Image segmentation by a multiresolution approach

Segmentation is the primary step to vision systems. It allows the generation of a compact description of image (edges, regions), but is often one of the most critical processing for subsequent treatments and especially for the matching process. This paper deals with extracting regions (segmentation) which can be used as primitives to facilitate the matching process in stereovision. In the literature, numerous segmentation techniques have been proposed, and they can be divided into three categories: classification methods, edge detection methods and region growing methods. Our technique consists in representing the initial image in a pyramidal form and then applying the region growing process at various resolutions. The linked pyramid is a very flexible data structure representing the image at different resolution levels (multi-resolution). The originality of our method resides on the use of region merging technique with the multi-resolution process. Experimental results for real stereo pair images are presented.