Multiprocessor Implementation of a Texture Segmentation Scheme for Satellite Radar Images

This paper presents a scheme for segmenting images on the basis of differences in localised measures of spatial texture. The scheme used was originally proposed by Wilson and Spann [1] but incorporates a new clustering algorithm which gives improved overall segmentation performance. The Wilson and Spann [1] algorithm uses a clustering algorithm which proved susceptible to initial input parameters and gave poor segmentation on our images. Our algorithm uses a modification of the Koontz, Narendra and Fukunaga [2] clustering algorithm. By linking the clustering to the resolution of the image, significant clusters were able to be realised, yielding a more robust segmentation scheme. The adaptation also resulted in a significant reduction in run-time. The paper is directed towards the problem of segmenting satellite synthetic aperture radar (SAR) images and we give comparisons of the techniques on SAR and other images.