Spatio-spectral dissimilarity algorithm for multiresolution texture segmentation

We present an integrated multichannel algorithm advantageously combining the power of three Gabor filtering procedures, focussed on salient global spectral peaks, multiscale spectral differences and multiresolution energy features, respectively. The approach includes a method specific design of tuned matched Gabor filters and spatio-spectral dissimilarity measures for analytical evaluation of each output channel´s information content. Based on these measures we propose a novel algorithm to select the most representative channel for each texture present in the image. After post-processing we achieve distinct segmentation along the boundaries. Our method operates in an essentially unsupervised mode and requires no a priori knowledge of the image. Results are given demonstrating the efficacy of our segmentation algorithm