Image Segmentation Based on Multiscale Local Region Belief Propagation Algorithm

Image segmentation approaches based on the conventional multiresolution Markov random field (MRF) often produced blocky artifacts. To solve this problem, a new multiscale local region belief propagation (BP) algorithm was proposed. This algorithm based on MRF model built local region messages with different scales, then the messages were propagated on MRF, and segmentation results were finally estimated by local region probabilities based on maximum a posterior (MAP) criterion. This algorithm combined local region features with global features, and multiscale messages were propagated on the finest MRF, which avoided misclassified result propagating between levels on the conventional multiresolution MRF model. Therefore, the proposed algorithm obtained not only more accurate segmentation results but also faster speeds. Experimental results on a wide variety of images had verified the effectiveness of this algorithm.