Measuring Overlap-Rate for Cluster Merging in a Hierarchical Approach to Color Image Segmentation
Shengrui Wang, Haojun Sun · 2004
Cluster overlapping is a phenomenon not yet well understood by researchers. It is a key factor that influences the performance of a clustering algorithm. In color image segmentation, it is appropriate to study the overlap phenomenon with the hypothesis that cluster distributions can be described by mixtures of Gaussians. This paper presents a new theory regarding cluster overlap, which allows for the computation of a cluster overlap rate that turns out to be a very good measure of similarity between clusters. Using this measure, we develop a new hierarchical algorithm for image segmentation that partially solves the problem of determining the best number of clusters. Experimental results demonstrate the effectiveness of the new algorithm.