ASIAN JOURNAL OF ENGINEERING, SCIENCES & TECHNOLOGY, VOL.

3, ISSUE 2 SEPTEMBER 2013

A Neural Network Based Pakistani License Plate

Recognition System and Controlling Mechanism
Syed Fahad Ali, Muhammad Moinuddin, Waseem Taj, Farwa Jaffri, Summiya Tasneem

Abstract:- In this paper a complete system of
recognition and controlling of any vicinity security
parameter is proposed based on the neural
network architecture for automation and self
decision making for the vehicle authentication and
permission for entering into the premises of Iqra
University Main Campus .This Implemented
Process is based on the Segmentation of the
characters from the Number Plate Of the Vehicles
entering into the premises, by using the well
known method of projection method and
recognition of the characters is done by using the
BPNN based algorithm designed for the said
problem.
Index Terms- Authentication, BPNN, Projection,
Segmentation , Recognition.
algorithm is written to extract the characters in the
whole image,the localization of the license plate was
carried out by using the projection method devised
by Zidouri [5],the sequential flow for the segmented
extraction of the license plate and then the characters
is shown in the Fig.[1].
II.IMPLEMENTED SYSTEM
The given procedure is carried out using the
MATLAB® software algorithm designed by
considering the following factors.
• Image Acquisition
• License Plate Detection
• LP Segmentation
• Character Detection
• Character Segmentation

I.INTRODUCTION

The implemented system is the first in its kind in

Pakistan for the security purposes of any premises
that particularly needs any special protocol of
security. This system uses License Plate detection
and recognition system using Artificial Neural
Network [1].
The Work for the recognition of License Plate has
been done in various parts of the world with different
mechanisms like Dr.Moin & M.Deriche implemented
the same purpose in Saudia Arabia for its license
plate with two different sizes of the License
Plates[2],Among the commercial mechanisms which
are widely used for the license plate recognition are
Car Plate Recognition(CPR) by Barosso[3],
Nijuhis[4] also gave a method on the basis of the Figure 1 Procedure For Character Extraction
Template matching technique in which whole A. IMAGE ACQUISITION
template is being matched with the original one to The first and most important task was to capture an
make decision. The first and most important task was image of the vehicle inbound to the premises of the
to capture the image via a high resolution camera, its Iqra University[6],the task was completed by using a
input is sent to the processing unit and their an digital High Definition camera and then its output is
transmitted to the Personal Computer having
Iqra University, Main Campus, Shaheed e Millat, Karachi
engr.fahad01@gmail.com MATLAB software enabled and installed into ,the

69
ASIAN JOURNAL OF ENGINEERING, SCIENCES & TECHNOLOGY, VOL. 3, ISSUE 2 SEPTEMBER 2013

picture is loaded into the directory of the MATLAB

in the form of a 3*3 matrix as the image captured is
in a form of RGB. Then the image is converted and
then transformed into binary by using a threshold
level by which maximum information can be used for
the processing in the segmentation and detection
phase of the system. Our system used a level of 0.57
determined by the MATLAB software command
"greythresh", as shown in Figure[2],the converted
image in the gray and binary are shown in the Fig.[3]
and Fig[4],the conversion is done by using the
MATLAB® command set for the conversion so that
the processing can be done for further decision. The
segmentation is done by using the thresholds and the
localizations check in the whole image by calculation Figure 4 Binary Image
of the "polyarea" of the each boundary defined by
the command "bwboundary" of each closed A. LICENSE PLATE SEGMENTATION AND
boundary used. The usage of the boundaries is to EXTRACTION
determine the exact location of the license plate in The licesne plate is then segmented and extracted by
the whole given image by evaluating the thresholds finding the conditions in whole image that satisifies
and checks for the allocation of the license plate, the the conditonal checks for the license plates set by
fixed area of the license plate is also determined by mathematical analaysis and the graphical input taken
using the command "ginput" and using the following by Matlab and the the area that satisfied the
equation. conditional checks were marked with another color
so that there is a differential ability between whole
𝐴𝑟𝑒𝑎 = 𝐴𝑐𝑡𝑢𝑎𝑙 𝐿𝑒𝑛𝑔𝑡ℎ ∗ 𝐴𝑐𝑡𝑢𝑎𝑙 𝑊𝑖𝑑𝑡ℎ … … . (2.1) image and the license plates Fig[5]
The next step was to extract the highlighted licesne
plate and for that particular process we used the
coordinate system and extracted the whole image by
using the axiom that the coordinates of the boundry
are found same for any two points for x axis and two
for the y axis then by using the extraction process for
the matrices then they are plotted to have the license
plate Fig[6]

Figure 2 Original Image

Figure 5 License Plate Boundary Detected

Figure 3 Gray Scale Image

70
ASIAN JOURNAL OF ENGINEERING, SCIENCES & TECHNOLOGY, VOL. 3, ISSUE 2
Figure 6 Extracted License Plate
Then the next step was to clean the license plate
because of the reason that it contained some points
and marks which may hinder the process of
classification and edge detection for doing that a
median filter of order 2 is used to eliminate the spots
and marks in order to have its optimum output we
used its impulse response to be of 2*2 instead of 3*3
which was its inbuilt value. The result was clear and
clean license plate For the calculation and analysis
the license plate is inverted by replacing the 1’s with
0’s to make the black area visible so that the values
of 1’s can be utilized in segmentation and processing
processes Fig.[8]
Figure 8 Inverted and Filtered Image
The next step is to remove the extra words like punch
holes and word "SINDH" placed in lower middle
center of each license plate in Pakistan, To do so a
projection method is used to eliminate the extra stuff
like alphabets in the image. The projection method
works on the simple principle of finding the
difference of one row/column with the preceding one
then finding the absolute one this will give the
unsorted positions of the image where there will be
black spots so by arranging it we have the required
positions which were spotless.
The same process in used to remove the word
“SINDH” and this process is used in both horizontal
and vertical dimension in remove the extra stuff like
the bolt holdings and etc Fig[9]
Figure 9 Words Removed
71
SEPTEMBER 2013
A. CHARACTER SEGMENTATION AND
EXTRACTION
The final and foremost important phase of the system
is to localize and segmentize the characters one by
one and effectively as the next step is to recognize
the extracted images from the neural network
designed. To do so another projection procedure is
commenced for the extraction of the characters from
the extracted license plate and by finding the correct
positions of the gaps between the character spaces it
is plotted as an image and all the characters are being
extracted from the license plate and in real from the
real image and this system can detect license plate
form all over the image of the inbound of vehicle to
the premises of Iqra University.
III.RECOGNITION SYSTEM
The next and the core task of the system is to detect
and recognize the characters being extracted from the
image processing algorithm to the installed database
automatically and to do so a neural network is
implemented. Neural network is a system that is
based on human neural system and it is trained on the
features of the characters extracted also with the
noisy data and complete data sets are provided as one
of the inputs to the network as shown in Fig.[10],the
neural network has governing equations and has a
non linear functions to be implemented for the
correct recognition and detection of the pattern or the
task given to it. The output of jth neuron v can be
calculated by summing it with weights of the
particular layer neuron denoted by w and the
activation function applied onto it n.
𝑣𝑗(𝑛) = ∑𝑙𝑘=0 𝑤𝑗𝑖(𝑛) 𝑦𝑖(𝑛) … . . (3.1)
Where l, k are the limits of the training of the neural
network and the error (cost function) can be
calculated of the system by using the following
relation.
𝑒𝑗(𝑛) = 𝑑𝑗(𝑛) − 𝑣𝑗(𝑛) … . . (3.2)
Where dj(n) is the desired output which should be the
output of system is provided and vj(n) is the actual
output from the neural network Fig.[10]
ASIAN JOURNAL OF ENGINEERING, SCIENCES & TECHNOLOGY, VOL. 3, ISSUE 2 SEPTEMBER 2013

The above table shows the recognized character

which is compared with the character that was used
to train the neural network system.

v. The Cost Function

Cost function is defined as the error performance of

the ratio of the weights updated and the performance
behavior of the neural system and the cost function is
ideally is a decay function with time when the
Figure 10 Multilayer Feed Forward Network weights update and the decision are more precise and
accurate then in this system the weights were
The neural network uses is Back Propagation Feed
decaying factor which depicts that the as the training
Forward Neural Network who’s input is the
session value is increased the weights are more
characters extracted by the extraction process, the
updated and the values are decreased which shows
system is trained for multiple shapes of characters for
that the system is now self reliable and self sufficient
a learning rate of .005 and the number of trials were
to make decisions in real time Fig.[11]
found to be 1500 this sets up the neural network to
perform at its peak value that is 88% in recognition.
This system used multiple values tested upon and
then final value at which each character was
recognized by the system effectively was proposed.

IV.EXPERIMENTAL RESULTS
The trained and designed neural network has an
efficiency of about 80% and it recognized the
characters of the test image very perfectly and the
system worked efficiently. The cost function graph is
shown in the Fig.[8]. The system was implemented
for about 3 vehicles parked inside the premises of
Iqra University and had a very good performance
ration in terms of recognition and segmentation from
the image as shown in Table[1]. Figure 11 Cost Function graph
Successful Character 98%
Extraction ACKNOWLEDGMENTS
Successful recognition 99% The authors would like to thank ALMIGHTY
Overall Success 98% ALLAH for his blessings and giving opportunity for
Table.1 Performance Report serving the university and for the country, Also
would like to thank the parents for their support and
The experimental results also show that the system
patience and for Dr. Muhammad Moinuddin for his
recognized the characters from the images very
clearly and the detection process was faster and complete guidance throughout the procedure and
reliability process is shown in the mentioned table. making of this system. Also the authors would like to
thank the administration of Iqra University for their
huge support in completion of the project and the
research work for Pakistani license plate recognition
system. Authors would like to thank Dr. Kamran
Raza for his permission to carry out the research and
would like to thank Sajjad Hussain Zaidi and
Table.2 Performance analyzing of Received and decided Jawwad Ahmed for their technical assistance and
Character guidance throughout the process.

72
ASIAN JOURNAL OF ENGINEERING, SCIENCES & TECHNOLOGY, VOL. 3, ISSUE 2 SEPTEMBER 2013

REFERENCES Telecommunication Department, his contact detail mail:[

dr.moin@iqra.edu.pk]
[1]Simon Haykin,"A comprehensive Neural
Network"2nd edition. Waseem Taj has done Bachelors in
Telecommunication from Iqra Univeristy
[2 ] Mohamed Deriche and Muhammad Moinuddin, Karachi he is currently associated with Saint
developed A Neural Networks based Saudi Arabian Micheal's Convent school Karachi,his contact
detail [email: coldfire_waseem@hotmail.com]
License Plate Recognition System Electrical
Engineering Department, King Fahd University of, Farwa Jaffri has also done Bachelors in Telecommunication from
Petroleum & Minerals. Iqra University and currently associated nowhere, Her work in this
system is to design the drafting.

[3] D.G.Bailey,D.Irecki,B.K.Lim and L.Ying,"Test Summiya Tasneem has done successfully bachelors in
bed for number plate recognition applications", telecommunication from Iqra University and currently associated
Proceedings of First IEEE international workshop on with different associations as an visiting faculty for intermediate
studies.
Electronic design, test and applications
(Delta02),IEEE Computer Society.,2002.

[4] J.Barosso,E.L.Dagless,A.Rafael and J.Bulas-

Cruz,"Number Plate recognition using computer
vision.,"Proceedings of IEEE international
Symposium on Industrial Electronics,vol 3.pp761-
767,1997.

[5] Iqra University established in 2000 in

Karachi,Sindh,Pakistan.One of the Major
Universities for graduate studies with most valuable
and perfect faculty of engineering and sciences.

[6] Zidouri and Sarfaraz," An Automated Saudia

Arabian Number Plate Recognition System,",
International Conference on systems, International
conference on Geometric Modeling and Graphics,16-
18 July 2003,pp.36-41,2003.

[7] S.Chang,L.S Chang,Y.C Chang and S.W

Chang,"Automated Licese Plate Recognition,"IEEE
Transcactions on Intelligent Transportation
Systems,vol,5,pp.42-53,March2004

Syed Fahad Ali (3-3-1991 till date) has done

B.E (Telecommunication) from Iqra University
and is now associated with different universities
as visiting faculty member and currently
persuading his MS(Comm.) from NUST
university. His area of interest are Image
processing and RF communication, his contact details are [email
id: engr.fahad01@gmail.com ]

Dr.Moinuddin is the supervisor of the complete

system and he is currently affiliated with Iqra
University as Associate Professor

73