Texture feature extraction and classification by combining statistical and
neural based technique for efficient CBIR. Paper presented at
2012 Int. Conf. on MulGraB 2012, the 2012 Int. Conf. on BSBT 2012, and the
1st Int. Conf. on Intelligent Urban Computing, IUrC 2012, Held as Part of the
Future Generation Information Technology Conference, FGIT 2012 Vol. 353
CCIS, p. 106-113
The final publication is available at
link.springer.com and via doi
http://dx.doi.org/10.1007/978-3-642-35521-9_14
Copyright 2012 Springer-Verlag
COPYRIGHT NOTICE
UB ResearchOnline
http://researchonline.ballarat.edu.au
Texture Feature Extraction and Classification by
Combining Statistical and Neural based Technique for
Efficient CBIR
Siddhivinayak Kulkarni and Pradnya Kulkarni
School of Science, Information Technology and Engineering,
University of Ballarat,
Mount Helen, Victoria-3353, Australia
{S.Kulkarni, P.Kulkarni}@ballarat.edu.au
Abstract. This paper presents a technique based on statistical and neural feature
extractor, classifier and retrieval for real world texture images. The paper is
presented into two stages, texture image pre-processing includes downloading
images, normalizing into specific rows and columns, forming non-overlapping
windows and extracting statistical features. Co-occrance based statistical
technique is used for extracting four prominent texture features from an image.
Stage two includes, feeding of these parameters to Multi-Layer Perceptron
(MLP) as input and output. Hidden layer output was treated as characteristics of
the patterns and fed to classifier to classify into six different classes. Graphical
user interface was designed to pose a query of texture pattern and retrieval
results are shown.
Keywords: Texture images, Multi Layer Percptron, Classifier, image retrieval
1 Introduction
Image databases are becoming very popular due to the large amount of images that
are generated by various applications and due to the advancement in storage devices,
image compression, scanning, networking etc. Retrieving the specific images based
on their content has become an important research area for the last decade. These
images are retrieved based on their content such as global colour, texture, shape as
low level features.
An image database may contain thousands of textured images. The main problem
user that user is facing of locating the images having similar texture pattern in the
given query. More specifically, this problem is considered in two main parts: a.
finding the images having the similar texture given in query and b. specifying a
texture in query.
A good image retrieval system dealing with textures must provide solutions to both
problems. In posing a query in terms of texture, it is not realistic to expect the user to
draw a texture that he or she wants to retrieve. Therefore, all the textures which are
extracted from the database are classified into different clusters and to pose a query in
terms of these textures. While retrieving the texture patterns, those are similar to the
query, only that particular cluster has been considered. The similarity is calculated
based on the query pattern and all the texture images which belong to the same class.
This technique reduces the search only for that particular class and effectively reduces
the searching time. This similarity is calculated based on weighted Euclidean distance
and presented to the user. This is the effective way to express the texture query and
getting the result based on that particular query
In this paper, gray level single textured images are used to extract the texture
features and construct a feature vector by using co-occurrence matrix for each
textured image. These statistical based extracted features are used an input and output
to the Multi Layer Perceptron and characertics was taken from hidden layer, which is
then fed to classifier to classify these features into six different classes for efficient
retrieval. The results obtained are very promising and some of the results are
illustrated in this paper.
The rest of the paper is organised as follows: Section 2 gives the brief idea
regarding analysis of prominent texture feature extraction techniques, Section 3
details the proposed research methodology for extracting texture features based on
statistical-neural technique, Section 4 describes experimental results and analysis and
the paper is finally concluded in Section 5.
2 Analysis of Prominent Texture Feature Extraction Techniques
In texture feature representations such as pixel neighborhood [1], a simple texture
feature can be constructed by comparing suitable properties of current pixel with the
properties of neighboring pixels. But the disadvantage of this technique is that these
features are not very accurate as the feature vectors entirely depend upon the center
pixel. In tamura features [2], all the six texture properties are visually meaningful so
this texture representation becomes attractive in image retrieval. These properties of
texture are easy to recognize by human but elusive when to be described
quantitatively by a machine.
Markov random fields are attractive because they yield local and parsimonious
texture descriptions. But MRF model use individual pixels based measurements and
are hence not easily flexible to change in image resolution [3], [4], [5]. With SAR
models, there are major difficulties in selecting the size of the dependent pixel
neighborhood and the appropriate window size in which texture is regarded
homogenous. MRSAR model was developed to overcome this problem. But MRSAR
is computationally a very expensive set of features [6]. The wold decomposition
model avoids the actual decomposition of images and tolerates a variety of in-
homogeneities in natural data, making it suitable for use in large collections of natural
patterns [7]. The statistical properties such as mean and variance are extracted from
the wavelet sub-bands as texture representations. To explorer the middle band
characteristics, tree structured wavelet transform is used to improve the classification
accuracy. The wavelet transform when combined with other techniques such as
Kohonen map achieves better results. Gabor transform provides an attractive
approach [8], which is well suited to texture classification and database retrieval.
These are far superior compared to co-occurrence features and less sensitive to noise.
But there are possibilities for either mistreatment or adaptation to suit specific data.
This technique is not generally applicable for segmentation or image analysis.
3 Research Methodology
This section deals with the retrieval of texture images in detail. To retrieve texture
images, it is important to pre-process these images. This preprocessing includes the
formation of a texture image database, extraction of texture features from these
images, classifying these features in appropriate classes for retrieval. Research
methodology is divided broadly into two sections: Section-I Texture images pre-
processing and section-II Feature Extraction and Classification. Figure 1 shows the
block diagram of the proposed technique.
Fig. 1. Block diagram of proposed statistical-neural technique for texture image classification
3.1 Texture Image Database
Texture image database was created by downloading images from World Wide Web
and consists of 500 texture images. Each image is re-arranged as 512 x 512 pixels.
These images consist of textures of both statistical and structural natures. Structural
textures are considered to be consists of texture primitives which are repeated
systematically within the texture. In statistical texture usually no repetitive texture can
be identified. These texture images mainly contain the texture of brick wall, wood,
sky, grass, glass and fire. Each image is divided into 16 non-overlapping sub-images
each 128 128 pixels in size, thus creating a database of (500 16) 8000 texture
images.
3.2 Co-occurance Matrix
Gray Level Co-occurrence Matrix (GLCM) is one of the texture feature extraction
methods, estimates the image properties related to the second order statistics. It
contains the information about grey levels (intensities) of pixels and their neighbours,
at fixed distance and orientation. Each entry (i, j) in GLCM corresponds to the
number of occurrences of the pair of gray levels i and j which are a distance d in
direction θ apart in original image. Figure 2 shows the distances and orientations of
pixel p for co-occurrence matrix.
Fig. 2. Distances and directions of pixel p for co-occurrence matrix
In order to estimate the similarity between the different gray level co-occurrence
matrices, Haralick [9] proposed 14 statistical features extracted from them. To reduce
the computational complexity, only some of these features were selected. The
description of four most relevant features that are widely used in the literature [10]
[11] are given in Table 1.
Energy is a measure of textual uniformity of an image. Energy reaches its highest
value when grey level distribution has either a constant or a periodic form. A
homogenous image contains very few dominant grey tone transitions, and therefore
the P matrix for this image will have fewer entries of larger magnitude resulting in
larger value of energy feature. Also, energy feature have smaller value if P matrix
contains large number of smaller entries.
Table 1. Features extracted from grey level co-occurance matrix.
Entropy measures the disorder of an image and it achieves its largest value when
all the elements in P matrix are equal. When the image is not textually uniform many
GLCM elements have a very small value, which implies the entropy is very large.
Entropy is inversely proportional to GLCM energy.
Energy
i j
d
jiP ,
2
Entropy
jiPjiP
d
i j
d
,log,
Contrast
i j
d
jiPji ,
2
Inverse
Difference
Moment
i j
d
ji
ji
jiP
,
,
2
