International Journal of Scientific Research Engineering & Technology (IJSRET)

Volume 2 Issue 10 pp 604-609 January 2014 www.ijsret.org ISSN 2278 0882

IJSRET @ 2014
Strategic Approach to Image Block Segmentation
Avanish Shrivastava, Prof.Mohan Awasathy
Department of Electronics and Telecommunication Engineering, SSCET, Bhilai,
Department of Electrical and Electronics Engineering, Institute of Technology, Korba.
CSVTU, Bhilai, Chhattisgarh, India
AbstractImage segmentation is the problem of
partitioning an image into its constituent components.
Image segmentation being an ill-defined problemthere
is no unique ground-truth segmentation of an image,
against which the output of an algorithm may be
compared. However, many segmentation algorithms have
a parameter that explicitly encodes the number of clusters
and yet, do not have well accepted schemes for its
selection. Choosing the appropriate segmentation scales
for distinct ground objects and intelligently combining
them together are two crucial issues to get the appropriate
segmentation result for target applications. We propose a
block-based method of segmentation that is capable of
dealing with intensities of blocks and processing any one
block at time, out of many blocks for the analysis of
background variations, illumination and comparing those
blocks on the basis of their intensities. This approach
leads to an issue of choosing block sizes. For block-based
classification, an image is divided into blocks, and a
feature is detected for each block by grouping statistics
extracted from the block. Specifically, images are
analysed on a non-overlapping block-by-block basis. The
standard deviation or mean or average of the pixel
intensities of the block can be used as features. It is taken
as a primary feature, for identifying background blocks.
Features and algorithms for the block-segmentation and
classification of regions into respective block are
evaluated for applicability to grayscale images only.
Keywords Segmentation, DCT, tracking, Block
processing.
I. INTRODUCTION
One of the fundamental and critical tasks in many
computer-vision applications is the segmentation of
foreground objects of interest from an image sequence.
Many image processing tasks require to know the
meaning (e.g. object or background) of the image pixels.
Image segmentation is an important process to furnish
such information to many image processing applications
such as pattern recognition and object identification.
Image segmentation is a process of dividing an image
into different regions such that each region is nearly
homogeneous, whereas the union of any two regions is
not (Unnikrishnan R. et. Al. 2007). It serves as a key in
image analysis and pattern recognition and is a
fundamental step toward low-level vision, which is
significant for object recognition and tracking, image
retrieval, face detection, and other computer-vision-
related applications. Note that rather than merely
providing a labelling of all regions in the image, the
segmentation process must extract the object of interest
from the background to support the subsequent feature
extraction and object classification processing. The
accuracy of segmentation can significantly affect the
overall performance of the application employing it
subsequent processing stages use only the foreground
pixels rather than the entire frame. The traditional
processing flow for image-based pattern recognition
consists of image segmentation followed by classification.
This approach assumes that the segmentation is able to
accurately extract the object of interest from the
background image autonomously. The performance of
this subsequent processing is strongly dependent on the
quality of this initial segmentation. This expectation of
ideal segmentation is rather unrealistic in the absence of
any contextual information on what object is being
extracted from the scene.
Previous Work on Segmentation
Various segmentation methodologies developed
previously have been glanced in this paper. Though they
all perform the segmentation operation but they differ
depending upon the application they are meant for. Hence
the different segmentation process are applicable to
respective application only, if they are applied to other
ones, the extent of error addition is unexpectedly high
including the changes in the image which vanishes their
knowledgeablity. These changes in the image can be
caused by variations in the object itself (i.e., different
color or texture), or by variations in the environmental
factors, such as the sensor, lighting conditions, and most
importantly shadow and highlight bands which cause
nonuniform changes in the appearance of the objects.
Again since the existing methods require homogeneity of
the object of interest, any nonuniform changes will lead
to a violation of the homogeneity assumption.
Segmentation is employed in diverse applications, such
as tracking ( Xiaofeng Liu, 2010), action recognition
(Johnson Andrew E. and Hebert M, 2005), gait
recognition (Wenjing Li et. Al. 2008), anomaly detection
(Lina Yi et.Al. 2010), content based video coding
(Xiaofeng Liu, 2010) and computational photography
International Journal of Scientific Research Engineering & Technology (IJSRET)
Volume 2 Issue 10 pp 604-609 January 2014 www.ijsret.org ISSN 2278 0882
IJSRET @ 2014
(Peijun Li et. Al. 2011).There is an abundance of
literature on image segmentation. Methods also have
been defined for post processing the low-level
segmentation to further regularize the segmentation
output. Many approaches have been proposed before,
including the clustering method (Doung G. And Xie M.
2005), active contours (Ping Han Lee, 2011), normalized
cuts (Tao W. and Jin H. 2007), graph-cut-based
approaches (TaoW. and Jin H. 2007)etc.The segmented
regions arerepresented by using the graph structures, and
the N-cut method is applied to perform globally
optimized clustering. Because the number of the
segmented regions is much smaller than that of the image
pixels, the method allows a low-dimensional image
clustering with significant reduction of the complexity
compared to conventional graph partitioning methods that
are directly applied to the image pixels. While
considering every region as only one graph node there are
some inappropriate partitioning, so an improved
segmentation strategy is needed.
Graph-based approach visual group is based on several
key factors such as similarity, proximity, and
continuation (Brejl M. And Sonka M. 2000). A graph is
partitioned into multiple components that minimize some
cost function of the vertices in the components or the
boundaries between those components. So far, several
graph cut-based methods have been developed for image
segmentations i.e. (Tao W. and Jin H. 2007) proposed a
general image segmentation approach based on
normalized cut (N-cut) by solving an eigensystem and
developed an image-partitioning approach by using a
complicated graph reduction. Besides graph-based
approaches, there are also some other types of image
segmentation approaches that mix the feature and spatial
information (Lina Yi et. Al. 2010). These methods are
basically data-driven approaches. The data-driven method
sometimes fails to produce satisfactory segmentation
when there are shadows, occlusion, cluttering, low
contrast area, or noise in the image. Incorporation of
contextual and prior information is very important to
improve segmentation under such situations.
The paper (Tao W. and Jin H. 2007) concerns an N-cut
method in a large scale. It has been empirically shown
that the N-cut method can robustly generate balanced
clusters and is superior to other spectral graph-
partitioning methods, such as average cut and average
association (Lee P.H. et. Al. 2011). The N-cut method has
been applied in scene detection (Johnson A.E. and Hebert
M. 2005) and cluster-based image retrieval (Doung G.
And Xie M. 2005). However, image segmentation
approaches based on N-cut, in general require high
computation complexity and therefore, are not suitable
for real-time processing.Previously, it was thought that an
efficient solution to this problem is to apply the graph
representation strategy on the regions that are derived by
some region segmentation method (Zhang L, and Ji Q,
2011). For example, (Brejl M. and Sonka M. 2000)
developed an image segmentation method that
incorporates region based segmentation and graph-
partitioning approaches. This method first produces a set
of over-segmented regions from an image by using the
watershed algorithm, and a graph structure is then applied
to represent the relationship between these regions.
Choosing the appropriate segmentation scales for distinct
ground objects and intelligently combining them together
are two crucial issues to get the appropriate segmentation
result for target applications.
After analysis of various methodologies of segmentation
and glancing it with the requirement in this work,
basically, there are mainly three limitations of existing
image segmentation algorithms. Firstly, Existing
segmentation algorithms are built upon the following two
common underlying assumptions (i) the object of interest
should be uniform and homogeneous with respect to
some characteristic, (ii)adjacent regions should be
differing significantly. These assumptions, however, are
rarely met in real-world applications. Secondly, there are
few metrics available for evaluating segmentation
algorithms. Some of the proposed measures of
segmentation quality include: edge-border coincidence,
boundary consistency, pixel classification, object overlap
and object contrast. However, none of these metrics can
be passed as to be widely accepted as ideal and many
require ground truth information. Some methods have
used multiple hand-segmentations from a number of
human experts to define a segmentation quality metric
that is really measuring the segmentation consistency
(Unnikrishnan R. et. Al. 2007). This metric would clearly
not be appropriate for any real-time application and also
requires a considerable offline effort. Thirdly, the final
limitation of existing segmentation algorithms is their
inability to adapt to real-world changes.
II. PROPOSED IMAGE BLOCK SEGMENTATION
The proposed technique has following main components.
1. Training
a. Divide testing image into non-
overlapping square blocks with equal
size and detect the texture of the image
separately.
b. Extract a feature (standard deviation) for
each block.
c. Estimate model parameters based on the
feature vector and their hand labelled
classes.
d. Assign each block a unique individual
number.
2. Testing
a. Compare the block segmented image
with texture bounded image.
International Journal of Scientific Research Engineering & Technology (IJSRET)
Volume 2 Issue 10 pp 604-609 January 2014 www.ijsret.org ISSN 2278 0882
IJSRET @ 2014
b. Blocks are mapped with textures of
image.
c. Unique block numbers involved in a
detected texture is given out.
This new algorithm is proposed for segmenting
images by classification of status of pixels within the
blocks. The general procedure for block-based
segmentation and classification partitions an image into
non-overlapping nn blocks. Block- segmentation and
classification algorithms can be based on the information
provided by a multidimensional feature space.
Algorithms can be improved by context-based
classification supplementing first pass classification from
feature information. The pixels of blocks are handled in
block by block manner, as illustrated in figure. A block of
pixels is generally defined as (qq) array of pixels. The
proposed segmentation algorithm decides one of the
following three possible outcomes for each block of
pixels.
(i) All the pixels within the blocks are classified as
object,
(ii) All the pixels within the blocks are classified as
background,
(iii) No change to pixel classification.
The arrangement of block used in this algorithm is just
one of many possible choices. The choice of arrangement
may have influence on the segmented image. However
the segmented images using different arrangement of
blocks are expected to be similar to each other because
each block is still tested for change of class.There are
three possible actions on the block viz.
(i) All the pixels changed to object.
(ii) All the pixels changed to background.
(iii) No change is made.
Issue of Block size
For block-based image classification algorithms, images
are divided into blocks, and decisions are made
independently for the class of each block. This approach
leads to an issue of choosing block sizes. In this
algorithm the size of blocks of pixels is an adjustable
parameter. In order to investigate the effect of block size
on the segmented image, simulation studies are carried
out on real images. A block of pixels is denoted as a qq
matrix. Three different block sizes have been investigated.
We dont want to choose a block size too large since this
obviously entails crude classification. On the other hand,
if we choose a small block size, only very local properties
belonging to the small block are examined in
classification. The penalty then comes from losing
information about surrounding regions. A known method
in signal processing to attack this type of problem is to
use context information (Brejl M. and Sonka M. 2000).
Here the image is firstly analysed on account of the pixel
content and safe marginal limit of block size such that the
blocks shouldnt lose their unique knowledgeable
information, hence in this way both the block sizes and
classification rules can vary according to context.
The next step towards the goal is to detect out the features
of block. Here, two methods are analysed and concluded
wether to use standard deviation or the mean operation on
pixels intensities of individual blocks. The individual
main features of these methods are that even after
operation their effect on pixel remains knowledgeable,
even in the worst situations. Secondly they give the
sufficient information about the region of object,
occupying the image space including the safe limits of
object border issue. The standard deviation of pixel
intensities () in an image block is a feasible feature for
classification of scanned images. It can be taken as a
primary feature for identifying background blocks; is
very small for background images. As cannot be used
to effectively discriminate pictures from text in postal
images. The related features variance and absolute
deviation of pixel intensities have also been mentioned as
classification features and should perform similarly to .
The mean of pixel intensities () in an image block is a
useful feature for classification, since it gives the
knowledge of average variation of pixel intensities. The
three classes, background, picture, and text, can be
separated into three bands based on . However, non-
white backgrounds and light pictures may cause
misclassification. The feature is a good supplemental
tool in classification, but should not be taken as primary
because of the variety of envelope background colors.
Visual inspection will be employed to evaluate the
goodness of segmented image. A good segmented image
is one that retains as many details of image as possible
and viewer will still be able to recognize the object
contained in image.
III. EXPERIMENTS
Each image is split into blocks which are considerably
smaller than the size of the image (e.g. 22, 44 . . .
1616) with each non-overlapping block with its
neighbours in both the horizontal and vertical directions.
In some cases block overlapping is preferred where each
block is overlapping its neighbours by a configurable
amount of pixels (e.g., 1, 2 . . . 8) in both the horizontal
and vertical directions. Block overlapping can also be
interpreted as block advancement. The most obvious
difference between the results of images with different
block size is the increase in number of black dots as block
size decreases.
International Journal of Scientific Research Engineering & Technology (IJSRET)
Volume 2 Issue 10 pp 604-609 January 2014 www.ijsret.org ISSN 2278 0882
IJSRET @ 2014
Fig. 1 As an investigation to the performance of
proposed segmentation algorithm, the above is the
selected image to be segmented.
(Source www.mathworks.com)
A(1,1
)
A(2,1) A(3,1) A(M,1
)
A(1,2
)
A(2,2) A(3,2) A(M,2
)
A(1,3
)
A(2,3) A(3,3) A(M,3
)


A(1,
N)
A(2,N
)
A(3,N
)
A(M,
N)
Fig.2 Co-ordinates of pixels in digital image
Fig.3 Test Image after finding the standard deviation of
pixel intensities of blocks after block segmentation.
Fig.4 Test Image after scaling and before block
segmentation.
Fig.5 Test Image after scaling and block segmentation.
Fig.6 Test Image after scaling and block segmentation
for k = 2.
Fig.7 Test Image after scaling and block segmentation for
k = 5.
Fig.8 Test Image after scaling and block segmentation
for k = 10
International Journal of Scientific Research Engineering & Technology (IJSRET)
Volume 2 Issue 10 pp 604-609 January 2014 www.ijsret.org ISSN 2278 0882
IJSRET @ 2014
Fig.9 Test Image after scaling and block segmentation
for k = 20
Fig.10 Test Image after scaling and block segmentation
for k = 40
When block size is large the segmentation results
resemble the true scene more. However the segmented
image will have staircase like edges. As the block size
decreases, the edge becomes smoother. Visual
inspections will be employed to evaluate the goodness of
segmented images. A good segmented image is one that
retain as much details of image as possible and viewer
will still be able to recognize the object contained in
image. For a digital image with height of N pixels and a
width of M pixels, the spatial coordinates are arranged as
shown in figure. The co-ordinates of pixels are organized
such that the vertical position is indexed from top to
bottom in terms of n, n = 1, 2.. N. The horizontal
position is indeed from left to right in terms of m, m = 1,
2.M. A pixel in image is denoted as A(m, n). It is
noticed that q = 1 is a special case where one pixel is
processed at a time. To decide the dimensions of the
block we introduce a scalar k which is helpful in
deciding block size without overlapping the other blocks
of same image. It is the maximum row and column
dimension for the block. By changing the values of k
from k = 1 to k = 40, the images are observed and
compared with input image of fig.1. From above
experiments we conclude that for the minimum value of k
the image can be understood after finding standard
deviation. As we proceed further by increasing the value,
image becomes foggier.
IV. APPLICATION
A. Block Features
Choosing features is a critical issue in classification
because features often set the limits of classification
performance. The intra-block features are defined
according to the pixel intensities in a block. They aim at
describing the statistical properties of the block. Features
selected vary greatly for different applications. Widely
used examples include moments in the spatial domain or
frequency domain and coefficients of transformations,
e.g., the discrete cosine transform (DCT). The inter-block
features are defined to represent relations between two
blocks, for example, the difference between the average
intensities of the two blocks. The use of the inter-block
features is similar to that of delta and acceleration
coefficients in speech recognition, in which there is
ample empirical justification for the inclusion of these
features.
B. Image Segmentation of Documents
The second application of the algorithm is the
segmentation of document images into text and
photograph. Photograph refers to continuous-tone images
such as scanned pictures, and text refers to normal text,
tables, and artificial graphs generated by computer
software. We refer to the normal text as text for
simplicity if the meaning is clear from context. Images
experimented with are 8 bits/pixel gray-scale images.
This type of classification is useful in a printing process
for separately rendering different local image types. It is
also a tool for efficient extraction of data from image
databases.
C. Comparing the Blocks
A second area of interest is to extend the use of the
algorithm for block comparison among various regions to
detect the intensity variations by including texture
features. This would be useful for other applications.
Another area of interest is to explore the integration of
more powerful low-level segmentation algorithms.
V. CONCLUSIONS
In our proposed method the images were divided into
blocks, and standard deviation over the pixels of block
was operated. The reason to use standard deviation is to
minimize the intensity variations of pixels of same block
to an extent as minimum as possible. Pixel-based
processing approaches to foreground detection can be
susceptible to noise, illumination variations, and dynamic
backgrounds, partly due to not taking into account rich
contextual information. In contrast, region-based
approaches mitigated the effect of above phenomena but
suffered from blockiness artifacts. The proposed
detection method belonged to a region-based category,
but at the same time was able to segment smooth
contours of foreground objects.
International Journal of Scientific Research Engineering & Technology (IJSRET)
Volume 2 Issue 10 pp 604-609 January 2014 www.ijsret.org ISSN 2278 0882
IJSRET @ 2014
AKNOWLEDGEMENT
I gratefully acknowledge the unmatchable, helpful
knowledgeable kind comments of Prof.MohanAwasthi
for improving the clarity of the paper and crystal clearing
the work.
REFERENCES
[1] Jia Li, Amir N. and Robert M.G. Feb. 2000, Image
Classification by a Two-Dimensional Hidden Markov
Model., Fellow, IEEE, IEEE TRANSACTIONS ON
SIGNAL PROCESSING, VOL. 48, NO. 2.
[2] Vikas R., Conrad S. and Brian C.L. JANUARY 2013,
Improved Foreground Detection via Block-Based
Classifier Cascade With Probabilistic Decision
Integration, IEEE TRANSACTIONS ON CIRCUITS
AND SYSTEMS FOR VIDEO TECHNOLOGY,
VOL. 23, NO. 1.
[3] Wen Gao, Jie Y. Weiqiang W. MAY 2008, Modeling
Background and Segmenting Moving Objects from
Compressed Video. IEEE, IEEE TRANSACTIONS
ON CIRCUITS AND SYSTEMS FOR VIDEO
TECHNOLOGY, VOL. 18, NO. 5.
[4] RanjithUnnikrishnan, Caroline Pantofaru, Martial
Hebert, JUNE 2007, Toward Objective Evaluation of
Image Segmentation Algorithms, Member, IEEE,
IEEE TRANSACTIONS ON PATTERN
ANALYSIS AND MACHINE INTELLIGENCE,
VOL. 29, NO. 6.
[5] Lina Yi, Guifeng Zhang and Zhaocong Wu, 2010, A
Scale-Synthesis Method for High Spatial Resolution
Remote Sensing Image Segmentation, IEEE
TRANSACTIONS ON GEOSCIENCE AND
REMOTE SENSING.
[6] Michael E. Farmer and Anil K. Jain, Fellow, IEEE,
DECEMBER 2000, A Wrapper-Based Approach to
Image Segmentation and Classification IEEE
TRANSACTIONS ON IMAGE PROCESSING,
VOL. 14, NO. 12.
[7] Lei Zhang, Member and QiangJi, Senior Member,
IEEE, SEPTEMBER 2011, A Bayesian Network
Model for Automatic and Interactive Image
Segmentation, IEEE TRANSACTIONS ON IMAGE
PROCESSING, VOL. 20, NO. 9.
[8] Wenbing Tao, Hai Jin and Yimin Zhang, Senior
Member, IEEE, OCTOBER 2007, Color Image
Segmentation Based on Mean Shift and Normalized
Cuts, IEEE TRANSACTIONS ON SYSTEMS,
MAN, AND CYBERNETICSPART B:
CYBERNETICS, VOL. 37, NO. 5.
[9] Guo Dong and Ming Xie, Member, IEEE, JULY
2005, Color Clustering and Learning for Image
Segmentation Based on Neural Networks IEEE
TRANSACTIONS ON NEURAL NETWORKS,
VOL. 16, NO. 4.
[10] Peijun Li, JiancongGuo, Benqin Song, and Xiaobai
Xiao, MARCH 2011, A Multilevel Hierarchical
Image Segmentation Method for Urban Impervious
Surface Mapping Using Very High Resolution
Imagery, IEEE JOURNAL OF SELECTED TOPICS
IN APPLIED EARTH OBSERVATIONS AND
REMOTE SENSING, VOL. 4, NO. 1.
[11] MarekBrejl and Milan Sonka, Member, IEEE,
OCTOBER 2000, Object Localization and Border
Detection Criteria Design in Edge-Based Image
Segmentation: Automated Learning from Examples,
IEEE TRANSACTIONS ON MEDICAL IMAGING,
VOL. 19, NO. 10.
[12] K.Z.Mao, Peng Zhao and Puay H.T., JUNE 2006,
Supervised Learning-Based Cell Image Segmentation
for P53 Immunohistochemistry, IEEE
TRANSACTIONS ON BIOMEDICAL
ENGINEERING, VOL. 53, NO.6.
[13] Hao Ma, Jianke Zhu, Michael Rung-TsongLyu,
Fellow, IEEE, and Irwin King, Senior Member, IEEE,
AUGUST 2010, Bridging the Semantic Gap Between
Image Contents and Tags, IEEE TRANSACTIONS
ON MULTIMEDIA, VOL. 12, NO. 5.
[14] Ruth Bergman and Hila Nachlieli, JUNE 2011,
Perceptual Segmentation: Combining Image
Segmentation With Object Tagging, IEEE
TRANSACTIONS ON IMAGE PROCESSING,
VOL. 20, NO. 6.
[15] Xiaofeng Liu, Member and Jerry L. Prince, Fellow,
IEEE, AUGUST 2010, Shortest Path Refinement for
Motion Estimation From Tagged MR Images, IEEE
TRANSACTIONS ON MEDICAL IMAGING, VOL.
29, NO. 8.
[16] L. Vincent and P. Soille, Jun. 2009 Watersheds in
digital spaces: An efficient algorithm based on
immersion simulation, IEEE Trans. Pattern Anal.
Mach. Intell., vol. 13, no. 6, pp. 583597
[17] Ping-Han Lee, Yen-Liang Lin, Shen-Chi Chen, Chia-
Hsiang Wu, Cheng-Chih Tsai, and Yi-Ping Hung,
DECEMBER 2011, Viewpoint-Independent Object
Detection Based on Two-Dimensional Contours and
Three-Dimensional Sizes. IEEE TRANSACTIONS
ON INTELLIGENT TRANSPORTATION
SYSTEMS, VOL. 12, NO. 4.
[18] Wenjing Li, George Bebis, and Nikolaos G.
Bourbakis, Fellow, IEEE, , NOVEMBER 2008. D
Object Recognition Using 2-D Views. IEEE
TRANSACTIONS ON IMAGE PROCESSING,
VOL. 17, NO. 11.
[19] Andrew E. Johnson and Martial Hebert,
Member, IEEE, MAY 2005, Using Spin Images for
Efficient Object Recognition in Cluttered 3D Video
Scenes. IEEE TRANSACTIONS ON PATTERN
ANALYSIS AND MACHINE INTELLIGENCE,
VOL. 21, NO. 5.