Graph-theoretic algorithms for image segmentation

Image segmentation partitions a digital image into disjoint regions, each region is homogeneous, while adjacent regions are not. A variety of methods have been used to perform segmentation, but only a few utilize graph theory. We introduce a new approximation method for partitioning based on cutsets. During domain-dependent feature analysis, a complete, weighted graph, K,, is produced. Nodes correspond to pixels or groups of pixels, and edge weights measure the similarity between nodes. Partitioning seeks to minimize the inter-segment and maximize the intra-segment similarity. Given such a weighted graph, our method determines a maximal spanning tree. Of the 2^n-1 possible partitions, only those fundamental cutsets corresponding to the edges in a spanning tree are evaluated. Our implementations include adaptation of three similarity measures using this approach. The effectiveness of the three similarity measures on a number of actual images is demonstrated