MODEL-FREE TEXTURE SEGMENTATION BASED ON DISTANCES BETWEEN FIRST-ORDER STATISTICS

This work describes several new image segmentation approaches, all bas ed on statistical edge detection. Although the involved statistical pa ttern recognition methods are well established, their adaptation to an edge-oriented segmentation is an extremely uncommon topic in the imag e processing literature. Thus, the subject of this paper is both tutor ial and innovative in nature, with respect to pattern recognition and to image segmentation, respectively. In this approach, the borders bet ween homogeneous regions are detected by evaluating a ''measure of div ersity'' between symmetrical and equal-sized subsets of the observatio n window. Each neighborhood is characterized by its complete local gra y value distribution. In case the characterization by the first-order statistics is not sufficient, information on the gray values' spatial relationships must be extracted from the second order statistics. The work focuses on the analysis of efficient methods for measuring a ''de gree of diversity'' following several approaches, namely: Distances be tween the vectors of the rank-ordered gray values: Minkowski-, Canberr a-, Tanimoto-distance, scalar product, composite distance (newly devel oped here). Distances between estimated gray value density functions: distances of Kolmogorov, Bhattacharyya, and Patrick-Fischer. The estim ation is performed by means of Parzen kernels. Measures of cluster sep aration between subwindows taken in the one-dimensional gray value spa ce or in the two-dimensional space obtained by also considering a spat ial distribution feature. Nonparametric Wilcoxon-type two-population t ests. Distances between the estimated locations and between the estima ted scales of symmetrical subwindows. Experimental results obtained on natural images with problematic textures (remote sensing, radar, ultr asound, nuclear medicine) and on synthetic images, with all the approa ches mentioned above, are shown and illustrated. (C) 1995 Academic Pre ss, Inc.