Image segmentation based on a multiresolution Bayesian framework

A multiresolution Markov random field (MMRF) approach to texture segmentation which makes use of both regional and boundary information is proposed in this work. The algorithm is designed to improve the efficiency and solve two problems: over-segmentation and under-segmentation. At each resolution level, regional information and boundary information between textured regions is extracted to describe each block using the multiresolution Fourier transform (MFT). The image is then modelled as a sequence of Markov random fields (MRFs) of varying resolution and a Gibbs sampling scheme is applied to label the constituent sites at each scale by seeking the configuration with minimal interaction energy. The interaction energy is defined as a function of the regional and boundary information. The constraints of region context, and boundary smoothness and connectivity are also encoded in the interaction energy. Once the algorithm converges at a given scale, the segmentation result is propagated down to the next resolution for further refinement.