Relevant Feature Subset Selection from Ensemble of Multiple Feature Extraction Methods for Texture Classification

Relevant Feature Subset Selection from Ensemble of Multiple Feature Extraction Methods for Texture Classification

Bharti Rana (School of Computer and Systems Sciences, Jawaharlal Nehru University, New Delhi, India), Akanksha Juneja (School of Computer and Systems Sciences, Jawaharlal Nehru University, New Delhi, India) and Ramesh Kumar Agrawal (School of Computer and Systems Sciences, Jawaharlal Nehru University, New Delhi, India)
Copyright: © 2015 |Pages: 18
DOI: 10.4018/IJCVIP.2015010103
OnDemand PDF Download:
$37.50

Abstract

Performance of texture classification for a given set of texture patterns depends on the choice of feature extraction technique. Integration of features from various feature extraction methods not only eliminates risk of method selection but also brings benefits from the participating methods which play complimentary role among themselves to represent underlying texture pattern. However, it comes at the cost of a large feature vector which may contain redundant features. The presence of such redundant features leads to high computation time, memory requirement and may deteriorate the performance of the classifier. In this research workMonirst phase, a pool of texture features is constructed by integrating features from seven well known feature extraction methods. In the second phase, a few popular feature subset selection techniques are investigated to determine a minimal subset of relevant features from this pool of features. In order to check the efficacy of the proposed approach, performance is evaluated on publically available Brodatz dataset, in terms of classification error. Experimental results demonstrate substantial improvement in classification performance over existing feature extraction techniques. Furthermore, ranking and statistical test also strengthen the results.
Article Preview

1. Introduction

Texture classification is a process of determining the class of texture, from a set of known texture classes, to which a given image or sub-image belongs. Texture is an inherent property of an image and plays a vital role in distinguishing images. In the last few decades, remarkable progress has been made in texture classification but it is still considered as a challenging problem. Wide range of real-time applications of texture classification such as automated detection of defects and quality control of texture images, medical diagnosis, microscope images, postal address recognition and interpretation of maps, remote sensing and geological images has motivated the research community to develop robust and efficient techniques.

The performance of texture classification mainly depends on the choice of: (i) features to represent an image and (ii) classifier to learn a decision model. In literature, numerous feature extraction techniques have been proposed, which can be classified into four categories (Tuceryan & Jain, 1998; Materka & Strzelecki, 1998; Zhang & Tan, 2002): structural, statistical, model based and signal processing based techniques.

Structural techniques represent textures as being composed of simple primitive structures called “texels” (or textons or texture elements) and placement rules that govern their spatial arrangement. They provide a good symbolic description of an image and are most effective for representing regular textures. But extraction of texels is a complex process (Zhang & Tan, 2002). On the other hand, statistical techniques characterize the texture in terms of statistical properties that measures correlation among the grey levels of an image (Materka & Strzelecki, 1998). However, they do not explicitly consider the hierarchical structure of the texture. Moreover, the complexity of computing higher order statistical features increases with the increase in number of grey levels. Model-based techniques, construct a model of an image that can be used to describe as well as synthesize the texture (Zhang & Tan, 2002). The parameters of the model describe the basic texture properties. However, parameter estimation and the choice of suitable model are two basic problems in these techniques. In signal processing based techniques, features are constructed by convolving images with spatial or (and) frequency domain filters. These techniques perform well with both random and regular textures. But a key problem of these techniques is to choose or design appropriate filters.

Complete Article List

Search this Journal:
Reset
Open Access Articles: Forthcoming
Volume 7: 4 Issues (2017): 3 Released, 1 Forthcoming
Volume 6: 2 Issues (2016)
Volume 5: 2 Issues (2015)
Volume 4: 2 Issues (2014)
Volume 3: 4 Issues (2013)
Volume 2: 4 Issues (2012)
Volume 1: 4 Issues (2011)
View Complete Journal Contents Listing