Ijetr011917

International Journal of Engineering and Technical Research (IJETR)
ISSN: 2321-0869, Volume-1, Issue-9, November 2013
6 www.erpublication.org

Abstract— This paper presents an unsupervised texture image
segmentation algorithm using clustering. Two criteria are
proposed in order to construct a feature space of reduced
dimensions for texture image segmentation, based on selected
Gabor filter set. An unsupervised clustering algorithm is applied
to the reduced feature space to obtain the number of clusters, i.e.
the number of texture regions [1]. A simple Euclidean distance
classification scheme is used to group the pixels into
corresponding texture regions. Experiments on a mixture of
mosaic of textures generated by random field model show the
proposed algorithm of using the clustering gives satisfactory
results in terms of the number of regions and region shapes. A
mathematical programming based clustering approach that is
applied to a digital platform of segmentation problem involving
demographic and transactional attributes related to the images.
It is the collection of sub images corresponding to different
image regions and scales is obtained. From the experiments, it is
found that the multistage sub image matching method is an
efficient way to achieve effective texture retrieval for image
segmentation.
Keywords- Image segmentation, Texture feature extraction,
sub-images matching, Clustering, Texture segmentation, Gabor
filter set.
I. INTRODUCTION
The multistage approach to extract colour features using
the Multistage Colour Coherence Vector (MCCV), manages
to handle the problem of sub-image colour retrieval
effectively. The idea of multistage feature extraction
technique has found much success in the area of colour
analysis. The multistage approach can be used to extract
texture information using any texture feature extraction
method, the Discrete Wavelet Frames (DWF). Wavelet-based
feature extractors are chosen because of their high
classification accuracy and low computational load [1].
In this fundamental process in digital image processing
which has found extensive application in areas such as image
retrieval, medical image processing, and remote sensing [2].
Texture is an essential key to the segmentation of image and
originates from the scale dependent spatial variability, and is
observed as the fluctuation of gray levels in images [3]. It is
well known that another source of the fluctuations exists in
images. Even though the human interpreter often is superior in
identifying strange details and phenomena in images, there is
still a need to automate this process.
Manuscript received November 01, 2013.
Mr.S.Raja is an Assistant Prof. of CSE, Panimalar Institute of Technolog
y, Chennai, TamilNadu, India.
Dr.S.Sankar is a Faculty in Department of EEE, Panimalar Institute of
Technology,Chennai.Tamil Nadu ,India.
Dr.S.Saravanakumar is a Prof. of IT, Panimalar Institute of Technology,
Chennai, TamilNadu, India.
The clustering leads to the concept of unsupervised
segmentation or classification of images. The notion of scale
gives an important hint to texture analysis. Wavelet packet
transform have received much attention as a promising tool
for texture analysis, because they have the ability to examine a
signal at different scales. The texture feature set derived from
the wavelet de-composition where the sampling between
wavelet levels is omitted is superior to that derived from the
standard wavelet decomposition [4], [5].The trend probably
therefore seems to be a concentration on the wavelet packet
transform which basically is the wavelet transform with sub
band decompositions not restricted to be dyadic. The discrete
wavelet packet transform are critically sampled multi rate
filter banks. However, critically sampled filter banks typically
imply inaccurate texture edge localization. Therefore, in this
paper, wavelet packet decomposition is employed to extract
local texture features from image.
From the sub-band filtering point of view, the difference
between the wavelet packet transform and the conventional
wavelet transform is that the former also recursively
decomposes the high frequency components, thus
constructing tree structured multi-band extension of the
wavelet transform wavelet packet transform decomposes an
image into sub images[6,7]. In the Fig.1 shows the standard
decomposition. H and L in Fig. 1 denote a high pass and low
pass filter respectively. The approximated image LL is
obtained by low pass filtering in both row and column
directions. The detail images, LH, HL, and HH, contain high
frequency components.
Fig.1 Decomposition of images into sub-images
Sub-image classification is done by indexing sub regions
and no specific objects. The image is partitioned in
overlapping rectangles of different sizes and this allows
dealing with the different scales of objects in the image. Each
of these regions is represented by global information and, for
efficiency reasons, instead of storing the whole global
features, only the differences of the feature vectors are stored.
The segmentation process is repeated at several different
resolutions or scales of the image. The result of this process is
a collection of sub-images each corresponding to different
regions of the parent image.
Investigation of Compressed Image Segmentation
S.Raja, S.Sankar, S.Saravanakumar

A. Sub-Image Querying
The sub-image querying problem can be defined as
follows: Given as input query a sub-image Q of an image I and
an image set S, retrieve from S those images Q’ in which
query Q appears (denoted Q<= Q’).The problem is made
more difficult than image retrieval by a wide variety of effects
(such as changing viewpoint, camera noise, etc.) that cause
the same object to appear different in different images . The
patterns belonging to each category will form a cluster in the
feature space which is compact and isolated from clusters
corresponding to other texture categories.
First stage: original image
Sub-image tilling
Fig.2 Sub-image querying
The above condition however might be suitable for
colour features, but might not work for texture features. By
non-linearly re-scaling the resolution of the input image, the
texture characteristics of the image might be affected, and
features extracted from it might not reflects an accurate
representation of the texture, hence resulting in incorrect
retrieval. Because of this, multiscale resolution is used from
the original size of the input image in the proposed algorithm.
Because of the possibly non-dyadic nature of the image size,
the algorithm ends up with a collection of slightly overlapping
(64×64) sub-images at the current resolution.
A suitable texture feature extraction can be performed
on all sub-images separately. Each time a feature vector of a
sub-image is produced, it is added to the final feature vector,
which represents the multiscale feature vector of the original
image. After all sub-images of the current scale have been
processed, the image is linearly rescaled and the whole
process is repeated.
B. Salient Location Identification
The wavelet decomposition is used because of its
computational simplicity. The decomposition algorithm is
based on the non-standard decomposition procedure.
Input image Salient Location Neighborhood
Identification Extraction
Sub-image
Descriptor Assignment
Fig.3 Neighborhood identification and extraction
The wavelet decomposition of the feature extraction
process from the sub-images are obtained from the
neighborhood region, and the descriptor vectors are
computed from each sub-image.
An input image yields three coefficient matrices, namely
the Horizontal (H ), Vertical (V ), Diagonal (D ), and the
compressed image to a quarter of the size of the original
image. The decomposition can be recursively performed by
using the output image from the previous iteration as the
input. At each decomposition level, wavelet coefficient
matrices are combined into one signal matrix S:
S (n, m) = | H (n, m)| +|V (n, m)|+|D (n, m) | (1)
The salient locations are identified by each element of the
last signal matrix (last wavelet decomposition) to the
untransformed input image. The localization approach
determines the location of a android when it travels near the
place that a reference image was taken as in (1). Hence, the
current query image is different from the corresponding
reference image by certain rotation angles. The neighborhood
region around each salient location is extracted to make it
rotationally invariant using the coordinates transformation in
equation (2). As a result, the sub-images of the same salient
point extracted from this method are having the same
appearance regardless of rotational deviations from an
original image.
Xn=Xf+ (u-M/2) cosaf (2)
Yn=Yf+ (v-M/2) sinaf
Where (u, v) are the transformed (new) coordinates,
(xn,yn) are untransformed coordinates in the original image,
and (fx, fy) is the coordinate of the salient point af=tan
-1 (
Xf / Yf
).
Classically, image segmentation is defined as the partitioning
of an image into overlapping, constituent regions which are
homogeneous with respect to some characteristic such as
intensity or texture. If the domain of the image is given by I,
then the segmentation problem is to determine the sets S<=I
whose union is the entire image I.
II. BACK PROPOGATION SYSTEM
In this back-propagation algorithm is a widely used learning
algorithm in Artificial Neural Networks. The Feed-Forward

International Journal of Engineering and Technical Research (IJETR)
ISSN: 2321-0869, Volume-1, Issue-9, November 2013
Neural Network architecture is capable of approximating
most problems with high accuracy and generalization ability.
This algorithm is based on the error-correction learning rule.
Error propagation consists of two passes through the different
layers of the network, a forward pass and a backward pass. In
the forward pass the input vector is applied to the sensory
nodes of the network and its effect propagates through the
network layer by layer. The actual response of the network is
subtracted from the desired response to produce an error
signal. This error signal is then propagated backward through
the network against the direction of synaptic conditions. The
synaptic weights are adjusted to make the actual response of
the network move closer to the desired response.
Fig.4. Input layer Hidden layers Output layer
The back propagation neural network is essentially a
network of simple processing elements working together to
produce a complex output. These elements or nodes are
arranged into different layers: input, middle and output. The
output from a back propagation neural network is computed
using a procedure known as the forward pass. The circuit of
propagation network is as shown in the Fig.4. The pictorial
representation of different steps as follows;
1.The input layer propagates a particular input vector’s
components to each node in the hidden layers.
2. Hidden layer nodes compute output values, which become
inputs to the nodes of the output layer.
3. The output layer nodes compute the network output for the
particular input vector. The forward pass produces an
output vector for a given input vector based on the current
state of the network weights.
4.Since the network weights are initialized to random
values, it is unlikely that reasonable outputs will result
before training. The weights are adjusted to reduce the error
by propagating the output error backward through the
network. This process is where the back propagation neural
network gets its name and is known as the backward pass:
5. Compute error values for each node in the output layer.
This can be computed because the desired output for each
node is known.
6. Compute the error for the middle layer nodes. This is done
by attributing a portion of the error at each output layer node
to the middle layer node, which feed that output node. The
amount of error due to each middle layer node depends on the
size of the weight assigned to the connection between the two
nodes.
7. Adjust the weight values to improve network performance
using the Delta rule.
8. Compute the overall error to test network performance.
The training set is repeatedly presented to the network
and the weight values are adjusted until the overall error is
below a predetermined tolerance. Since the Delta rule follows
the path of greatest decent along the error surface, local
minima can impede training. The momentum term
compensates for this problem to some degree.
Neighborhoods from the sub-band images and with the same
centre coordinates are represented. This process is repeated
for all unknown wavelet coefficients at the highest level. Once
all of the wavelet coefficients have been generated, the
synthesised image is inverse transformed to give an image that
should resemble that of the sample texture.
Note that, in order to avoid problems with boundary
conditions, it is necessary to pad each sub image with zeros
before performing the algorithm. This padding should be
removed prior to inverse transform.
Input image low poss 75 degree 45 degree
15 degree -15 degree -45 degree -75 degree
Fig.5 Synthesized image transformation
The unsupervised texture segmentation algorithm, in
terms of low computational load is considerable. The Fig.5 of
size m * n and the image to be synthesised Is of size M * N. Let
p 2 Is be a pixel to be synthesised and let w be the width of the
neighbourhood of the square spatial neighbourhood
surrounding p. For each pixel to be synthesised in algorithm
needs to search up to nm locations.
At each of those locations, 4w2 operations need to be
performed to calculate the weighted sum squared difference.
This is 4nmw2 operations in total for each searched region.
Therefore to generate Is of size M* N the algorithm will have
to perform 4NMnmw2
operations. In comparison, the
algorithm proposed in this paper synthesises texture at the
level of the complex wavelet transform.
At this level the dimensions of the sample image are
nm=16 and the image to be synthesised are NM=16. Since all
the sub images at this level must be searched, the total number
of operations is given as NM=16 /nm=16 /4w1
2
*6. Here w1 is
the neighbourhood size for this process and is typically
smaller than the needed.
The multi-resolution compositing technique is to turn
into an image which is a combination of an original image and
a synthetic texture derived from target; the combination is
controlled by a mask. The algorithm remains the same as
before, except that at each iteration of the algorithm, the
original is composite back into the noise image according to
the mask using that avoids blurring and aliasing.
For nonuniform textures, the texels within a region may
vary randomly in orientation and the position and shape of the
texels may be perturbed by the end of the algorithm, noise will
have been converted into a synthetic texture. First,

Match-Histogram is used to force the intensity distribution of
noise to match the intensity distribution of target. Then an
analysis is constructed from the target texture. The pyramid
representation can be chosen to capture features of various
sizes by using the features of various orientations. Then, the
noise are modified in an iterative manner. At each iterations, a
synthesis is constructed from noise, and Match-Histogram is
done on each of the sub-bands of the synthesis and analysis.
Fig.6 matching of sub-textures.
In the Fig. 6, shows that sub-texture field corresponding
to the concrete between the bricks. That show some of the
connected components have been used to generate this
sub-texture field left of Fig 6. The sub-texture fields are
synthesized in preprocess by using any existing texture
synthesis technique.
III.RESULTS AND DISCUSSIONS
In this section, the results obtained with the previously
described texture synthesis and sub image matching
technique. Fig.5 illustrates an example of synthesis result. The
top row shows from left to right: the input, the resulting cells
and the corresponding mesh. The second row shows the
arrangement map and the resulting synthesized mesh, that row
is the resulting texture for a low (left) and high (right) amount
of classes. Using a high amount of classes causes connected
components to be very small, which results in nearly no
computed sub textures. The left column represents the model,
the middle shows existing techniques from top to bottom:
texture quilting the feature matching synthesis technique of
pixel the right column the result shows. In all cases, the
pre-processing time required only once for a given texture is
below m*n this includes segmentation, texture-mesh
generation and sub texture field synthesis.
Simulations have been carried out using a database of
aquatic images. The proposed scheme for indexing the still
texture content is employed on the image database.
IV. CONCLUSIONS
This method of each textured regions are obtained.
Simulation results show that the proposed algorithm is
effective in texture image retrieval on the bases of texture
content. The segmentation and it allows efficient proposed
algorithm can be applied to large real-world image database
for efficient textured-based image indexing. In our
propagation algorithm, features of the segmented regions are
calculated as the mean of features of all blocks it contains,
features more efficient for indexing purpose can be found
later.
BIOGRAPHY
Mr.S.Raja completed his B.Tech degree in
Information Technology at Sri Balaji Chokalingam
Engineering College, Arni and M.Tech degree in
Computer Science and Engineering at Dr.M.G.R
University, Chennai. He has presented number of
papers in National and International Conferences.
His areas of interests are Computer Networks, Data
Mining, Java Programming & Database
Technologies. He is working as an Assistant
Professor in Department of Computer Science and
Engineering at Panimalar Institute of Technology,
Chennai.
Dr.S.Sankar obtained his B.E Degree in Electrical
& Electronics Engineering at Sri Venkateswara
College of Engineering, from Madras University
and M.E (Power System) Degree from Annamalai
University Chidambaram. He has done his Ph.D in
the area of FACTS controllers in 2011. His
research interests are in the area of FACTS,
Electrical Machines, Voltage stability, power
quality, Power system security and Power System
Analysis.
Dr S.SARAVANAKUMAR has more than 10
years of teaching and research experience. He did
his Postgraduate in ME in Computer Science and
Engineering at Bharath engineering college,anna
university chennai, and Ph.D in Computer Science
and Engineering at Bharath University, Chennai.
He has guiding a number of research scholars in the
area Adhoc Network, ANN, Security in Sensor
Networks, Mobile Database and Data Mining
under Bharath University Chennai, , Sathayabama
University and Bharathiyar University.
References
1. Li-Yi Wei and Marc Levoy. Fast texture synthesis using
tree-structured vector quantization. Proceedings of SIGGRAPH 2012,
pages 479–488, 2012.
2. Unser, M.: Texture Classification and Segmentation Using Wavelet
Frames. IEEE Trans. Image Processing. Vol. 4 (11) (2005), pp.
1549-1560.
3.Fukuda, S., Hirosawa, H.: A Wavelet-based Texture Feature Set
Applied to classification of Multifrequency Polarimetric SAR Images.
IEEE Trans. Geosci. Remote Sensing. Vol. 37 (2009), pp. 2282-2286
4.B. S. Manjunath and W. Y. Ma, "Texture Features for Browsing
and Retrieval of Image Data", IEEE .
5 W. Niblack, "The QBIC Project: Querying Images by Content
using Color, Texture and Shape", In SPIE
6.K. Jain and F. Farrokhnia, "Unsupervised Texture Segmentation
Using Gabor Filters", Pattern Recognition, Vol. 24, No. 12, 1991,
pp. 1167-1186.
7.Y. Rubner and C. Tomasi, “Texture Matrices". Proceedings of
IEEE International Conference on Systems, Man and Cybernetics,
San-Diego, USA, 1998, pp. 4601-4607.

Ijetr011917

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Ijetr011917 (20)

More from ER Publication.org (20)

Recently uploaded (20)

Ijetr011917