Application of partial modeling techniques for texture segmentation

We describe an approach to texture segmentation that uses robust partial modeling procedures. First, a method for encoding (modeling) amplitude spectra of the texture images by means of sets of bivariate Gaussian functions is described. This procedure involves adaptive determination of a low-pass filter, clustering of the residual high-pass spectrum, and robust parametric encoding of separate spectral segments. Based on this model, a small set of Gabor filters tuned to the channels of high activity in the image Fourier spectrum is selected and applied to generate feature images for texture segmentation. The resulting feature images are segmented by an algorithm that uses the notion that homogeneous texture image regions can be defined as having unimodal texture feature histograms. This algorithm then applies a robust partial modeling technique to encode the feature image histograms as mixtures of univariate Gaussians. The estimated parameters of the univariate Gaussian functions that compose such mixtures are next used for segmenting feature images based on a maximum-likelihood decision rule. Several examples are presented to demonstrate the performance of the proposed approach to texture segmentation.