Texture classification using non-separable two-dimensional wavelets

In this paper, a new feature to the characterization of texture coarseness at multiple resolutions is proposed for texture classification. The feature is characterized by the extremum number of 2-D non-separable wavelet transforms (NSWT) estimated at the output of the corresponding filter bank. On a set of twelve Brodatz textures, the performances of texture classification based on pyramidal decomposition will be comparatively studied using the variance, entropy, extremum number and entropy of extremum as features, respectively. Experimental results show that the extremum number-based measure performs best among the features. In addition, we suggest that the wavelet coefficients of local extremum represent the original texture image instead of the entire wavelet coefficients. To explore the suitability of the NSWT for the texture characterization, the time varying, rotation invariant, and discriminatory characteristics are further investigated. It is shown that the textures have no time varying property and the NSWT is not rotation invariant to arbitrarily chosen degree of sample rotation. Finally, we show that the discriminatory characteristics of features do spread more in lower frequency subbands evaluated by a novel evaluation function based on genetic algorithms (GA).