Evaluation of texture features based on mutual information

Abstract

This article describes a study on features selection methods for classification purposes. A special attention is paid to method based on mutual information known from information theory. For experiments a set of 16 different homogeneous texture images from Brodatz album was selected. Texture features obtained based on mutual information technique was compared to those estimated using two techniques: Fisher coefficient and combined probability of classification error with average feature correlation respectively. Performed experiments shown advantage of features selected using mutual information based approach on texture classification. For additional evaluation of feature selection methods unexampled coefficient, based on classification results for every 1, 2, and 3 feature subsets is proposed. Based on this coefficient it is demonstrated that mutual information value indicates which feature is statistically better for classification. It is also possible to determine the optimal number of histogram bins for discretization of feature values.