SEGMENTATION OF RANDOM-FIELDS VIA BORROWED STRENGTH DENSITY-ESTIMATION

In many applications, spatial observations must be segmented into homo geneous regions and the number, positions, and shapes of the regions a re unknown a priori. information about the underlying probability dist ributions for observations in the various regions can be useful in suc h a procedure, but these distributions are often unknown. Furthermore, while there may be a large number of observations overall, the antici pated regions of interest maybe small with few observations from the i ndividual regions. This paper presents a technique designed to address these difficulties. A simple segmentation procedure can be obtained a s a clustering of the disjoint subregions obtained through an initial low-level partitioning procedure. Clustering of these subregions based upon a similarity matrix derived from estimates of their marginal pro bability density functions yields the resultant segmentation. It is sh own that this segmentation is improved through the use of a ''borrowed strength'' density estimation procedure wherein potential similaritie s between the density functions for the subregions are exploited. The borrowed strength technique is described and the performance of segmen tation based on these estimates is investigated through an example fro m statistical image analysis.