Texture segmentation on two high-performance computers

An implementation of a texture segmentation algorithm on two high-performance computers, the Connection Machine CM-2 and the Convex mini-supercomputer, is presented. Texture segmentation is the process of identifying regions with similar texture and separating regions with different textures and is one of the early steps towards identifying surfaces and objects in an image. A segmentation algorithm is described which first extracts texture tokens from the input image, then computes the Voronoi tessellation of the extracted tokens and measures shape features (moments of area) of the resulting Voronoi polygons. Feature similarity is used to obtain an initial labeling of texture tokens as interior or border with four quantized directions. This labeling is then refrained using probabilistic relaxation labeling. The computation of the Voronoi tessellation and the probabilistic relaxation labeling process, which are highly data-parallel procedures, are discussed. Substantial speedups were obtained over a sequential (Sun-4/280) implementation.<
>