-sum
of all the base classifiers in the ensemble.
The steps of the AdaBoost algorithm are reproduced here
for the sake of clarity in the presentation of present analysis.
We consider the AdaBoost solution for a two-class problem.
The Base Classifier
In this work, we have taken the base learning algorithm to
be a simple threshold classifier which uses only one feature
dimension for class discrimination. The selection of that
particular feature dimension is made by searching over all the
feature dimensions while monitoring the rate of
misclassification according to certain threshold criterion. The
dimension with minimum rate of misclassification is selected
for defining the base classifier. The error rate for each
dimension is calculated as the sum of weights of all
misclassified training samples as defined in Step 2 above. The
threshold is a moving hyperplane which is perpendicular
bisector to feature components of consecutive training
samples. The selection of minimum error dimension with

COPO305-2
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

associated threshold definition and samples weight distribution

Feature Set 1
specify the base classifier in each round of the AdaBoost run.
In the present calculations we used the AdaBoost package Principal
available in MatLab [10]. Component
Analysis

Vector Autoscaling
III. MULTIPLE PREPROCESSORS AND INFORMATION FUSION Data Set 2

Fusion of information means to combine or assemble

Data Space Input Feature Space
information measures obtained from different sources, or Fusion Data Fusion
extracted by different methods from the same data source, or
both, in such a way that the fused information becomes Dimensional Autoscaling
Principal
potentially more efficient for target representation. In this Component
Data Set 1
work, we analyze the effects of information fusion on the Analysis
Principal
performance of AdaBoost algorithm as outlined schematically Component
in Fig. 1. The information fusion is implemented in two Feature Space 1
Analysis
Feature Space 2
alternate ways which have been referred to in the schematic as
Feature Set 2
data space fusion and feature space fusion. The two most
AdaBoosting 1 AdaBoosting 2
commonly used mean centering and data scaling methods
called the vector autoscaling and the dimensional autoscaling
have been used. These are defined as follows. Let M denote Error Minimization Error Minimization

the dimensionality of raw data, and N before the number of

training examples of samples. The training data set will thus be Fig. 1. Schematic of data preprocessing, feature extraction and adaboosting.
a N M matrix with samples in rows and variables in
columns.
dimensional autoscaled data separately. Then, the fusion is
Vector autoscaling done by a similar concatenation of the feature components as
The matrix elements are mean-centered and variance- explained above. The final feature spaces created by the
normalized for each dimension separately (column wise) as methods are of the same size, N 2M matrix with N training
xij xi examples with increased dimensionality equal to 2M .
xij
i

1 N 1 N
where xi xij and ( xij xi ) 2 . IV. VALIDATION
i
N j 1 N j 1 Four data sets of two class problems have used in the
validation analysis. These data were collected from the UCI
Dimensional autoscaling machine learning repository. When analyzed by a single strong
The matrix elements are mean-centered and variance classifier (backpropagation neural network) in combination
normalized for each sample separately (row wise) as with dimensional autoscaling and PCA feature extraction, the
xij x j classification rate for the two data sets (sonar and heart) were
xij typically <60% and that for the other two (Haberman breast
j cancer and Pima Indian diabetes) were typically >70%. The
1 M 1 M analysis of the same data sets were done by the proposed
where x j xij and j ( xij x j ) 2 . method consisted of the weak threshold classifier based
Mi1 M i 1 AdaBoost algorithm and the two methods of feature extraction.
The feature extraction has been done by the principal The division of the available data between the training and the
component analysis (PCA). test sets were done by random selection in nearly 50-50 ratio.
In data space fusion, the sample vectors transformed by the The description of the data sets is given Table 1.
two methods were fused by simple concatenation of the vector Table 2 presents the best classification results obtained by
components. That is, if the i-th training vector processed by the AdaBoost algorithm with the linear threshold base
the vector autoscaling is x1i {x1ij } {x1i1, x1i 2 ,..., x1iM } and that classifiers as described in the preceding. The data are
processed by four combinations of preprocessing, fusion and
processed by the dimensional auto scaling is
feature extraction strategies before AdaBoosting. These
xi2 {xij2} {xi21, xi22 ,..., xiM
2
} then the i-th fused data vector is combinations are: vector-autoscaling + PCA; dimensional-
autoscaling + PCA; (vector-autoscaling + dimensional-
defined by z i {x1i1, x1i 2 ,..., x1iM , xi21, xi22 ,..., xiM
2
} . The feature
autoscaling) data space fusion + PCA; and, (vector-autoscaling
extraction is done by PCA of the new N 2M dimensional + PCA) + (dimensional-autoscaling + PCA) feature space
data matrix. fusion. It can be seen that in all cases the performance of the
In feature space fusion, two alternate feature spaces are AdaBoost algorithm has improved after the multiple
created first by the PCA of the vector autoscaled and the

COPO305-3
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

preprocessor based data space or feature space fusions. The vector autoscaling dimensional autoscaling
amount of improvement however depends on the type of data. 1 1

In the Haberman breast cancer data and the pima-indian

error rate

error rate
diabetes data the improvements are very marginal. However, 0.5 0.5
in case of sonar data and heart data the improvements are
substantial after data space fusion.
0 0
20 40 60 80 20 40 60 80
TABLE I
DATA SETS USED IN PRESENT ANALYSIS data space fusion feature space fusion
1 1
Data Classes Sample Attribute Remark
s s

error rate

error rate
Sonar 2 208 60 Classes: sonar returns from a 0.5 0.5
metal cylinder and from a
similarly shaped rock.
Attributes: integrated energy
within a frequency band. 0 0
20 40 60 80 20 40 60 80
Heart 2 267 22 Classes: cardiac normal and X - axis: number of threshold classifiers in AdaBoost ensemble
abnormal condition.
Attributes: SPECT images. Fig. 1. Variation of error rate for sonar test data with ensemble size in
Haberman‟s 2 306 3 Classes: patients‟ survival after AdaBoost algorithm for linear threshold classifiers.
Breast-Cancer 5 years and death within 5 years
of breast cancer surgery. vector autoscaling dimensional autoscaling
1 1
Attributes: Age, year of
operation, positive axillary
nodes.

error rate

error rate
Pima-Indian 2 768 8 Classes: signs or no-sign of 0.5 0.5
Diabetes diabetes pima-indian females
above 21 years age.
Attributes: patient‟s history and 0 0
physiological parameters 20 40 60 80 20 40 60 80

TABLE II data space fusion feature space fusion

ADABOOST CLASSIFICATION RESULTS OBTAINED UNDER DIFFERENT 1 1
CONDITIONS OF PREPROCESSING AND FUSION
error rate

error rate
Classification Rate (%) 0.5 0.5

Data Data Feature

Vector Dimensional
Space Space
Autoscaling Autoscaling 0 0
Fusion Fusion 20 40 60 80 20 40 60 80
X - axis: number of threshold classifiers in AdaBoost ensemble
Sonar 53.66 59.76 75.61 53.66
Fig. 2. Variation of error rate for heart test data with ensemble size in
Heart 66.31 67.92 79.68 70.06 AdaBoost algorithm for linear threshold classifiers

Haberman vector autoscaling dimensional autoscaling

Breast- 75 71.7 76 75 1 1
Cancer
error rate

error rate

Pima- 0.5 0.5

Indian 66.7 77.7 76.7 78.4
Diabetes
0 0
20 40 60 80 20 40 60 80
The effect of present boosting strategy of the AdaBoost
data space fusion feature space fusion
algorithm can be seen from the error plots shown in Fig. 1 1 1
through Fig. 4 for the four sets of test data used in the analysis.
The error rate is defined as the fraction of misclassified test
error rate

error rate

samples, and the plots show the variation of error rate with the 0.5 0.5

number of base classifiers in the ensemble for AdaBoost. A

notable point from these plots is that the boosting strategy
0 0
yields good results only when the base classifier is weak. For 20 40 60 80 20 40 60 80

example, the selected threshold classifier results in the initial X - axis: number of threshold classifiers in AdaBoost ensemblle

error rates for the sonar and heart data close to 50%, and for
the breast-cancer and diabetes data close to 25%. The
AdaBoosting reduces error rate significantly in case of former,
Fig. 1 and Fig. 2, but not so much in case of latter, Fig. 3 and
Fig. 4. The similar impact on the classification rates is
Fig. 3. Variation of error rate for Haberman test data with ensemble size in
AdaBoost algorithm for linear threshold classifiers.
apparent from the results in Table II. Another notable point is
that the data space fusion facilitates better boosting.

COPO305-4
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
vector autoscaling
1 1
error rate
error rate
0.5 0.5
0 0
20 40 60 80
data space fusion
1 1
error rate
error rate
0.5 0.5
0 0
20 40 60 80
X - axis: number of threshold classifiers in AdaBoost ensemble
Fig. 4. Variation of error rate for pima-indian diabetes test data with
ensemble size in the AdaBoost algorithm for linear threshold classifiers.
V. DISCUSSION AND CONCLUSION
Sonar data is a complex distribution of energy of chirped
sonar signals in different frequency bands returned from two
types of targets (metallic cylinder and cylinder shaped rock).
All the attributes are therefore of the same kind. Besides, there
could be appreciable correlation between different attributes in
raw data. The best result for this data set is obtained by the
combination of data space fusion with AdaBoosting. The error
rate on the training data set drops to 0 after 20 rounds of base
learner creation. On the test data set however the error rate
continued to decrease up to 80 rounds, Fig. 1. The use of other
preprocessing methods did not produce much boosting for
ensemble sizes. The use of dimensional autoscaling and
feature space fusion reduced the error rate significantly after
only few rounds of iteration. Later, the error rate increased.
The vector autoscaling did not produce boosting effect in any
condition, Fig. 1.
Heart data is diagnostic cardiac data based on SPECT
(single proton emission computed tomography) images for
patients belonging to two categories: normal and abnormal.
The classification result in Table II and error result in Fig. 2
indicates a trend similar to that for sonar data. The
combination of data space fusion with AdaBoosting yields the
best result. The attributes of image data are also likely to be
correlated.
In contrast, the Haberman‟s survival data after breast cancer
surgery and the Pima Indian Diabetes data consist of patient‟s
history and physiological parameters like number of positive
axillary nodes and body glucose level. The variables in these
data sets are of different types and do not seem to have direct
correlation. The AdaBoosting in any combination of
preprocessing fusion strategy does not yield significantly
enhanced classification rate.
It appears therefore that the strategy of multiple
preprocessors based fusion of information enhances the
efficiency of the AdaBoost algorithm in those multivariate
situations where the variables are of the same type, and are
possibly correlated. The data preprocessing by the vector
dimensional autoscaling
autoscaling and the dimensional autoscaling followed by data
space fusion and feature extraction by principal component
analysis enhances the classification efficiency of the
AdaBoosted linear threshold classifier by 12% to 22% for the
sonar and heart data compared to the AdaBoosting without
fusion. We thus conclude that by bringing in diversity in the
preprocesssing methods for data representation to the feature
20 40 60 80
extractor yields more accurate feature set, which further
feature space fusion
enhances the efficiency of the AdaBoost algorithm.
ACKNOWLEDGMENT
This work is supported by the Government of India,
Defence Research & Development Organization Grant No.
20 40 60 80
ERIP-ER-0703643-01-1025. The authors are thankful to Mr.
Prashant Singh and Miss Prabha Verma for their support and
several useful discussions.
REFERENCES
.
[1] A. K. Jain, R. P. W. Duin, and J. Mao, “Statistical pattern recognition: a
review,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol.
22, no. 1, pp. 4-37, Jan. 2000.
[2] R. J. Schalkoff, Pattern Recognition – Statistical, Structural and
Neural Approaches, Wiley & Sons, 1992, ch. 1.
[3] T.G. Dietterich, “Ensemble learning,” in The Handbook of Brain
Theory and Neural Networks, 2nd ed. M.A. Arbib, Ed. Cambridge,
MA: The MIT Press, 2002, pp. 405-408.
[4] R. E. Schapire, “The strength of weak learnability,” Machine Learning,
vol. 5, no. 2, pp. 197-227, 1990.
[5] T. G. Dietterich, “Ensemble methods in machine learning,” in Lecture
Notes in Computer Science, J. Kittler, F. Roli, Eds. Berlin Heidelberg:
Springer-Verlag, vol. 1857, 2000. pp. 1–15.
[6] Y. Freund, Y. Mansour, R. Schapire, “Why averaging classifiers can
protect against overfitting,” in Artificial Intelligence and Statistics 2001
(Proc. of the Eighth International Workshop: January 4-7, 2001, Key
West, Florida), T. Jaakkola, T. Richardson, Eds. San Francisco CA:
Morgan Kaufmann Publishers, 2001.
[7] D. Chen, J. Liu, “Averaging weak classifiers,” in Lecture Notes in
Computer Science, J. Kittler and F. Roli, Eds. Berlin Heidelberg:
Springer-Verlag, vol. 2096, 2001, pp. 119-125.
[8] G. Levitin, “Threshold optimization for weighted voting classifiers,”
Naval Research Logistics, vol. 50, 2003, pp.322-344.
[9] Y. Freund, R. E. Schapire, “A decision-theoretic generalization of on-
line learning and an application to Boosting,” Journal of Computer and
System Sciences, vol. 55, no. 1, 1997, pp. 119–139.
[10] Cuneyt Mertayak (2007, may 25). AdaBoost, version 1.0. Available:
http://www.mathworks.com/matlabcentral/fileexchange/21317-
adaboost.
[11] R. G. Osuna, H. T. Nagle, “A method for evaluating data preprocesssing
techniques for odor classification with an array of gas sensors,” IEEE
Trans. Syst. Man Cybern. B, vol. 29, May 1999, pp. 626–632.
COPO305-5
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Real-time Hand Gesture Recognition Techniques for

Human-Computer Interaction
ANSHUL SHARMA
Department of Electronics and Communication Engineering, MNNIT Allahabad
Email ID: getanshulsharma@gmail.com

Abstract- This paper presents some real-time including the complicated nature of static and
video processing techniques for intelligent and dynamic hand gestures, complex backgrounds, and
efficient hand gesture recognition, with an aim occlusions. Attacking the problem in its generality
of establishing a virtual interfacing platform for requires elaborate algorithms requiring intensive
Human-Computer Interaction (HCI). The first computer resources. Due to real-time operational
step of the process is colour segmentation based requirements, we are interested in a
skin detection, followed by area-based noise computationally efficient algorithm.
filtering. If a gesture is detected, the next step is
to calculate a number of independent Previous approaches to the hand gesture
parameters of the available visual data and recognition techniques include the use of markers
assign a distinct range of values of each on various points on the hand, including fingertips.
parameter to a predefined set of different Calculation and observation of relative placement
gestures. The final step is the hierarchical and orientation of these markers specifies a
mapping of the obtained parameter values to particular gesture. The inconvenience of placing
recognise a particular gesture from the whole set. markers on the user’s hand makes this an infeasible
Deliberately, the mapping of gestures is not approach in practice. Another approach is to use
exhaustive, so as to prevent incorrect mapping sensor- fitted gloves to detect the orientation and
(misinterpretation) of any random gesture not other geometrical properties of the hand. The
belonging to the predefined set. The applications demerit of this approach is its cost ineffectiveness.
of the same are inclusive of, but not limited to,
Sign Language Recognition, robotics, computer The approach proposed in this text is quite user-
gaming etc. Also, the concept may be extended, friendly as it does not require any kind of markers
using the same parameters, to facial expression or special gloves for its operation. Also, the
recognition techniques. memory requirements are low because the
subsequent video frames are not stored in memory,
they are just processed and overwritten. Obviously,
1. INTRODUCTION it adds a new challenge to make the algorithm very
Gestures and gesture recognition are terms fast and efficient, fulfilment of which is ensured by
increasingly encountered in discussions of human- using low-complexity calculation techniques. For
computer interaction. The term includes character the ease of implementation, the proposed algorithm
recognition, the recognition of proof readers is based on three basic assumptions:
symbols, shorthand, etc. Every physical action
involves a gesture of some sort in order to be 1. The background should be dark.
articulated. Furthermore, the nature of that gesture 2. The hand should always be at a constant
is generally an important component in establishing distance from the camera.
the quality of feel to the action. The general 3. There should be a time gap of at least 200
problem is quite challenging due a number of issues ms between every two gestures.

COPO306-1
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
2. DATA EXTRACTION
The concept of video processing is based upon
data extraction from subsequent frames of the video.
The video is accessed frame by frame and the
following actions are performed on every
subsequent frame: (Some frames may be skipped,
without loss of information, because at an average
frame rate of 30 fps, any two subsequent frames
will have almost the same data.)
2.1 Hand Block Detection:
The first step of the gesture recognition process is
to detect if there is a gesture at all. It is done by skin
colour detection and area based detection of the
hand block. Also, the gesture should only be taken
into consideration if the whole hand is captured in
the video frame. This is achieved by calculating the
geometrical centroid of the hand block and
imposing a range restriction on the coordinates of
the centroid.
2.2 Elliptical Approximation:
The concept of elliptical approximation is the
heart of the gesture recognition process discussed in
this text. Once the hand block is detected, i.e. the
presence of a gesture (valid or invalid) is ensured, it
is approximated by an ellipse having the same
second-moments as the hand block detected.
Having obtained such a mathematical modelling,
we can interpret a number of independent
parameters from the geometrical properties of this
ellipse. This concept is depicted in Figure 1.
Figure 1: Elliptical approximation of hand block
The Orientation of the hand object is simply the
inclination of the major axis of the ellipse thus
obtained, from the positive horizontal axis.
Orientation is measured in a range of -90 to +90
degrees, as depicted in Figure 2.
Figure 2: Orientation of the hand block
The Eccentricity of the ellipse is simply the ratio
of semi-major axis (the distance between the two
foci) to the major axis. The value of eccentricity of
an ellipse lies between 0 and 1, both inclusive. An
ellipse with zero eccentricity is actually a circle.
The other extremity (eccentricity = 1) represents a
straight line. Thus, the eccentricity is a measure of
the narrowness of the ellipse. Higher be the
eccentricity, narrower is the ellipse (and vice versa).
2.3 Counting the number of fingers:
A very fast approach to count the number of
fingers in a gesture is discussed in this text, using
the data extracted by the means of elliptical
approximation as discussed already. With the vertex
of the ellipse as centre and a radius equal to 0.35
times the length of major axis, a circle is drawn.
The number of times the periphery of this circle
intersects any part of the hand block gives a direct
measure of the number of fingers in the gesture.
Figure 3 explains the concept. Clearly, if the circle
intersects the hand block at N places, then (N-1) is
the number of fingers in the gesture. As far as the
selection of the radius is concerned, 0.35 times the
major axis length was experimentally found to be
COPO306-2
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
the most optimum. A larger scaling factor (greater
than 0.4) entirely covers the smaller fingers and
hence, the presence of thumb and the little finger in
a gesture can’t be detected. On the other hand, a
small scaling factor (less than 0.3) is unable to
distinguish between two fingers separated by a
small gap, because the gap between the fingers is
lesser near the palm.
Figure 3: Counting the number of fingers
2.3 Perimeter Estimation:
Another distinguishing characteristic of a gesture
is the length of the periphery of the hand object.
Perimeter is calculated by simply counting the
number of pixels on the boundary of the hand
object.
Figure 4(a)
Figure 4(b)
Figure 4: Gestures having equal Areas but unequal Perimeters
Its significance can be appreciated by
acknowledging the fact that two gestures with same
area can have different perimeters. The hand
objects in Figure 4(a) and Figure 4(b) cover
approximately the same area, but the perimeter of
the hand object in Figure 4(a) is larger than that of
the hand object in Figure 4(b).
3. GESTURE RECOGNITION PROCESS
Having calculated the various parameters of the
hand block by mathematical modelling using
elliptical approximation, we now assign a unique
set of values of all these parameters to every gesture,
to be able to distinguish each one uniquely. Instead
of calculating all the parameters for every gesture, it
is efficient to categorize the gestures into different
hierarchical levels, so that only relevant properties
of every gesture are calculated, to make the
algorithm time-efficient. Statistically speaking, the
most optimum order is achieved by adopting the
following hierarchical recognition process:
1. Eccentricity. The first step is to mathematically
model the gesture by elliptical approximation and
calculate the eccentricity of the ellipse thus
obtained. If the eccentricity is too low (less than
0.5), it shows that the ellipse is wide enough
(approaching a circular shape), as is the case with
the gestures shown in Figure 5.
COPO306-3
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

three. This is obvious from Figure 4, which shows

one case in which perimeter estimation is required
as a distinguishing feature. Another similar case is
complimentary to these gestures, with negative
Orientation.

(a)

Figure 6: Gestures with Eccentricity > 0.95

4. Orientation. Orientation calculation is necessary

for every gesture, irrespective of the values of the
other parameters associated with it. This is because
the proposed gesture set is so made that every
gesture has a complimentary gesture whose all the
properties are the same as the former, with the only
difference being that its Orientation is the negative
(b) of that of former. Since Orientation calculation is a
Figure 5: Gestures with Eccentricity < 0.5 must for every gesture, this step may be executed at
any level, without any effect on the complexity of
On the other hand, if the eccentricity is too high the algorithm. Keeping it at the end of the process is
(greater than 0.95), it shows that the ellipse is very just one of the several equally efficient options.
narrow (almost a straight line), which is the case
with the gestures shown in Figure 6. In both these 4. THE ENTIRE GESTURE SET
cases, there is no need to count the number of
fingers, as it is obviously zero. The algorithm and the parameters already
discussed altogether can uniquely recognise a
2. Number of Fingers. If the value of Eccentricity gesture out of a closed set of 32 distinct gestures.
lies between 0.5 and 0.95 (both exclusive), the The complete gesture set is shown in Figure 7.
algorithm to calculate the number of fingers is
applied to the gesture frame. 5. APPLICATIONS
3. Perimeter Estimation. Perimeter estimation is Real-time gesture recognition is an emerging and
necessary as a distinguishing parameter only in a challenging field of Computer Vision. The main
few cases, when the number of fingers is less than applications of the same include, but are not limited
to:

COPO306-4
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
Figure 7: The Entire Gesture Set
Human- Computer Interaction (HCI).
Wireless robotic control mechanism.
Virtual Keyboard Interfacing Platform.
Security systems.
6. SCOPE FOR FUTURE WORK
The proposed gesture recognition techniques can
prove to be an integral basis for research work in
certain emerging fields, viz. Computer Vision, Real
Time Signal Processing, HCI, etc. Some major
fields that provide a challenging scope for the
future work of this text are discussed below:
Motion Gestures: So far we have discussed only
the recognition of static gestures. The above [3].Rafael C. Gonzalez, Richard Eugene Woods: “Digital
mentioned techniques, coupled with a concept Image Processing”, Third Edition, Prentice Hall Publications,
of motion recognition, can be used for
recognition of motion gestures.
No Background Limitation: The above
mentioned techniques are based on the
assumption that the background is always dark.
However, using dynamic colour segmentation
schemes, the same techniques can be applied
without this limitation also.
Double-Handed Gestures: All the same
concepts can be easily extended to gestures
involving both the hands at once. Doing so will
increase the size of the gesture set by about 5
times (with a substantial increase in complexity,
though).
Three- dimensional Gestures: The proposed
techniques deal with the recognition of only
two-dimensional gestures (hence the
assumption that the hand should always be at
constant distance from the camera, as discussed
earlier). However, if we use an additional
camera, orthogonal to the first one, we can
obtain the information about the distance of the
hand from the camera. This can be helpful for
efficient recognition of three-dimensional
gestures.
Facial Expression Recognition: The concept of
mathematical modelling of gestures using
Elliptical Approximation can be quite efficient
for the recognition of facial expressions.
Artificial Intelligence: Real-time Gesture
Recognition, coupled with a concept of Neural
Networks, may prove to be an integral
milestone in the field of Artificial Intelligence.
7. REFERENCES
[1]. Juan P. Wachs, Helman Stern, Yael Edan: “Cluster
Labeling and Parameter Estimation for the Automated Setup
of a Hand-Gesture Recognition System”, IEEE Transactions
on Systems, Man and Cybernetics—Part A: Systems and
Humans, Vol. 35, No. 6, November 2005.
[2]. Hyung-Ji Lee, Jae-Ho Chung: “Hand Gesture Recognition
Using Orientation Histogram”, 1999 IEEE Tencon.
2008.
COPO306-5
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Forensic Research by Face Reading Systems

Pramod Kumar Pal
Department of Computer Science & Engineering,
Radha Govind School of Engineering,
Radha Govind Group of Institutions
(U.P. Technical University)
Meerut-250004
Abstract: A contemporary definition for
The nature of computer vision causes the personality is offered by Carver and
fact that not only computer science Scheier: “Personality is a dynamic
researchers are interested in it, but organization, inside the person, of
neuroscientists and psychologists, too. psychophysical systems that create a
One of the main interests for psychology person’s characteristic patterns of
is identification of person’s behavior, thoughts, and feelings” [1]. In
psychological traits and personality other words, personality is a complex
types which can be accomplished by combination of traits and characteristics
different means of psychological testing: that determines our expectations, self-
questionnaires, interviews, direct perceptions, values and attitudes, and
observations, etc. Though that is a predicts our reactions to people, subjects
general tendency of people to read and events.
character into a person’s physical form, According to the Diagnostic and
especially face. In relation to Statistical Manual of the American
psychological characteristics Psychiatric Association [2], personality
recognition, face provides researchers traits are “enduring patterns of
and psychologists with instrument of perceiving, relating to, and thinking
obtaining information about personality about the environment and oneself that
and psychological traits that would be are exhibited in a wide range of social
much more objective than questionnaires and personal contexts.” Traits and
and neuropsychological tests and could characteristics are the same thing, as a
be obtained remotely using person’s trait is a distinguishing characteristic,
facial portrait, with no need for personal feature or quality. Theorists generally
involvement. The paper describes assume that a) traits are relatively stable
approaches to psychological over time, b) traits differ among
characteristics recognition from facial individuals, and c) traits influence
image such as physiognomy, phase behavior. Many psychologists have
facial portrait, ophthalmogeometry, and studied personality traits, and have
explains the need in automating it. attempted to identify or define all of the
Keywords: human traits. Some examples of
psychometrics, psychological personality traits identified and studied
characteristics, personality, Myers- by psychologists are: warmth, emotional
Briggs typology, ophthalmogeometry, stability, independence, dominance,
physiognomy, pattern recognition, face impulsivity, sensitivity, imagination,
recognition, facial expression introversion-extroversion,
recognition . suspiciousness, etc.
Introduction There is continuing debate about how
many different personality traits exist.

COPO307-1
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
The most popular model of
psychological traits is a five-dimension
personality model named as the “Big
Five” and proposed by Lewis Goldberg:
Extraversion, Neuroticism,
Agreeableness, Conscientiousness,
Openness to experience.
The difference between personality type
and personality trait should be
considered. Personality types are
distinguished from personality traits,
which come in different levels or
degrees. According to type theories, for
example, there are two types of people,
introverts and extraverts. According to
trait theories, introversion and
extraversion are part of a continuous
dimension, with many people in the
middle.
Identification of psychological
characteristics is the task widely used in
theoretical and practical psychological
research, education, coaching, career
guidance and hiring process, business
and political affairs, psychotherapeutic
diagnostics, self-exploration and
awareness, etc. Teachers evaluate
student’s personality traits and cognitive
abilities to know how to represent
material better and how to establish
communication and learning process in
more efficient way. Lawyers are
beginning to use personality testing for
criminal behavior analysis, litigation
pro- filing, witness examination and jury
selection. Medical stuff analyses
personality characteristics and observes
patient’s psychological state in regard of
its influence on medical treatment
process. Even websites design and
software interfaces are now developed
based on knowledge of user’s
personalities and their preferences.
It’s hard to underestimate the role of
identification of psychological
characteristics in modern society, as
everyone needs to know people’s
psychological traits to understand or
predict their reactions to various subjects
and situations.
2. Face analysis and other methods for
psychological characteristics
recognition
The question of methods and
instruments which may be used for
effective psychological characteristics
identification exists. To measure human
characteristics or identify personality
types psychological researchers apply
psychometrics which is the field of study
including the theory and technique of
psychological measurement primarily
concerned with the study of differences
between individuals. Psychometrics
involves such research tasks as the
construction of instruments and
procedures for measurement; and the
development of theoretical approaches to
measurement. For instance, the study
states that a useful psychological
measure must be both valid (i.e., actually
measures what it claims to measure) and
reliable (i.e., internally consistent or give
consistent results over time).
The most accepted means of measuring
personality using psychometrics study
is psychological questionnaires and
tasks. There are two major types of such
personality tests. Projective tests assume
that personality is primarily unconscious
and assess an individual by how he or
she responds to an ambiguous stimulus.
The idea is that unconscious needs will
come out in the person’s response.
Objective tests assume that personality is
consciously accessible and measure it by
self-report questionnaires. Research on
psychological assessment has generally
found that objective tests are more valid
and reliable than projective tests.
Examples of testing questionnaires and
tasks are Holland Codes, Rorschach test,
COPO307-2
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
Minnesota Multiphasic Personality
Inventory, Myers-Briggs Type Indicator,
Enneagram Type Indicator, NEO PI-R,
Thematic Apperception Test, Kelly’s
Repertory Grid, Online Depression
Screening Test, Online Screening for
Anxiety and so on [3].
Except psychological questionnaires and
tasks, interviews and direct observation
tests are applied. Interviews often
contain some observation and questions
about person’s past, abilities and
attitudes. Direct observation of people as
they complete activities may be with
clinical purpose, such as to observe the
nature of a parent-child interaction in
order to understand a relational disorder.
The Parent-Child Interaction
Assessment-II [4] is an example of a
direct observation procedure that is used
with school-age children and parents: the
parents and children are videotaped
playing at a make-believe zoo. Direct
observation procedures are also used in
research, for example to explore
sequences of behavioral interaction.
There is a range of approaches to
measuring personality and psychological
characteristics which are based on
human physical appearance. They are
described below:
Evaluation based on face
The face plays a crucial role in human
social cognition. Facial expressions
are important signals of internal states –
emotions and intentions. Humans also
see in the face signals of internal
qualities that are more stable over time,
like attractiveness as a mate [5,6,7] or
dominance [8,9,10], etc. According to
Liggett [11], “There can be little doubt
that the face plays a crucial part in our
everyday assessment of our fellows. Not
only does it enable us to identify
transient emotions – flashes of pleasure
and rage, disappointment and hatred – it
can also help us to make useful
judgments about more durable and
lasting qualities of personality and
character”. Judging personality from
face comes from ancient times [12,13] to
nowadays. The belief that the face
reveals information about underlying
character cuts across national, cultural
and geographical boundaries. Several
classical Greek and Roman scholars,
including Plato and Aristotle, argued that
faces contain clues about people’s
underlying personalities and
dispositions. For over 2500 years many
notable philosophers, historians and
physicians openly supported the idea
that a person’s disposition is reflected in
their face.Widespread interest in
physiognomy – the study of the face and
its relationship to human ability,
potential and character
– peaked at the end of the eighteenth
century when a physician and pastor
named Johann Kaspar Lavater produced
a formal classification system and set of
rules specifying the relationship between
the face and the mind [14]. Lavater
believed that character is assessed just as
well from the shape and size of the
forehead, for example, as from
observations of behavior. Indeed,
eighteenth and nineteenth century courts
commonly used physiognomy to assess
immoral tendencies in suspected
criminals.
Among contemporary applications of
face analysis not only psychological
interpretations are accomplished, but
also medical results are obtained. For
example, more than 700 genetic issues
influencing facial structure and facial
features are known, and special software
for disease identification from face had
been developed [15,16].
Body types
COPO307-3
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

William Sheldon classified personality type of mimics and voice may define
according to body type [17]. He called personality traits. For example, it’s used
this a person’s somatotype and identified in socionics (see Table 2) that is a
three main somatotypes shown in branch of psychology based on Carl
Table 1. Sheldon’s somatotypes and Jung’s work on Psychological Types.
character interpretations Moreover, many socionics experts use
the visual method of personality
Sheldon’s Character Shape
Somatotype
characteristics identification as a main
method for personality traits and types
Endomorph Relaxed, sociable, Plump, recognition.
[viscerotonic] tolerant, buxom,
comfort-loving, developed
Table 2. Example of some outer
peaceful visceral appearance characteristics and their
structure interpretation
Mesomorph Active, assertive, Muscular
[somatotonic] vigorous,
OUTER APPEARANCE
combative No Physical Sensoring Intuitive
characte
Ectomorph Quiet, fragile, Lean,
[cerebrotonic] restrained, delicate,
r
non-assertive, poor 01 The Short and Lengthy
sensitive muscles form of thick, and thin,
bones muscles are muscles
Person is rated on each of these three and pronounced aren’t
dimensions using a scale from 1 (low) muscles pronounce
to 7 (high) with a mean of 4 (average). d
Therefore, for example, a person who is 02 Form of Sen Sen Intu Intui
a pure mesomorph would have a score of the nose sori sori itive tive
1-7-1. ng ng + +
In Ayurvedic medicine (used in India + + Ethi Logi
since ˜3000 BC) there are three main Log Eth cal cal
metabolic body types (doshas) – Vata, ical ical
Pita, & Kapha – which in some way «triangle
correspond to Sheldon’s somatotypes. Horizonta with peak
Body types have been criticized for very l line in on the top»
weak empirical methodology and are not the nose «triangle
generally used in Western psychology bridge with peak
(they are used more often in alternative in the
therapies and Eastern psychology and bottom»
spirituality).
Complex physical appearance Neuropsychological tests
evaluation Around the 1990s, neuroscience entered
This is approach of evaluation of face the domain of personality psychology.
and body parts in complex, and it is It introduced powerful brain analysis
considered to be physiognomy too. tools like Electroencephalography
Physical appearance characteristics such (EEG), Positron Emission Tomography
as appearance of some facial features, of (PET), Functional Magnetic Resonance
the skull, shoulders, hands, fingers, legs, Imaging (fMRI) and structural MRI
including diffusion tensor imaging (DTI)

COPO307-4
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

to this study. One of the founders of this are often able to distort their responses.
area of brain research is Richard This is particularly problematic in
Davidson of the University of employment contexts and other contexts
Wisconsin-Madison [18]. Davidson’s where important decisions are being
research lab has focused on the role of made and there is an incentive to present
the prefrontal cortex and amygdala in oneself in a favorable manner. Social
manifesting human personality. In desirability is a tendency to portray self
particular, this research has looked at in a positive light, and faking bad also
hemispheric asymmetry of activity in happens, that is purposely saying ’no’ or
these regions. Neuropsychological looking bad if there’s a ’reward’ (e.g.
studies have illustrated how hemispheric attention, compensation, social welfare,
asymmetry can affect an individual’s etc.). Work in experimental settings
personality. [20,21] has shown that when student
In contemporary psychological research samples have been asked to deliberately
there should be an instrument which fake on a personality test, they
would provide a maximum amount and demonstrated that they are capable of
type of objective/unbiassed information doing this.
about personality in as short a time as Though several strategies have been
possible, preferably with no participation adopted for reducing respondent faking,
of person whose characteristics are this is still a problem for such traditional
identified. Comparison of approaches to psychological testing instruments like
identi- fication of psychological questionnaires, interviews, direct
characteristics described above is observations. Surprisingly,
represented in Table 3. neuropsychological tests are prone to
Some comparison of approaches to respondent faking, too [22,23]. Faking
identification of psychological response styles include faking bad
characteristics (malingering), faking good
Criterion Psy Inte Face, Neurops (defensiveness), attempts at invalidation,
chol rvie body ycholog
ogic w, evalua ical
mixed responding (faking good and
al dire tion tests bad), and a fluctuating, changing style
que ct that occurs within one evaluation
stio obs
nnai erva
session. These response styles lead to
res tion getting incorrect results.
Easy and not – – + – Concerning face and facial features,
time-consuming
for person who is
faking becomes much more complicated:
tested it’s impossible to change the shape of a
Person may not – – + – nose or cheekbones just when person
participate in
testing process
wants. Besides, it is often unknown to a
High validity and + – ? – holder what his/her face reveals exactly.
reliability [19] Theoretically people can “fake” facial
Practically no – – + – features intentionally changing their
possibility for
respondent faking shape, color, texture, for instance, using
No need in + + + – plastic surgery, and identifying personal
expensive hi–tech psychological characteristics becomes
hardware
much harder in this case, though it may
In psychological testing there is
be also accomplished.
considerable problem that respondents

COPO307-5
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
Face is the first subject that is unique for
people and used for people recognition.
Thus, face is the most available means of
evaluation among other instruments
based on questionnaires, interviews,
neuropsychological tests. People in
general may not participate in testing
process, identification of personality
characteristics may be done remotely,
even by exterior parties.
Summarizing, face provides researchers
and psychologists with instrument of
obtaining information about personality
and psychological traits that would be
much more objective than questionnaires
and neuropsychological tests (as we
can’s change facial features just when
such desire appears) and could be
obtained remotely using person’s facial
portrait, with no need for personal
involvement.
If such instrument is working
automatically (system gets facial
portrait, processes it and in result gives
out information about personality
characteristics) and has straight-forward
layout, then: 1) psychological testing
becomes more accurate, fast, objective
and available for different kinds of
research and applications; 2) deep
knowledge in interpretation of facial
features, which is rather rare in modern
society, isn’t needed to administer and
use the instrument. Methods and
algorithms originally developed for face
detection, face recognition and facial
expression recognition research fields as
well as contemporary trends (applying
standard face images, multimodality,
three-dimensionality) should be applied
and adjusted to so-called Automatic
Psychological Characteristics
Recognition from Face. From its side,
Automatic Recognition of Psychological
Characteristics from Face is believed to
bring scientific benefits to face
recognition, facial expression
recognition, face animation, face
retrieval, etc., and finally contribute to
development of human-computer
interaction on higher level. Thus, the
relations between such research areas as
face recognition, facial expression
recognition and psychological
characteristics recognition are mutually
beneficial.
3. Approaches to psychological
characteristics recognition from face
There are three main approaches to
psychological characteristics recognition
from face: physiognomy, phase facial
portrait and ophthalmogeometry, see
Fig.1. The first originally interprets
different facial features, the second
works with angles of facial features and
facial asymmetry, and the third extracts
and interprets eye region parameters.
Methods developed for these approaches
are described below. Physiognomy is a
theory based upon the idea that the
assessment of the person’s outer
appearance, primarily the face, facial
features, skin texture and quality, may
give insights into one’s character or
personality. Physiognomy has flourished
since the time of the Greeks
(Empedocles, Socrates, Hippocrates and
Aristotle), amongst the Chinese and
Indians, with the Romans (Polemon and
Adamantius), in the Arab world
(including Avicenna), and during the
European renaissance (Gerolamo
Cardano and Giovanni Battista della
Porta). It faded in
Figure 1. Approaches to psychological characteristics
recognition from facial portrait
COPO307-6
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
popularity during the 18th century, was
eclipsed by phrenology in the 19th and
has been refreshed by personologists in
the 20th century.
During 20th century attempts had been
made to perform scientific experiments
concerning validity of different facial
features interpretations and high
accuracy results had been claimed [24],
though they are mostly aren’t accepted
by official science [25]. At the same
time, science step by steps proves some
physiognomy beliefs. For instance,
correlations have been established
between IQ and cranial volume
[26,27,28,29]. Testosterone levels,
which are known to correlate with
aggressiveness, are also strongly
correlated with features such as finger-
length ratios and square jaws [30,31].
Interpretation of facial features
based on physiognomy has been
implemented into psychological
characteristics diagnosis tools such as
“Visage” Project [32] developed by Dr.
Paul Ekman and “Digital physiognomy“
software [33] developed by Uniphiz Lab.
“Visage” is a project for collecting
and organizing information about
relatively permanent facial features. It
includes methods for storing, retrieving,
and inspecting the data. Visage is a
unique database schema for representing
physiognomy and the interpretation of
physiognomic signs. The Visage
demonstration application illustrates
limited variations of some facial features
in the following categories: forehead and
eyebrows (see the Fig.2), eyes and
eyelids, nose, mouth and jaw, cheeks,
chin, ears. User should select features
that are distinctive about the face that is
going to be interpreted and then click the
“Get...” button. The application retrieves
information from the database relevant
to description of physiognomy,
including an estimation of the accuracy
of the sources of information.
“Digital physiognomy“ software
determines a person’s psychological
characteristics based on temperament
types, intellect, optimism – pessimism,
conformism – adventurism, egoism –
altruism, philanthropy – hostility,
laziness, honesty, etc.
Figure 2. Example of the table and interface of Visage
demonstration application: facial features in the
forehead and eyebrow area [34]
and then presents a detailed person’s
character analysis in a graphic format.
The tool works like a police sketch
(photo robot), so user has to select
different parts of the person’s face, and
doesn’t need to have a person’s
photograph, see Fig. 3. It’s claimed that
only the facial features that can be
interpreted with high accuracy were
used, and the confidence factor is
calculated for each interpretation by the
tool. It should be noted that “Digital
physiognomy“ tool also uses visual
systematic classification of 16
personality types based upon Myers-
Briggs typology, see Fig. 4.
“Visage” and “Digital Physiognomy”
projects are some of the first attempts
to develop physiognomic database and
use modern technology for
physiognomic interpretations. In spite of
having value for psychological diagnosis
based on physiognomy, both projects use
manual selection of facial features, and
thus, can’t be used extensively and
applied in scientific research.
COPO307-7
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Phase facial portrait approach to

psychological characteristics recognition
from facial portrait is primarily based on
calculating of angles of facial features
lines directions, as shown on Fig.5.
Video-computer psychological diagnosis
and correction method [35] had been
invented by professor Avtandil
Anuashvili. It’s remote method Figure 4. Socionics visual interpretation by “Digital
physiognomy“ based upon Myers-Briggs typology
for personal psycho-type identification,
method of one of the brain hemispheres
dominance identification, method of
psycho-diagnostics and psycho-
correction. The method is based on the Figure 5. The representation of phase facial portrait
concept [37]
thesis that face is an informational
these oscillatory processes among
background projecting in biological and
themselves [36]. Based on these
psychological terms the results of joint
meanings, the software classifies a given
brain hemispheres functioning. The
person to one of 49 psychological types
system of 49 types of people in
and gives out complete personal
coordinates “left-side thinker –
characteristic, professional characteristic
right-side thinker”, “psychologically
and recommendations on harmonization,
stable personality – instable personality”
effective interaction with other people
had been created and correlated to other
and environment.
personal typologies, see Fig. 6.
Video-computer psychological diagnosis
Software developed to apply video-
and correction method is one of the
computer psychological diagnosis and
primary instrumental psychological
correction method accepts facial image
methods concerning examination and
and determines: 1) dominance of one of
usage
the brain hemispheres on the basis of a
difference of amplitudes of oscillatory
processes occurring in right and left
hemispheres and 2) degrees of a
coordination of

Figure 6. Table of 49 psychological types system

invented used for video-computer psychological
diagnosis and correction; here L – logic (practical
mind), I – intuition (spirit, principles), S – stability, D
– destability [38]
Figure 3. Fragment of “Digital physiognomy“
software

COPO307-8
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
Figure 7. Translated picture from Muldashev’s book
[39]: here two parameters of facial eye region are used
for recognition of some basic psychological traits, e.g.
strong will and fearfulness, etc.
Figure 8. Ophthalmogeometrical pattern extraction 40]
of brain asymmetry phenomena and face
asymmetry. Although Anuashvili claims
that application developed for video-
computer psychological diagnosis and
correction method is entirely automated,
practically it may be considered to be
semi-automated as manual selection of
facial points on image is required. This
limits usage of such application for
extensive research and other purposes.
Concerning ophthalmogeometry
approach, it is based on idea that
person’s emotional, physical and
psychological states can be recognized
by 22 parameters of an eyes part of the
face [39], see Fig. 7. phthalmogeometry
phenomenon has been discovered by
prof. Ernst Muldashev. Apart from other
interesting facts, E. Muldashev has
found that in 4-5 years after birth the
only practically constant parameter of
human body is the diameter of the
transparent part of cornea which equals
10±0,56 mm. He also represented an
idea that ophthalmogeometrical pattern
is unique for people. The procedure of
this pattern identification and calculation
is described by Leonid Kompanets [40],
see Fig. 8.
Ophthalmogeometry is based on
interesting ideas and may be applied to
psychological, medical research as well
as to biometrics, though this is not very
deeply investigated area of facial
analysis which primarily needs
automation of ophthalmogeometric
pattern extraction and further
investigation.
4. Conclusion
The paper represents general idea that
face provides researchers and
psychologists with objective instrument
of obtaining information about
personality and psychological traits. An
up-to-date survey of approaches and
methods in psychological characteristics
recognition from facial image is
provided.
In perspective new research task of
automating procedures in applications of
psychological characteristics recognition
from face should be explored. Various
approaches and methods developed
within face recognition, facial
expression recognition, face retrieval,
face modeling and animation may be
applied and adjusted for recognition of
psychological characteristics from face.
Undeniably, such automated system of
psychological characteristics recognition
from face will get countless
psychological, educational, business
applications. It may be used also as part
of medical systems: 1) patient’s
psychological state and traits influence
the process of medical treatment, and it
should be taken into consideration and
researched; 2) patient’s psychological
characteristics should be taken into
account to reflect and construct the
psychosomatic model of disease in the
environment, which includes biological,
psychological, and social factors.
References
[1] Carver C. S., Scheier M. F. Perspectives on
personality (4th ed.) Boston: Allyn and
Bacon, 2000, page 5.
[2] DSM, Diagnostic and Statistical Manual of Mental
Disorders, can be found at
http://www.psych.org/ research/ .
[3] Hampson S. E. Advances in Personality
Psychology. Psychology Press, 2000.
[4] Holigrocki R. J., Kaminski P. L., Frieswyk S. H.
(2002). PCIA-II: Parent-Child Interaction
COPO307-9
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
Assessment Version II. Unpublished manuscript,
University of Indianapolis.
(Update of PCIA Tech. Rep. No. 99-1046. Topeka,
KS: Child and Family Center,
The Menninger Clinic.) (Available from Dr. Richard J.
Holigrocki or Dr. Patricia
L. Kaminski).
[5] Nigel Barber. The evolutionary psychology of
physical attractiveness: Sexual selection
and human morphology. Ethology and Sociobiology,
Volume 16, Issue 5, September
1995, pages 395-424.
[6] John P. Swaddle, Innes C. Cuthill. Asymmetry and
Human Facial Attractiveness:
Symmetry May not Always be Beautiful. Proceedings:
Biological Sciences, Vol. 261,
No. 1360 (Jul. 22, 1995), pages 111-116.
[7] Thomas R. Alley, Michael R. Cunningham.
Averaged faces are attractive, but very
attractive faces are not average. Psychological Science
2 (2), 1991, pages 123-125.
[8] Leslie A. Zebrowitz, Gillian Rhodes. Sensitivity to
”Bad Genes” and the Anomalous
Face Overgeneralization Effect: Cue Validity, Cue
Utilization, and Accuracy in Jud
72 Ekaterina Kamenskaya, Georgy Kukharev
ging Intelligence and Health. Journal of Nonverbal
Behavior Volume 28, Number 3
/ September, 2004, pages 167-185.
[9] Caroline F. Keating. Gender and the Physiognomy
of Dominance and Attractiveness,
Social Psychology Quarterly, Vol. 48, No. 1 (Mar.,
1985), pages 61-70.
[10] Ulrich Mueller, Allan Mazur. Facial Dominance
of West Point Cadets as a Predictor
of Later Military Rank. Social Forces, Vol. 74, No. 3
(Mar., 1996), pages 823-850.
[11] J. Liggett, The human face. New York: Stein and
Day, 1974, page 276.
[12] Phisiognomics, attributed to Aristotle. Cited in
J.Wechsler (1982), A human comedy:
Physiognomy and caricature in 19th century Paris
(p.15). Chicago: University
of Chicago Press.
[13] A. Brandt. Face reading: The persistence of
physiognomy. Journal Psychology Today,
1980, December, page 93.
[14] Sibylle Erle. Face to Face with Johann Caspar
Lavater, Literature Compass 2 (2005)
RO 131, pages 1 -4.
[15] Stefan Boehringer, Tobias Vollmar, Christiane
Tasse, Rolf P Wurtz, Gabriele
Gillessen-Kaesbach, Bernhard Horsthemke and
Dagmar Wieczorek. Syndrome identification
based on 2D analysis software. European Journal of
Human Genetics
(2006), pages 1-8.
[16] Hartmut S Loos, Dagmar Wieczorek, Rolf
PWürtz, Christoph von der Malsburg and
Bernhard Horsthemke. Computer-based recognition of
dysmorphic faces, European
Journal of Human Genetics (2003) 11, pages 555-560.
[17] Irvin L. Child, William H. Sheldon. The
correlation between components of physique
and scores on certain psychological tests. Journal of
Personality, Vol. 10, Issue 1,
September 1941, page 23.
[18] Richard Davidson, Ph.D. Vilas Professor of
Psychology and Psychiatry. Can be
found at https:// psychiatry.wisc.edu/
faculty/FacultyPages/Davidson.htm.
[19] The Validity of Graphology in Personnel
Assessment. Psychological Testing Centre.
Found at www.psychtesting.org.uk, November 1993
reviewed April 2002.
[20] Chockalingam Viswesvaran, Deniz S. Ones.
Meta-Analyses of Fakability Estimates:
Implications for Personality Measurement,
Educational and Psychological Measurement,
Vol. 59, No. 2, 1999, pages 197-210.
[21] Deniz S. Ones, Chockalingam Viswesvaran,
Angelika D. Reiss. Role of Social Desirability
in Personality Testing for Personnel Selection: The
Red Herring. Journal
of Applied Psychology, 1996. Vol. 81, No. 6, pages
660-679.
[22] Hall, Harold V.; Poirier, Joseph G.; Thompson,
Jane S. Detecting deception in
neuropsychological cases: toward an applied model.
From: The Forensic Examiner,
9/22/2007.
[23] Allyson G. Harrison, Melanie J. Edwards and
Kevin C.H. Parker. Identifying students
faking ADHD: Preliminary findings and strategies for
detection. Archives of
Clinical Neuropsychology. Volume 22, Issue 5, June
2007, pages 577-588
[24] Naomi Tickle. You Can Read a Face Like a
Book: How Reading Faces Helps You
Succeed in Business and Relationships, Daniels
Publishing, 2003.
[25] Robert Todd Carroll. The Skeptic’s Dictionary: A
Collection of Strange Beliefs,
Amusing Deceptions, and Dangerous
Delusions.Wiley; 1st edition (August 15, 2003)
[26] J. Philippe Rushton, C. Davison Ankney. Brain
size and cognitive ability: Correlations
with age, sex, social class, and race. Psychonomic
Bulletin & Review, 1996, 3
(1), pages 21-36.
[27] Michael A. McDaniel. Big-brained people are
smarter: A meta-analysis of the re
Recognition of Psychological Characteristics from
Face 73
lationship between in vivo brain volume and
intelligence. Intelligence, Volume 33,
Issue 4, July-August 2005, pages 337-346.
[28] J. Philippe Rushton. Cranial size and IQ in Asian
Americans from birth to age
seven. Intelligence, Volume 25, Issue 1, 1997, pages
7-20.
[29] John C. Wickett, Philip A. Vernon, Donald H.
Lee. Relationships between factors of
COPO307-10
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

intelligence and brain volume. Personality and

Individual Differences, Volume 29,
Issue 6, December 2000, pages 1095-1122.
[30] John T. Manning. Digit Ratio: A Pointer to
Fertility, Behavior, and Health. Rutgers
University Press, 2002.
[31] Bernhard Fink, Karl Grammer, Philipp
Mitteroecker, Philipp Gunz, Katrin Schaefer,
Fred L. Bookstein, John T. Manning. Second to fourth
digit ratio and face
shape. Proceedings of the royal society, Volume 272,
Number 1576 / October 07,
2005, pages 1995-2001.
[32] Visage Project for Physiognomy Data, can be
found at
http:// face-and-emotion.com/ dataface/ visage/
about\_visage.jsp. Date: 20
January 2008.
[33] Digital Physiognomy Software: Match person’s
face to his or her character. Can be
found at http://www.uniphiz.com/ physiognomy.htm.
Date: 20 January 2008.
[34] Facial Features in the Forehead & Eyebrows
Area, can be found at
http:// face-and-emotion.com/ dataface/ visage/
visage\_forehead.jsp. Date: 20 January
2008.
[35] Anuashvili Avtandil. Fundamentals of
Psychology. Scientific, Philosophic and Spiritual
Fundamentals of Psychology. The Institute for Control
Problems Press, Moscow,
2001 (In Russian).
[36] Video-computer system for psychodiagnosis and
psychocorrection is the
first objective tool for psychotherapist and
psychologist, can be found at
http://www.anuashvili.ru/Video-
computer\%20psychodiagnostics/ video-compute.
htm. Date: 20 January 2008.
[37] The audio-presentation for video-computer
psychological diagnosis and correction,
can be found at http://www.tsarvgolove.ru (in
Russian). Date: 20 January 2008.
[38] Avtandil Anuashvili’s website, can be found at
http://www.anuashvili.ru/ . Date:
20 January 2008.
[39] Muldashev Ernst R. Whom did we descend
from?, “OLMA-PRESS”, Moscow 2002
(In Russian)
[40] Leonid Kompanets. Biometrics of asymmetrical
face. From Biometric Authentication:
First International Conference, By David Zhang, Anil
K. Jain, ICBA 2004 ,
pages 67-73.

COPO307-11
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
An Automatic Machine Learning and Particle Filtering
Based Approach to Real Time Human Tracking in Videos
Chandra Mani Sharma1, Alok Kumar Singh Kushwaha1, Ashish Khare2 and Sanjay Tanwani1
1
School of Computer Science and IT, DAVV Indore. 2 Department of Electronics and Communication, University of Allahabad, Allahabad.
chandramani.sharma@yahoo.com, alokkumarsingh.jk@gmail.com, akhare@allduniv.ac.in,
sanjay_tanwani@hotmail.com
Abstract:
Object Tracking is an important task in video
processing because of its various applications like
visual surveillance, human activity monitoring and
recognition, traffic flow management etc. Multiple
object detection and tracking in outdoor environment
is a challenging task because of the problems raised
by poor lighting conditions, occlusion and clutter.
This paper proposes a noble technique for detecting
and tracking the multiple humans in a video. A
classifier is trained for object detection using haar-
like features from the training image set. The human
objects are detected with the help of this trained
detector and are tracked with the help of a particle
filter. The experimental results show that the propose
technique can detect and track the multiple humans in
a video adequately fast in the presence poor lighting
conditions, clutter and partial occlusion and the
technique can handle varying number of human
objects in the video at various points of time.
Keywords: Human detection, Automatic multiple
object tracking, Haar-like features, Machine learning,
Particle filter.
1. Introduction:
Detecting humans and analyzing their activities by
vision is key for a machine to interact intelligently
and effortlessly with a human inhabited
environment .The aim of visual surveillance is the
real-time observation of targets such as human
beings or vehicles in some environment which leads
to the description of objects’ activities with the
environment or within them. Visual surveillance has
been used for security monitoring, anomaly
detection, intruder detection, traffic flow measuring,
accident detection on highways, and routine
maintenance in nuclear facilities[1,2,3,4]. Hu et.
al.[1], provide a good survey on visual surveillance
and its various applications. Detecting and Tracking
humans in a video is a step to take in the process of
analyzing and predicting their behavior and
intention. This is a good practice to detect objects in
a video sequence before tracking them. The problem
of multiple object tracking is more complex and
challenging than the single object tracking because
of the issue of the management of multiple tracks
caused by newly appearing objects and the
disappearance of already existing targets. Viola et
al[5] proposed an object detection framework for
and used it for first time for detecting the human
faces in an image. The adaptive boosting can be
used to speed up a binary classifier [6] and this can
be used in machine learning for creating the real-
time detectors. After detecting an object or multiple
objects in a video sequence captured by the
surveillance camera, the very next step is to track
these objects (human, vehicle etc.) in the subsequent
frames of the video stream. The particle filter based
tracking techniques are gaining popularity because
of their ease of implementation and capability to
represent a non-linear object tracking system and
non-Gaussian nature of noise. Various object
tracking techniques based on particle filtering are
found in literature [7,8,9,10,11]. These approaches
fall under two main categories; single object
trackers and multiple object trackers. Lanvin et.
al.[7], propose an object detection and tracking
technique and solve the non-linear state equations
using particle filtering. Single object trackers suffer
from the problem of false positives when severe
COPO4O1-1
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
occlusions occur because of hidden first order
Markov hypotheses [8]. The problem of tracking
multiple objects using particle filters can be solved
in two ways. One is by creating multiple particle
filters for each track and another one is by having a
single particle filter for all tracks. The second
approach works fine as long as the objects under
tracking are not occluded but in case of occlusion
when objects come close by, these techniques fail to
track the objects. In [8] a multi-object tracking
technique using multiple particles has been
proposed. Algorithms which attempt to find the
target of interest without using segmentation have
been proposed for single target tracking based on
cues such as color, edges and textures [12].
Chen et. al.[9] propose a color based particle filter
for object tracking. Their technique is based on
Markov Chain Monte Carlo(MCMC) particle filter
and object color distribution. Sequential Monte Carlo
techniques also known as particle filtering and
condensation algorithms and their applications in the
specific context of visual tracking, have been
described in length in the literature[13,14,15]. In
[10], object tracking and classifications are
performed simultaneously. Many of the particle filter
based multiple object tracking schemes rely on
hybrid sequential state estimation.
The particle filter developed in[16] has multiple
models for the objects motion, and comprises an
additional discrete state component to denote which
of the motion models is active. The Bayesian
multiple-blob tracker[17] presents a multiple tracking
system based on statistical appearance model. The
multiple blob tracking is managed by incorporating
the number of objects present in the state vector and
the state vector is augmented as in [18] when a new
object enters the scene. The joint task of object
detection and tracking comes with a heavy
computational overhead but for visual surveillance
applications the speed is one of the most important
factors. Many previously proposed techniques are
slow in speed so cannot be good for visual
surveillance and some fail to cope with dynamic
outdoor environment conditions. This paper presents
a noble technique based on machine learning and
particle filtering to detect and track the humans in a
video. The humans are detected in video frames using
the object detection framework proposed by Viola
et.al.[5] using Haar-like features and then tracked
using a simple particle filter in the subsequent video
frames. Binary adaptive boosting [6] reduces the
training time of the human detector. The earlier
detection of the objects simplifies the process of
tracking and sheds the load of detection from tracker.
The exhaustive dataset used for training the detector
makes the system detect and track the multiple
objects in critical lighting conditions, in dynamic
background, partial occlusion and clutter.
The rest of the paper is organized as follows;
section(2) discusses the proposed technique for
human detection and tracking. In section (3) the
various experimental results using the proposed
technique are given which prove the validity and
novelty of the method. Section(4) comprises of the
conclusions and at last references are given.
2. Methodology
Here we solve two sub-problems. One is to detect
the humans in the video and another one is to track
them in the subsequent video frames. The sub
problem of object detection is solved using
machine learning approach for which we train our
human detector using binary adaptive boosting.
Algorithm for multi human detection and
tracking :-
 Let Z be the input video to the algorithm.
 In first frame  0
of Z, detect humans using
the human detector. Let  be the number
of detected humans.
 Initialize trajectories  j
, 1≤ j≤  with
initial positions  j,0
.
 Initialize the appearance model (color
histogram)  j
for each trajectory from
the region around  j,0
.

 For each subsequent frame of input
i
video,
(a) For each existing trajectory  j
,
COPO4O1-2
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

i. Use motion model to predict the
distribution  (
over locations for human j in
frame i, creating a set of
candidate particles
≤ K.
ii. Compute the color histogram
speed up the process. We collected positive and
 │
j ,i
),
j , i 1
negative image samples for training. The
positive images are those comprising the human
beings and negative images do not contain any
human being. We cropped the humans from the

(k )
j ,i
, 1 ≤k positive images and resized to the dimensions
of 40*40. Our positive dataset consisted of
2,000 images while the negative dataset


(k )
and likelihood consisted of 2,700 images. Fig.1 consists of
some cropped human images from positive
 ( │  , )
k (k )
for each samples used for human detector training. This
j ,i j human detector is nothing but a binary classifier
particle k using the appearance having two classes: human and non-human.
model. Next step after the sample collection is Haar-
iii. Resample the particles according to
their likelihood. Let k* be the like feature extraction from these samples. We
index of the most likely use here the rectangular Haar-like features and
particle. these features have their intuitive similarity to
iv. Perform confirmation by the Haar-wavelets. Integral Image
classification: run the Humans representation is used for fast feature evaluation
detector on the location (see fig. 3).


*
( k )
. If the location is
j ,i
classified as a human, reset
 j
 0; else increase

 j
  1.
j

v. If  j
is greater than threshold,
remove trajectory j.

(b) Search for new humans in frame

 i
and compute the Euclidean

distance  j ,k
between each
newly detected human k and each

Fig. 1: Sample positive human images used for detector
existing trajectory j
. When training.

 j ,k
 for all j, where  is a
threshold (in pixels) less than the
width of the tracking window,
initialize a new trajectory for
detection k.
2.1 Sample Creation and Human Detector
Training

We use a machine learning approach for object

detection based on Haar-like features, originally
proposed by Viola et. al.[5]. The training is
Fig2: Rectangle Haar-like features
performed using binary adaptive boosting to

COPO4O1-3
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

2.2.1 Predict: We predic  (  │ j ,i

) ,a
j , i 1

distribution over human object j's position in

frame i given our belief in its position in frame i-
1.

2.2.2 Measure: For each propagated particle

k, we measure the likelihood
 ( │ j , i , j ) using
k (k )
Fig3: Integral Image for fast feature evaluation a color histogram-

The Integral Image, ii ( x, y ) at location based appearance model. After computing the
likelihood of each particle we treat the
( x, y ) contains the sum of all pixels above and left likelihoods as weights, normalizing them to sum
of ( x, y ) and can be computed in single pass over to 1.
image using the following pair of equations.
2.2.3 Resample:- We resample the particles to
 ( x, y)  ( x, y 1)  i( x, y) (2.1.1) avoid degenerate weights. Without re-sampling,
over time, the highest-weight particle would tend
to a weight of one and the other weights would
ii( x, y)  ii( x  1, y)  ( x, y) (2.1.2) tend to zero. Re-sampling removes many of the
Where,  ( x, y) is the cumulative row sum and low weight particles and replicates the higher-
i ( x, y ) is the original image. We use adaptive weight particles. We thus obtain a new set of
equally-weighted particles. We use the re-
boosting technique[6] for training the human detector sampling technique described in [13].
system. The parameters and their values used in the
training set up are given in the table 1. 2.3 Motion and Appearance Models:-
Table 1: Human detector training parameters and their values.
Our motion model is based on a second-order auto-
regressive dynamical model. The autoregressive
Parameters Values Description
Npos 2,000 Number of positive images
consisting only human beings
model assumes the next state  t
of a system is a
Nneg 2,700 Number of negative images not function of some number of previous states and a

Nstages 20
consisting any human
Number of training stages
noise random variable  t
Minhitrate 0.991 Per stage minimum hit-rate

  f ( t 1,  t  2,.........,  t  p , i )
(99.10%)
maxfalsealarm 0.5 Maximum false alarm rate per t
(2.3.1)
stage (50%)
Mode All Use upright and tilted features
Width*height 40*40 Size of training images
We assume the simple second-order linear
Boosttypes DAB Discrete Adaptive Boosting autoregressive model

  2  j ,i 2  i (2.3.2)
j ,i j ,i 1

2.2 Particle Filter for Tracking:-

For tracking we use a particle filter. We use an Our appearance model is based on color histograms.
approach in which the uncertainty about an human's
state (position) is represented as a set of weighted
We compute a color histogram  j
in HSV space

particles, each particle representing one possible for each newly detected human and save it to
state. The filter propagates particles from frame i-1 to compute particle likelihoods in future frames. To
frame i using a motion model, computes a weight for compute a particle's likelihood we use the
each propagated particle using an appearance model, Bhattacharyya similarity coefficient between model

 
then re-samples the particles according to their (k )
histogram and observed histogram as
weights. The initial distribution for the filter is j
centered on the location of the object the first time it follows, assuming n bins in each histogram:
is detected. Here are the steps in more detail:

COPO4O1-4
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

 ( │ j , i , j )  e  
(k )
k (k ) d ( , )
j (2.3.4)

And ,

d ( ,  
n
) 1  
(k ) (k )
(2.3.5)
j j ,b b
b 1

 and  
(k )
and 
(k )
and b
denote bin b of ,
j ,b j Frame 100 Frame 125
respectively. A more sophisticated appearance model
based on local histograms along with other
information such as spatial or structural information
would most likely improve our tracking performance,
but we currently use a global histogram computed
over the entire detection window because of its
simplicity.

3. Experimental Results:-
Frame 150 Frame 175
We have experimented the proposed automatic
human detection and tracking technique on a number
of videos. The results with some of the representative
videos are given here. The human detection and
tracking starts automatically without providing any of
the initialization details unlike many other tracking
techniques in which operator intervention is required.
Our proposed detector detects the human beings
irrespective of their body poses and locations in
video (fig. 4). The results given in fig. 5 show that
the proposed technique performs fairly good in
outdoor environment in low lighting conditions and Frame 200 Frame 225
is quite suitable for visual surveillance applications.

Frame 1 Frame 25 Frame 250 Frame 275

Frame 300 Frame 325

Frame 50 Frame 75

COPO4O1-5
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
Frame 350 Frame 375
Frame 400 Frame 425
Fig. 4: Detection and tracking results with the proposed method on
a real-time video in outdoor environment. The video was shot in
evening in dim lighting conditions.
References
[1] Weiming Hu and Tieniu Tan , “A Survey on Visual
Surveillance of Object Motion and Behaviors ”, IEEE Trans.
Systems, Man, and Cybernetics, Vol. 34, No. 3, pp. 334-352,2006.
[ 2] E. Stringa and C.S. Regazzoni, ”Real-time Video Shot
Detection for Scene Surveillance Application,” IEEE Trans. Image
Processing, Vol. 9, pp. 69-79, 2009.
[3] I. Haritaoglu, D. Harwood, and L. Davis, “W4: Who? When?
Where? What? a Real-time System for Detecting and Tracking
People,” in Proc. IEEE Conference on Face Gesture Recognition,
pp. 222–227, 1998.
[4] D. Koller, J. Weber, T. Huang, J. Malik, B. Rao G. Ogasawara,
and S. Russell, “Toward Robust Automatic Traffic Scene Analysis
in Real-time,” In Proceedings of Int. Conference on Pattern
Recognition, Vol. 1, pp. 126–131, 1994,.
[5] Paul Viola and Michael Jones, “Rapid Object Detection using a
Boosted Cascade of Simple Features,” In the Proceedings of IEEE
International Conference on Computer vision and Pattern
Recognition, 2001.
[6] Freund, Y. and Schapire, R.” A decision theoretic
generalization of on-line learning and an application to boosting,”
Journal of Computer and System Sciences 55 pp.119–139, 2007
[7] P. Lanvin, J-C. Noyer, M. Benjelloun,” Object Detection and
Tracking using the Particle Filtering,” In Proc. of IEEE 37th
Asilomar Conference on Signals, Systems and Computers, Vol.
2,pp.1595-1599, 2003.
[8] Jingling Wang, Yan Ma, Chuanzhen Li, Hui Wang, Jianbo Liu,
“An Efficient Multi-Object Tracking Method using Multiple
Particle Filters” In the Proceedings of World Congress on
Computer Science and Information Engineering, pp. 568-572,
2009.
[9] Yuan Chen, Shengsheng Yu, Jun Fan, Wenxin Chen,
Hongxing Li, “An Improved Color-based Particle Filter for Object
Tracking,” In Proceedings of IEEE Second International
Conference on Genetic and Evolutionary Computing, pp. 360-363,
2008.
[10] Francois Bardet, Thierry Chateau, Datta Ramadasan, “
Illumination Aware MCMC Particle Filter for Long-term Outdoor
Multi-Object Simultaneous Tracking and Classification,” In
Proceedings of IEEE 12th International Conference on Computer
Vision, pp. 1623-1630, 2009.
[11] Cheng Luo, Xiongcai, Jian Zhang, “ Robust Object Tracking
using the Particle Filtering and Level Set Methods: A Comparative
Experiment,” In the Proceedings of IEEE 10th Workshop on
Multimedia Signal Processing, pp. 359-364, 2008.
[12] P. Brasnett, L. Mihaylova, N. Canagarajah, and D. Bull,
“Particle Filtering for Multiple Cues for Object Tracking in Video
Sequences,” in Proc. of the 17th SPIE Annual Symposium on
Electronic Imaging, Science and Technology, vol. 5685, pp. 430–
440,2005
[13] M. Isard and A. Blake, “Contour Tracking by Stochastic
Propagation of Conditional Density,” in European Conf. on Comp.
Vis. , pp. 343–356, Cambridge, UK,1996
[14] M. Isard and A. Blake, “Condensation–Conditional Density
Propagation for Visual Tracking,” Intl. Journal of Computer
Vision, vol. 28, no. 1, pp. 5–28, 1998.
[15] P. P´ erez, J. Vermaak, and A. Blake, “Data Fusion for
Tracking with Particles,” Proceedings of the IEEE, vol.92, no. 3,
pp. 495–513, March 2004.
[16] M. Isard and A. Blake, “A Mixed-State Condensation
Tracker with Autmatic Model-Switching,” in Proc. Int. Conf.
Computer Vision, pp. 107–112,1998.
[17] M. Isard and J. MacCormick, “Bramble: a Bayesian Multiple
Blob Tracker,” in Proc. Int. Conf. Computer Vision, pp. 34–
41,2001
[18] J. Czyz, B. Ristic, and B. Macq, “A Color-Based Particle
Filter for Joint Detection and Tracking of Multiple Objects,” in
Proc. of the ICASSP, 2005
COPO4O1-6
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

An Efficient Image Segmentation Method Based on Normalized Cut.

Vaibhav Sharma, Asstt. Professor RKGIT , Member, ICEIT

Abstract

An efficient method for segmentation of images has been proposed by incorporating the advantages of the
normalized cut partitioning method.The proposed method pre-processes an image by using the normalized cut
algorithm to form segmented regions and applied to form the weight matrix W. Since the number of the
segmented region nodes is much smaller than that of the image pixels. The proposed algorithm allows a varied
dimensional images with significant reduction of the computational complexity comparing to conventional
Ncut method based on direct image pixels. The experimental results also verify that the proposed algorithm
behaves an improved performance comparing to Ncut algorithm. This paper presents different ways to
approach image segmentation, explain an efficient implementation for each approach, and show sample
segmentations results. We show that an efficient computational technique based on a generalized eigen value
problem can be used to optimize this criterion. We have applied this approach to segmenting static images and
found the results to be very encouraging.

Index Terms---Normalized cut (Ncut);Grouping; image segmentation; graph partitioning.

I. Introduction algorithm for implementation method. For instance, a

The information we know directly about an image is
the intensity/color and location of each pixel, and
therefore these two variables are the central source of
data for any segmentation. Edges that look clearly
defined from far away are often blurry and hard to
distinguish at the pixel level. Similarly, an area that
looks like a solid color at a distance may be seen as
mix of many colors at the pixel level and may be
interpreted as a number of segments. The importance
and difficulties of perceptual grouping have been very
well studied in vision research and image processing
literature. How to partition image to subsets, when
there are many possible ways of doing it? The
intuitive answer would be a partitioning based on the
available prior knowledge. The difficulty is specifying
the knowledge in a compact and useful way i.e.
defining the descriptors of an image. Over the
decades, different scientific communities proposed
various approaches to generalized image segmentation
problem. Tree-based hierarchical clustering structures,
region-based merge and split image processing
algorithms and statistical approach to image
segmentation introduced additional constraints on a
possible solution. Before any algorithm
implementation we have to determine what criterion
we want to optimize and if there is an efficient
general image segmentation approach based on
normalized cut (Ncut) by solving an eigensystem,
which was proposed by Shi and Malik [4] .In recent
years, Ncut algorithm is widely used for image
processing and related fields, such as the motion
picture , medical imaging, vector field of the
segmentation. However, this approach is NP-hard
problem, the more pixels in the image, the larger the
generated graph of nodes is, which bring difficulties
to solve the algorithm.
Considering the above issues come from the two
algorithms, this paper designed a combination of both
methods, the specific algorithm can be find in the
following part.
II. Normalized Cuts
The normalized cut method was introduced by Shi
and Malik in their paper Normalized Cuts and Image
Segmentation [4]. The paper is the one of the most-
cited papers in computer vision according to Google
Scholar, which indicates that the method is used
frequently in practice.
COPO4O2-1
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
The normalized cuts algorithm defines the optimal
segmentation in terms of a segmentation that divides
the image into two parts. To achieve a full
segmentation, it finds an optimal binary segmentation
and then recursively finds an optimal segmentation
within each segment. The way the normalized cuts
algorithm defines an optimal segmentation is quite
intuitive. Given some measure of similarity between
pairs of pixels, we define the optimal segmentation to
be the one that maximizes the similarity between pixels
that belong to the same segment, and minimizes the
similarity between pixels that belong to different
segments. Formally, we can define the optimization
problem as follows:
Let G = (V,E) be a graph, where V is the set of pixels
P, and E is the set of all edges {p, q} between every
pair of pixels p and q. We set the weight of an edge,
denoted w{p,q}, equal to a similarity measure between
the pixels p and q, where a higher weight corresponds
to pixels that are more similar.
The similarity measure used by Shi and Malik [4] is the
following:
(1)
where X(i) is the spatial location of i and F(i) is the
vector based on intensity, color, or texture information
at i. This definition defines the similarity between two
pixels to be proportional to the probability of pixel i
being generated by a Gaussian distribution centered at
pixel j. Notice that with this definition we do not need a
fully connected graph, but instead only place edges
between two pixels that have a positive weight, i.e.
when ||X(i) − X(j)||2 < r2.To measure the similarity
between segments we use a normalized cut. Given a
graph as specified above, and sets A and B, a
normalized cut, denoted Ncut, is defined to be:
(2)
(3)
The associativity and cut both measure how similar
two sets are to each other. The only difference is that
a cut requires that the two sets are a partition of V ,
while associativity generalizes this to measure the
similarity between any two sets. A normalized cut
measures how similar A is to B compared to how
similar A is to the whole image and B is to the whole
image. Thus, when the value of the normalized cut is
at a minimum, we are finding the segmentation
where the difference between A and B is maximized.
It may seem a bit odd that we do not consider the cut
alone but rather the ratio of cut over associativity.
However, this is necessary to avoid a bias resulting
from a different amount of edges in the cut
depending on the relative size of the segments.
Unlike in graph cuts, where at least one terminal-
link from every node must be in the cut, here if a
segment is just one pixel, the value of the cut is just
the sum of at most |P| − 1 edges out of the pixel.
This sums to significantly fewer edges than a cut
that divides the image in half and cuts at most (P/2)
edges. If we consider just the cut alone then we will
falsely encourage one of the two parts to contain
very few pixels.
To measure the similarity between pixels within the
same segment we formalize something similar to the
normalized cut called the normalized associativity.
The normalized associativity between sets A and B,
denoted Nassoc(A,B), is the following:
COPO4O2-2
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
The eigenvector associated with the smallest

(D W ) y
eigenvalue will minimize  and is
Dy
equivalent to Ncut(A,B). But, if we assume it for
now, we can observe a number of properties of λ and
(4) the associated eigenvectors.

First, Ncut cannot be negative because all the

weights in the graph are positive, so

(D W ) y
Precisely, the optimization problem is equivalent to  = 0 . Also, we always have the
finding the eigenvector with the second smallest Dy
eigenvalue of the following generalized eigenvalue eigenvalue = 0 associated with the eigenvectors
       
equation: y 1 and y 0 because D1 W 1 and D0 W 0 .
But, in both of these cases, the vector only has
one value and all the pixels are assigned to the same
segment. The normalized cut of this segmentation is
undefined because we need non-empty sets A and B
so that we do not divide by zero in the normalized
(5) cut. Therefore, the smallest eigenvector is λ = 0, and
we have that the second smallest
where D is a |V |×|V | diagonal matrix with di,i eigenvalue is the one that minimizes the normalized
= W is a |V| X |V| matrix with wi,j = w{i,j}, cut.
W is a |V| X |V| matrix with wi,j =w(i,j), 
(D W ) y
Now, to see that  is the same as the value
   i A d i ,i Dy
y x b(1 x ) and b
i B d i ,i of Ncut(A,B) when (and ) are discrete we
  notice that:
The vector y is a linear combination of x and
 
therefore y will define the same segmentation as x .
Below we will give an explanation, but not a
complete proof, as to why this is the correct equation
to solve. The explanation expands and clarifies the
complete proof found in the paper by Shi and Malik
[4]. The paper also provides evidence that the Because can only have the values 1 and 0, we
eigenvalue problem is O(n(3/2)) where n is the have that = −b(1− ) can also only take on
number of pixels in the image, when using a sparse only two values. If xi =1( i A ) then yi = 1. If xi = 0
matrix where most of the edge weights are 0. 
( i B )then yi = −b. Also note that y TW multiplies
the ith column of W by yi. Because column i contains
To give some intuition for why the eigenvector the weights of all the edges going into a vertex i, we
associated with the second smallest eigenvalue of 
have that y TW multiplies problem by throwing out
minimizes Ncut(A,B), consider a such vectors and not dividing the segment further. It
rewriting the equation as follows: would be problematic if the first partition resulted in
such a vector because then we would not be able to
find any segmentation at all. The paper does not
address this issue aside from briefly presenting a
non-recursive solution for which they show no
results.
(6)
In summary, the normalized cut algorithm can be
written as follows:
COPO4O2-3
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
• Decide on a similarity measure and compute matrices memory is wasted. Each vector is converted into a
D and W. square 2D texture by the application program.
• Solve the generalized eigenvalue equation from Vectors are padded with zero entries if they do not
Equation (5) for the eigenvector entirely fill the texture.

associated with the second smallest eigenvector.
 Preparing the data
• Find a splitting point of such that the normalized
cut is minimized and use it to partition the graph. We represent a image of size nXm as a column
vector V R N where N =n˟ m
Recursively partition each segment until some We compute the weight matrix W R N N where W(i,j)
sufficient measure of similarity is reached inside the represent relationship between pixel i and j in V.
segment We can use Gaussian function that take into
III. Efficient Image Segementation Algorithm consideration of distance, intensity different, crossing
The multiplication of matrix-matrix and matrix- edge information. Since each pixel is connected to its
vectors is one of the most central operations applied limited number of neighbors, this matrix is very sparse.
in linear algebra. Parallelization of numerical solvers (,)Wij Now given a weight matrix W R N N and k
would greatly accelerate a variety of algorithms that classes. We compute the degree matrix , where and
have a primary bottleneck of waiting for these each element is the sum of the corresponding rows in
calculations and on the availability of computational W The proposed method consists of the following
resources. Speeding up matrix operations by itself Steps :-
presents a challenging problem. Many techniques 1 Find the optimal eigensolution Z* `by solving the
are theoretical or run across accuracy limitations on leading k eigenvectors V[ k ] [v1, v2 , , vk ] for
the GPU. Graphics hardware is specialized in a 1 1 1
manner that makes it easier for these types of D 2WD 2 v v.Z * D 2V[ k ]
problems than using a standard CPU. For the 1
purpose of this proposal, ~
2 Normalize Z* by X * Diag(diag 2 ( Z *Z *T ))Z *
The representation of a 2D matrix as a set of
diagonal vectors, and finally as a set of 2D textures 3 Initialize orthonormal matrix R* R N k for finding
is shown in the optimal discrete solution:
~ ~
R1* [ X * (i,1), , X * (i, k )]T where random i [N ]
C 0 N 1 , a , a zero vector
For j=2 to k do {
~
C C abs ( X * R*j 1 )
~
R*j [ X * (i,1), , X * (i, k )]T , where i=argmin C[i]
{
4 Initialize convergence monitoring parameter * =0.
5 Find the optimal discrete solution X * {0,1}N k
~ ~
X X * R*
1
X * (i, l ) if l=argmax
0

figure1. 6 Find the optimal orthonormal matrix R* RN k

Kuger and Westermann represent matrices as set of

diagonal vectors and we store vectors in 2D texture ~ ~
SVD( X *T X * ) U U T
maps (see Figure above). To every diagonal starting
at the i-th entry of the firrst row in the upper right =trace (∑)
*
part of the matrix, its counterpart starting at the (N- If | | machine precison,then stop and output
i)-th entry of the first column in the lower left part of X *

the matrix is appended. In this way, no texture

COPO4O2-4
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
*
= directly applying the Ncut algorithm to the image
~ pixels. With the partitioning parameter K=5 and
R* UU T
table I depicts the weight matrix W of the region
7. Go to step 5
nodes. Fig. 1(d) is the display results of the Ncut
During the whole segmentation process, we have
algorithm partitioning. With the partitioning
very heavy computation of matrix-vector
parameter K=5. Fig. 2(d) shows the results of the
multiplication, for example, in solving the k leading
1 1
proposed algorithm partitioning.The Fig. 1 image
2 2 includes the animal in the foreground and the
eigenvectors of D WD , finding the orthonormal
bulrush in the background. Therefore, an effection
matrix, etc. However, most of the time, the matrix in
segmentation algorithm should distinguish the
the matrix-vector multiplication stay the same.
animal from the bulrush background. In Fig. 2(b),
Therefore, we can take advantage of the computing
many separated regions are segmented by Ncut
speed of matrix-vector multiplication on GPU to
algorithm, it is not a good segment result. In Fig.
enhance the overall segmentation speed.
2(c), the Ncut algorithm is directly applied to the
IV. EXPERIMENTAL RESULTS AND ANALYSIS
image pixels, the image is partitioned into two parts.
In this paper, we demonstrate the superiority of the
from Fig. 2(c) and Fig. 2(d). the segment result
proposed method by comparing the performance of the
distinguish the animal from the bulrush background.
proposed approach to existing methods using two
So the proposed method provides an improved
images.
performance compared to the Ncut algorithm.
The Fig. 2(a) size is 100×100. Set h=(x,y)=(8,1) .
With the partitioning parameter K=5. the analysis of
the segment result is the same as the Fig. 1.
Finally, we consider the computational cost of the
proposed algorithm and compare it with that of the
Ncut method. A PC with a 2.8-GHz Pentium CPU
and 1G memory is applied to carry out the
computation of the proposed method which consists
of two sections. The first section is to segment the
where F(i)={ R(i) ,G(i) ,B(i)} is the color vector of original image using the Ncut method. The second
region i, if the image is the gray-level image, then section is to partition the region nodes by the Ncut.
F(i) is the gray value and X(i) is the spatial location The Ncut method [4] requires 8-9 s to process the
of region node i. in this paper, we defined the F(i) is image of 100x100 pixels. When use the proposed
the average gray of the region i, and the X(i) is the method, an image has to be decimated into a size of
center of the region i. in this case, the select of F(i) 120×120, the required time is only 8-10 s. Table II
and X(i) are more representative. In order to compares the computational cost between the
compare other method to the proposed algorithm, we proposed algorithm and the Ncut method for the
use the example of two image, as depicted in Fig. 1 images depicted in Fig. 1 and Fig. 2.
and Fig. 2. Fig. 1 image size is 100×100. Fig. 1(a)
shows the original image, and Fig. 1(b) depicts the
resultant image after applying the Ncut algorithm,
with the red contours depicting the boundaries
between the regions. We set h=(x,y)=(8,1) . As a
result, the Ncut algorithm produces 8 regions. This
eight regions of R(i) in F(i) is
34,51,54,75,78,82,84,and 193 respectively (from
small to large). The G(i) and B(i) are not list.
Therefore, the constructing the weight matrix W of
the graph based onregion nodes using the Ncut
algorithm is easier and smaller than that based on
pixels. Fig. 1(c) shows the partitioning results by
COPO4O2-5
CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

(a)
(a)

(b)

(b)

(c)
(c)

(d)
Fig. 1. (a) Original image. (b) The resultant image after applying
the Ncut Algorithm for edge computation. (c)The segmented
results by directly applying the Ncut algorithm to the image pixels.
(d)The results of the Ncut algorithm partitioning .
(d)
Fig. 2. (a) Original image. (b) The resultant image after applying
the proposed Algorithm for edge computation. (c)The segmented
results by applying the proposed algorithm to the image pixels.
(d)The results of the proposed algorithm partitioning .

COPO4O2-6
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

1 2 3 4 5 6 7 8
1 0 0 0 0 0 0 0 0
2 0 0 0 0 0 0 0 0
3 0 63.96937 63.04625 0 0 0 0 0
4 0 0 0 61.45696 0 0 0 0
5 0 0 0 0 56.69138 0 0 0
6 0 0 0 0 0 52.0375 53.44258 54.79843
7 0 0 0 0 0 0 0 0
8 0 0 0 0 0 0 0 0

TABLE I. WEIGHT MATRIX W OF ALL REGION NODES

No. of Time of Time of

Image Segementation Ncut Proposed
Size location(x) location(y) gradient(gx) gradient(gy) Threshold desired Method(s) Method(s)
Fig.1 100x100 8 1 20 3 0.02 5 7.0159 4.8593
Fig.2 150x150 8 1 20 3 0.02 5 NIL 12.1754
Fig.3 200x200 8 1 20 3 0.02 5 NIL 14.61
Fig.4 250x250 8 1 20 3 0.02 5 Nil 46.4527

TABLE II. COMPARED COMPUTATIONAL COST BETWEEN

THE PROPOSED METHOD AND THE NCUTMETHOD

V CONCLUSION minimization in vision. IEEE Transactions on Pattern

In this paper, efficient image segmentation algorithm
has been proposed and designed based on the
conventional Ncut algorithm . The effectiveness and
robustness of the proposed algorithm have verified by
some experimental results to express an improved
performance compared to the Ncut algorithm. Also, the
significant reduction to the computational cost of the
proposed algorithm in the experiments is favorable for
practical applications.
VI Future Work
We plan to deliver a tool that will allow for a
computational fast and efficient image segmentation
that can be applied to a video sequence.
References
[1]. Boykov,Y. and Kolmogorov, V. Computing geodesics
and minimal surfaces via graph cuts. In International
Conference on Computer Vision, vol. I:26-33, 2003.
Analysis and Machine Intelligence, 26(9):1124-1137,
2004.
[3]. Shi, J. MATLAB Normalized Cuts Segmentation
Code, Online. Internet. University of Pennsylvania. 15
May 2007 hhttp://www.cis.upenn.edu/ jshi/software/i.
[4]. Shi, J. and Malik, J. Normalized cuts and image
segmentation. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 22(8):888-905,2000.
[5]. J. Shi and J. Malik, “Motion segmentation and tracking
using normalized cuts,” Proceedings of the Sixth
International Conference on Computer Vision (ICCV),
Bombay, India, pp. 1154-1160, 1998.
[6]. Georgescu, B. and Christoudias, C. Code Available to
Download, Online, Internet.Rutgers University:
Robust Image Understanding Laboratory. 15 May 2007

[2]. Boykov,Y. and Kolmogorov, V. An experimental

comparison of min-cut/maxflow algorithms for energy
COPO4O2-7
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

OPTIMIZATION OF WIRED NETWORKS USING LOAD

BALANCERS
Yaduvir Singh1, Vikas Gupta2, Rahul Malhotra3
1
Thapar University, Patiala, Punjab
2,3
Adesh Group of Institutions, Sri Muktsar Sahib
1
yad_pra@yahoo.com
3
blessurahul@gmail.com
Abstract: - Computer Network is an interconnection of computers
for sharing of data and network resources. With the evolution of
Internet, the networking technology has not been just confined to
resource sharing but has boomed in the arena of cellular and
mobile communication as well. Rapid development in the field of
very large scale integration of complex circuits on to a smaller chip
has led to the evolution of high speed computer networks. In this
paper, performance analysis of the wired computer networks
through simulation has been attempted using OPNET as network
simulating tool. For wired networks, the performance parameters
like delay and throughput have been investigated with varying
transmission links. The load-balancing has been analyzed through
traffic sent, traffic received and CPU utilization parameters.
Keywords: OPNET, Load Balancers, Ethernet.
I. INTRODUCTION
Networks are being upgraded or deployed from scratch
all over the world. The field of network planning is thus
becoming all the more significant. Networks have grown rapidly
over the past few decades, with a challenge of connecting the
relevant networks together, so that the users can share the
network’s wealth. The undying thrust of humans to excel and fly
beyond the limits of horizon has provided the impetus for the
up-gradation of different technologies and Wireless Technology
is the blessed outcome of this effort. Change is the law of
nature. Over a span of time the wireless Technology has
overshadowed the wired technology. Wireless Technology is a
rapidly growing segment of the communications industry, with a
potential to provide high-speed, high quality information
exchange between the portable devices located anywhere in the
world. A wireless communication is a flexible data
communication system implemented as an extension to or as an
alternative for a wired communication. Wireless Local Area
Networks (WLANs) have been developed to provide users in a
limited geographical area with high bandwidth and similar
services supported by the wired Local Area Network (LAN).The
IEEE (Institute of Electrical and Electronics Engineers) 802.11
standards offer performance nearly comparable to that of
Ethernet. In addition, they provide scalability, mobility,
flexibility and ease of wireless access. Unlike wired networks,
wireless LANs, based on IEEE 802.11 standards, transmit and
receive radio waves through the air medium. The
communication can be between a wireless client and an Access
Point (AP), as well as among two or more wireless clients
within a certain range of each other. A wireless LAN basically
consists of one or more wireless devices connected to each other
in a peer-to-peer manner or through APs, which in turn are
connected to the backbone network providing wireless
connectivity to the covered area. As a result of wireless
technology, the traditional ways of wired networks have become
inadequate in meeting the challenges of present arena posed by
our collective lifestyles. Now a day’s wireless technology has
become a part of our life with major of the present applications
being upgraded to support wireless technology and all upcoming
applications are being manufactured with integrated wireless
support. Wireless technology provided a boon to the users of
computers, mobiles and other wireless technology enabled
gadgets by providing cheap and flexible wireless
communication access, without using the cables. Wired LAN is
slowly getting replaced by wireless LAN due to the ease in
access, mobility and scalability provided by it.
Among a wide range of applications of wireless
technology, few prominent areas of application are
telecommunications, security systems, industrial automation,
wireless sensor networks, distance learning education, health-
care and retail, as it is also easy to install on campuses, airports,
in stock markets, offices, hospitals, and other places.
II. WIRED LOCAL AREA NETWORKS
Wired Local Area Networking includes several
technologies such as Ethernet, token ring, token bus, FDDI and
ATM LAN. Some of these technologies survived for a while,
but Ethernet is by far the dominant technology. Evolution from
a 10Mbps Standard Ethernet to bridged Ethernet and then to a
switched Ethernet paved a way for faster Ethernet. Ethernet is a
contention media access method. In its purest form, contention
means that the computers are contending for use of the
transmission medium (wire). Any computer in the network can
transmit at any time (first come, first serve).
IEEE 802.3 standard specifies CSMA/CD as the access
method for first-generation 10-Mbps Ethernet, a protocol that
help devices share the bandwidth evenly without having the two

COPO5O1-1
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

devices transmit at the same time on the network medium. This

CSMA/CD protocol was created to overcome the problem of
collisions that occur when the packets are transmitted
simultaneously from different nodes over the same medium.
OPNET has been used for the simulation of these
computer networks. It provides a rich library of models for
implementing wireless and wired network models. In this
chapter we begin with description of the performance metrics
used in the thesis work. The performance metrics focused upon
in wired computer networks are:
Ethernet- Delay (sec): This statistic represents the end to end
delay of all packets received by all the stations.
FTP- Traffic Received (bytes/sec): This statistic means
average bytes per second forwarded to all FTP applications by Figure 2 Wired LAN Network Model
the transport layers in the network.
Ethernet – Traffic Received (bits/sec): This statistic defines 3.1 Performance Analysis of Wired LAN Model
the throughput (bits/sec) of the data forwarded by the Ethernet The objects used to design the network model of wired LAN
layer to the higher layers in this node.
are:
HTTP – Page response time (sec): Specifies time required to
retrieve the entire page with all the contained inline objects.  One Server: Web Server
 Two Switches: ethernet16_switch node model
 Two Circle Annotations
 Thirty Users: 10 users per LAN.
 Link Types: 10 Base T, 100 Base T, 1000 Base X
 Application Configuration: To define the type of
network load.
 Profile Configuration: To create new profiles and
assign the configured applications

3.2 Performance Analysis using Link Type Variation

The network design shown in Figure 2has been
modeled. In this network design the links have been varied as
shown in Table 1.
Figure 1 Performance analysis metrics of Wired LAN Table 1 Performance Analysis scenarios for Link type variation

III. PERFORMANCE ANALYSIS Parameter Scenario 1 Scenario 2 Scenario 3

In this simulation scenario, comparison has been done Type of Link used 10 Base T 100 Base T 1000 Base X
by varying the types of media (Ethernet links) used in the
No. of 30 (10 Per 30 (10 Per 30 (10 Per
networks for communication between the server and the clients Workstations LAN) LAN) LAN)
in a wired LAN. The network design as shown in Figure 2has For performance analysis, following metrics have been
been modeled. A new project is created with a new scenario.
investigated:
This new scenario model created has logical network scale with
Ethernet topology. The object palette of the selected Model Ethernet- Delay (sec): This statistic represents the end
family is used to select the objects required to create the to end delay of all packets received by all the stations.
network model (using drag and drop method) based on the
required technology. The graphical user interface helps to
arrange and place the objects as required and also provides an
ease to connect these objects conveniently.

COPO5O1-2
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

Figure 3 Ethernet Delay (sec) Figure 4 Ethernet Throughput (bits/sec) for link variation

As shown in Figure 3, the delay occurs due to the 3.3 Performance Analysis using Load-Balancer
heavy traffic that passes through the network. The cause of
heavy traffic is the large number of users (30) as shown in the In another simulation for performance analysis, the
Figure 1. The amount of traffic sent and traffic received number of stations in a network is varied as shown in the
increases/ decreases depending on the number of users and the figure5, in which there is only one server, a load-balancer, a
amount of data accessed. In first scenario according to Table firewall and other network objects as per the specifications of
4.1, the type of Ethernet link used is 10 Base T. The graphs in the network required.
Figures 3, show the delay statistics (which is maximum than the
other two types of links). The reason is the low data rate of the
10 Base T links as compared to the other two types of links
listed in Table 1 i.e. 100 Base T and 1000 Base X. Another
parameter used for performance analysis is:
Ethernet – Traffic Received (bits/sec): This statistic
defines the throughput (bits/sec) of the data forwarded by the
Ethernet layer to the higher layers in this node.
The Ethernet throughput depends on the amount of
delay that occurs in the transmission of packets. Figure 3
indicates that the delay in case of 100 Base T and 1000 Base X
is less as compared to a network with 10 Base T. Hence the
throughput achieved in the case of 100 Base T and 1000 Base X
is much higher as compared to throughput achieved in case of
Figure5 Wired LAN Network Model using Load Balancer
10 Base T as shown in following figure 4. Moreover, the links –
100 Base T and 1000 Base X provide more bandwidth as Then scenarios described in Table 2 have modelled and
compared to the 10 Base T. Hence the traffic received through simulated using the network design shown in figure 5. The table
the Ethernet increases with the increase in bandwidth. shows the variations in the number of nodes:

Table 2 Performance Analysis scenarios for varying number of nodes

Parameters Scenario Scenario 5 Scenario Scenario 7
4 6
No. of Workstations 10 20 30 30
No. of Servers 1 1 1 3
Load Balancing None None None Random
Policy policy
When the load balancer receives a packet from a client
machine, it must choose the appropriate server to handle the
request. The load balancer will use the load balancing policy to

COPO5O1-3
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

determine which server is most appropriate. Following load

balancing policies can be used:
Random: The load balancer chooses one of the candidate
servers at random.
Round-Robin: The load balancer cycles through the list of
candidate servers.
Server Load: The load balancer chooses the candidate server
with the lowest CPU load.
Number of Connections: The load balancer keeps track of the
number of connections it has assigned to each server. When a
new request is made, it chooses the server with the fewest
connections.
The Simulation study on load balancer has been done for the
following reasons: Figure 7 Response time (sec) for varying number of nodes
To determine, how much load on the individual servers can be
The response time associated to the highest load level
reduced, using a load balancer, in a network with heavy
under balancing is lower than any other single-server
Internet applications.
configuration (including the one with the smaller number of user
To determine the best load balancing for improving the
sessions) as is visible from figure 7.
performance of wired networks.

The following statistics are investigated for performance

analysis:
Traffic Sent (in Bytes/sec or Packet/sec): The amount of
traffic sent to a particular application server.
Traffic Received (in Bytes/sec or Packet/Sec): The amount
of traffic received from a particular application server.
As can be seen in Figure 6, as the number of user
increases, more traffic is generated. It is also observed that the
introduction of load balancing for a single server has no effect
on the level of generated traffic.

Figure 8 A wired LAN network model with load- balanced multiple servers

Using the network design modeled in Figure 8, four

Figure 6 Traffic sent (bytes/sec) for varying number of nodes
But as can be seen in figure 7, the use of load-balancing for
multiple servers (in figure 8) can significantly improve
performance.
scenarios (Table 3) have been designed. Firstly, the scenario
with random policy in which the load balancer randomly
chooses from one of the three application servers according to
the specified weight. Secondly, load balancer with the round-
robin policy in which the load balancer chooses each application
server in turn depending on the server weight. Thirdly in
scenario 9, the load balancer has server-load policy, in which the
load balancer chooses the server with the lowest load at the time
when a request is made. Then in scenario 10 where the load
balancer with number-of-connections policy tracks the number
of connections it has with each server. When it needs to
forward a request, the load balancer chooses the server with the
least connections.
Table 3 Performance Analysis scenarios with load-balancer

COPO5O1-4
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

For performance analysis, following statistics are collected at

Parameters Scenario Scenario Scenario Scenario 10
the load balancer and the server:- 7 8 9
The statistics at the load balancer are collected per No. of
30 30 30 30
application and per server. When load balancing is Workstations
enabled the statistics collected are: No. of Servers
3 3 3 3
Traffic Sent (in bytes/sec or packet/sec): The Load Random Round Server Number of
amount of traffic sent to a particular application server. Balancing policy Robin Load connections
Traffic Received (in bytes/sec or packet/sec): The Policy Policy Policy
amount of traffic received from a particular application
server.
The analysis from figure 9 shows averages for traffic
associated to each balanced server. In this section of
performance analysis, the simulation has been carried out for
10.0 hours.

Figure10 Traffic sent & Received (Bps) using No. of connections Policy

The investigations from Figure 10 showing average

traffic sent and Figure 4.17 showing average traffic received by
each balanced server indicates that the number of connections
policy provides faster load stability and the servers converge to
same average load over a period of 3 to 4 hours.
Figure 9 Traffic sent & received (bytes/sec) using server-load policy

The investigations from figure 9 for traffic sent

(bytes/sec) and traffic received (byte/sec) reveal that after a
period of more than four hours, all the balanced servers seem to
converge to similar average loads. This analysis has been done
using server-load balancing policy.
For further analysis of the network design presented in
Figure 8, a performance comparison of other load balancing
policies like random policy, round-robin policy and number of
connections policy, under the same user activity, has also been
done.
Figure 10 shows comparison using number of
connections load balancing policy on three load balanced
servers. Similarly Figure 4.17 and 4.18 show comparisons for Figure 11 Traffic sent & Received (bytes/sec) using Round-Robin policy
traffic sent and traffic received using round – robin policy and Figure 11 shows the average traffic sent and traffic
Figure 4.19 and Figure 4.20 show comparison of traffic sent and received (in bytes/sec) by each balanced server under the round-
received using random load balancing policy. robin policy. This policy assigns each incoming requests to the
next server in the sequence. Though the server 1 and server 2
with similar selection weights as assigned in the load balancer,
tend to have same average activity, the sever 3 with more
selection weight shows variance in the average server load.

COPO5O1-5
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011
Figure 12 Traffic sent & Received (bytes/sec) using Random policy
The statistics of the average traffic sent and received in
bytes/sec under the random policy shows the same phenomenon
as shown by the round robin policy in Figure 12. This is due to
the selection weights assigned to the balanced server using the
load balancer. Another comparison of the load-balancing
policies has been done by simulating the effect of all these
policies on the same server.
Figure 13 Traffic sent Received (bytes/sec) using different load balancing policies
The investigations present that while average traffic
received or sent (in bytes/sec) with the number of connections
policy is more than others, the average traffic received or sent
(in bytes/sec) with the server-load policy is lesser than others
(Figure 13).
The following statistic is collected at each of the balanced
servers.
CPU load (%): The percentage load on server.
The following figures present the percentage load on
server. This statistic reports the utilization, in percentage (%),
of the 'simple CPU'. The 'simple CPU' is used to model the IP
packet forwarding delays and application processing delays.
This statistic also takes into account the background CPU
utilization specified in the "CPU background utilization"
attribute. This statistic measures the utilization of central CPU
only. The analysis of figures 14 (which defines the impact of
number of connections policy & random policy on all the
servers), Figure 15(which defines the impact of round-robin
policy & server-load policy on all the servers) reveal that the
load on all the balanced servers averages to be approximately
same using the number of connections policy and server-load
policy.
Figure 14 CPU Utilization using number of connections policy & Random policy
But when the random policy and round robin load balancing
policies are used, Figures 14 shows that the CPU utilization (%)
of server 3 is more while there is little or no variance of CPU
load for server 1 and server2. The cause of more utilization of
server 3 is the assigning of selection weights in the load server.
Figure 15 CPU Utilization using Round Robin policy& Server load policy
The selection weight is the attribute which is applicable
only when the load balancing technique is either set to "round-
robin" or "random". It is used when determining the server to
choose. The probability of choosing a given server is: (selection
weight) / (total weight). The value of selection weight assigned
to server 3 is 10, which is twice the value of selection weights
assigned to the server 1and server 2. The overall analysis
indicates that the CPU utilization is minimum using server load
balancing policy.
4. CONCLUSIONS
COPO5O1-6
 CONFERENCE ON “SIGNAL PROCESSING AND REAL TIME OPERATING SYSTEM (SPRTOS)” MARCH 26-27 2011

In this paper work performance analysis of the wired network [11] Dr. Reinhard Kuch, “Studienbrief 2: Simulation of
configuration through simulation was started with the networks using OPNET ITGURU Academic Edition v
investigation of the network performance using various types of 9.1,” version: 1.0, 26-04-2009.
[12] “Opnet_Modeler_Manual,” available at
links. The impact of various network configurations on the
http://www.opnet.com
network performance was analyzed using the network [13] T.Velmurugan, Himanshu Chandra and S. Balaji,
simulator- OPNET. It has been investigated that performance of “Comparison of Queuing disciplines for Differentiated
the wired Networks is good if high speed Ethernet links are used Services using OPNET,” IEEE, ARTComm.2009, pp.
under heavy network loads. Moreover, the mechanism of load 744-746, 2009.
balancing also improves the performance by reducing and [14] Yang Dondkai and Liu Wenli, “The Wireless Channel
balancing the load equally among multiple servers. Modeling for RFID System with OPNET,” in the
Proceedings of the IEEE communications society
REFERENCES sponsored 5th International Conference on Wireless
communications, networking and mobile computing,
[1] Behrouz A Forouzan, Data Communications and
Beijing, China, pp. 3803-3805, September 2009.
Networking: Tata McGraw Hill, 4th Edition, 2007.
[15] Ikram Ud Din, Saeed Mahooz and Muhammad Adnan,
[2] Sameh H. Ghwanmeh, “Wireless network performance
“Performance evaluation of different Ethernet LANs
optimisation using Opnet Modeler,” Information
connected by Switches and Hubs,” European Journal of
Technology Journal, vol. 5, No 1, pp. 18-24, 2006.
Scientific Research, vol. 37, No 3, pp. 461- 470, 2009.
[3] Sarah Shaban, Dr. Hesham, M.El Badawy, Prof. Dr.
Attallah Hashad, “Performance Evaluation of the IEEE
802.11 Wireless LAN Standards,” in the Proceedings
of the World Congress on Engineering-2008 , vol. I,
July 2-4, 2008.
[4] A. Goldsmith, Wireless Communications: Cambridge
University Press, 1st Edition, August 2005.
[5] Mohammad Hussain Ali and Manal Kadhim Odah,
“Simulation Study of802.11b DCF using OPNET
Simulator,” Eng. & Tech. Journal, vol. 27, No. 6, pp.
1108-1117, 2009.
[6] Hafiz M. Asif, Md. Golam Kaosar, “Performance
Comparison of IP, ATM and MPLS Based Network
Cores Using OPNET,” in 1st IEEE International
Conference on Industrial & Information Systems
(ICIIS 2006), Sri Lanka, 8-11 August, 2006.
[7] Dibyendu Shekhar, Hua Qin, Shivkumar
Kalyanaraman, Kalyan Kidambi, “Performance
Optimization of TCP/IP over Asymmetric Wired and
Wireless Links” in the Proceeding of conference on
Next Generation Wireless Networks: Technologies,
Protocols, Services and Applications (EW-2002), –
Florence, Italy, February 25-28, 2002.
[8] Born Shilling, "Qualitative Comparison of Network
Simulation Tools,” Institute of Parallel and Distributed
Systems (IPVS), University of Stuttgart, 2005.
[9] “OPNET IT Guru Academic Edition” available at
http://www.opnet.com/university_program/itguru_acad
emic_edition
[10] Ranjan Kaparti, “OPNET IT Guru: A tool for
networking education,” REGIS University.

COPO5O1-7