A novel information theoretic approach to wavelet feature selection for texture classification

Abstract

In this research we address the problem of discriminant subband selection for texture classification. A novel Effective Information based Subband Selection (EISS) algorithm is proposed which utilizes the intra-class and inter-class distributions. Essentially these distributions are used to calculate the class-based entropy for a given subband. This class-based information is incorporated in the total information content of the training images to develop a robust Effective Information (EI) criterion. Only the subbands with the top EI criteria are allowed to participate in the classification process. The proposed EISS algorithm is evaluated on Brodatz texture database and has shown to outperform the most relevant method based on mutual information criterion.