Texture descriptors based on co-occurrence matrices

Abstract

This paper focuses on the problem of texture classification using statistical descriptors based on the co-occurrence matrices. A major part of the paper is dedicated to the derivation of a general model for analysis and interpretation of experimental results in texture analysis when individual and groups of classifiers are being used, and a technique for evaluating their performance. Using six representative classifiers; that is, second angular moment f1, contrast f2, inverse difference moment f5, entropy f9, and information measures of correlation I and II, f12 and f13, we give a systematic study of the discrimination power of all 63 combination of these classifiers on 13 samples of Brodatz textures. The conclusion that can be drawn from our study is that it is useful to combine classifiers up to a certain order. Here it turned out that groups of four classifiers are optimal.