A study of region partitioning using reciprocal estimation of region models and pixel motion

This paper presents a novel framework for precise region partitioning and motion estimation. Content-based image processing like MPEG-4 has attracted much attention because it is useful for achieving high coding efficiency. Motion analysis is one of the most powerful region-partitioning methods, but in general situations, motion cannot be estimated accurately especially near the object boundary. In MPEG-4 applications like image databases, arbitrary-shaped object image sequences must have more precise boundaries than computer vision applications. We propose a new framework to generate region pieces, which well reflect the actual object boundary. It uses reciprocal maximum likelihood estimations of region shape and pixel motion. Region shape estimation is done using measured pixel motion and polynomial region models. Pixel motion estimation is done using observed region boundary information. The results of segmentation using the simple region-piece-grouping method seems to have a precise boundary and should be improved if we use more advanced grouping techniques like texture analysis. This framework is useful as a preprocessing method for several traditional segmentation techniques