ANALYSIS OF MULTICHANNEL NARROW-BAND-FILTERS FOR IMAGE TEXTURE SEGMENTATION

This paper studies the increasingly popular class of numerical texture analysis/segmentation paradigms that decompose images into multiple oriented spatial frequency channels. Optimality properties for effective texture segregation filters are considered first using idealized assumptions on the image textures. The functional uncertainty of the channel filters is shown to define a tradeoff between spectral selectivity and accuracy in boundary localization which is optimized by the 2-D Gabor functions. Next, a simple model of textured image formation is used to support the use of logarithmic (homomorphic) texture processing. The idealized texture model is then relaxed to enable the analysis of textural perturbations interpreted as localized amplitude (contrast) and phase variations. It is shown that the effects of these perturbations can be effectively ameliorated through the use of postdetection smoothing. Finally, space-variant textures and complex texture aggregates not generically well analyzed by any current technique are studied. While any texture analysis procedure making use of spatiospectral measurements can encounter profound difficulties in computing quantifiable and useful texture attributes from complex textures, under certain conditions local frequency estimation approaches can be used to analyze space-variant textures, and multiple-component hypotheses can be defined that identify sets of suprathreshold components in a texture.