Image segmentation via multi-scaled belief propagation

Image segmentation plays an important role in computer vision and image analysis. In this paper, we develop a novel algorithm which can automatically segment an image into regions with relative uniform texture or color without the need to decide the region number in advance. In this work, the segmentation is formulated as a labeling problem in the Markov random fields (MRFs) model. An efficient multi-scale belief propagation (BP) algorithm is used to find the solution to the MRF estimation. Extensive experiments have shown the effectiveness of our approach.