Image Segmentation Using Graphical Models

Image segmentation is a very important technique in image processing. However, it is a very difficult task and there is no single unified approach for all types of images. This paper uses graphical models to design a segmentation algorithm and tests it on some nature images. First, the algorithm over-segments an image into small regions, called superpixels. For each superpixel, we model the pixels within it by a Markov random field (MRF). Then the parameters of each MRF are estimated. The coarse segmentation of the image is obtained by clustering these superpixels based on their MRF parameters. Then an undirected graphical model, in which each superpixel is a node, is used to add interactions among the superpixels. The result shows that the algorithm can generate decent segmentation for images with clear edge cues.