Parallel image segmentation with adaptive mesh

Abstract

AbstractWe propose a parallel segmentation algorithm for general images which does not require advance knowledge of the image and where no initial assumptions about the number and position of regions are needed. This algorithm splits the image into a set of small triangular regions with uniform image properties, and the hierarchical construction is based on a parallel adaptive mesh algorithm. Based on a recursive binary traversal of discontinuities of image properties, hierarchical binary split and merge processing is performed. In this paper, a region is defined as space of the image where the features have smooth variation. By evaluating the homogeneity of the region using local differential properties representing the smoothness of the image, the problems of the starting point of processing and of order dependence are solved. Matching between blocks is ensured by sharing the small overlapping space that is needed to preserve the continuity of local properties. Further, it is possible to extract a region boundary line that matches the discontinuities of the local properties. Experiments performed on artificial intensity images, color real images, and range real images confirmed the universality and effectiveness of the proposed algorithm. © 2002 Wiley Periodicals, Inc. Syst Comp Jpn, 33(10): 95–104, 2002; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.1161