Robust Classification of Texture Images using Distributional-based Multivariate Analysis

Classification of texture images has been recognized as an important task in the field of image analysis and computer vision through the last few decades. A plethora of research papers have appeared in the literature trying to cope with effective ways to extract faithful distributions that accurately represent the inner content and attributes of texture images. An issue of great importance is, also, the incorporation of a valid similarity measure that can successfully estimate how close these distributions are with respect to some pre-classified texture categories. The basic operations that need to be carried out in order to estimate the similarity between texture images and thereafter assess the classification problem are (a) choose an appropriate feature space for texture representation, (b) construct a theoretically valid distribution in the texture feature space, i.e. the texture signature, which provide a representation of the texture image in a multivariate feature space, (c) perform pairwise comparisons between corresponding texture signatures that constitute the consequent content distributions of the texture images and (d) choose an experimentally valid classifier for the subsequent evaluation. The scope of this chapter is the survey of a recently introduced methodology for distributional-based classification of texture images (Pothos et al., 2007), its enhancement via the incorporation of a self-organizing module and its adaptation so as to work in multivariate feature spaces. The original approach is based on an efficient strategy for analyzing texture patterns within a distributional framework and the use of a statistical distributional measure for comparing multivariate data, also known as the multivariate Wald-Wolfowitz test (WW-test) (Friedman & Rafsky, 1979). By combining the flexible character of the original methodology with the learning abilities of neural networks, we build a general-purpose platform for the efficient information management and classification of texture patches without any restriction regarding the exact image content. Here, we will first describe the enrollment of standard feature extraction techniques for summarizing texture information and structuring multivariate texture spaces. These techniques include wavelet analysis, discrete cosine transform (DCT), Gabor filters and edge histogram descriptor. The above methods have been considered as golden standards for extracting appropriate distributions from texture images. In the following stage, we will test the applicability of some multivariate distributional-based measures for estimating the