Texture analysis with a texture matched M-channel wavelet approach

Abstract

A novel method of texture analysis based on the wavelet transform is described. An M-channel extension of the existing two-channel biorthogonal wavelets is proposed. The extension offers a compact and efficient decomposition, a higher degree of freedom in the design of the filter coefficients, and the facility of an iterative linear solution. In contra-distinction to the classical, purely mathematical design procedures of wavelet filters, the proposed design takes into account texture-relevant features. Texture-matched, asymmetric separable 2-D FIR (finite impulse response) filters are obtained which permit the decomposition of the image into texture-feature-dependent pyramid structures downsampled by a factor of M for each direction. The performance of the new filters is tested with the Brodatz textures and compared with the results of other wavelet approaches.