SUPERVISED SEGMENTATION USING A MULTIRESOLUTION DATA REPRESENTATION

A method for segmenting multiple feature images is presented. It builds on two well established methodologies: Bayes' decision rule and a multiresolution data representation. The proposed method is aimed at image analysis applications requiring routine processing. The goal is to improve the classification reliability of conventional statistical based pixel classifier by taking account of spatial contextual information conveyed in an image via multiresolution structure. Our method is closely related to the Spann and Wilson quadtree segmentation algorithm. However, we shall demonstrate that a supervised formulation under the assumption that image classes are normally distributed, leads to a significant simplification of Spann and Wilson's algorithm and gives more consistent segmentation results. In order to incorporate Bayes' decision rule with a multiresolution data structure, the class statistics at each resolution level must be available. However, we point out that unbiased estimate cannot be achieved by direct calculation. This leads to the development of an efficient method to acquire the parameters of class distributions at each resolution level. It involves estimating the class statistics on training sites at the full image resolution. The corresponding parameters at lower resolution are computed by predetermined scaling factors. The segmentation scheme is validated on synthetic data. To clarify the applicability of the proposed method a set of experiments of its use in texture image segmentation is illustrated in which Brodatz textures and seismic images are used.