A Parallel, Non-parametric, Non-iteratrve Clustering Algorithm With Application To Image Segmentation

This paper describes a parallel, non-parametric, non-iterative clustering algorithm with application to segmentation of textured/ color/multispectral images. It groups multidimensional features extracted from the image into distinct clusters. Segmentation is done by mapping back these clusters into the spatial domain. The process starts with generation of a feature vector histogram in the multidimensional space. A parallel hillclimbing procedure is then utilized to identify local peaks which correspond to the center of clusters. Each peak must pass a cluster validity test or get merged with another nearby peak. Application of this technique to textured images is presented. The textural features extracted from local regions in the image are six estimated parameters of random field models fitted to the region and two other local statistics namely: sample mean and variance (total of eight features). The implementation of the algorithm on a SEQUENT S81 MIMD parallel processor is discussed.