Automatic Image Semantic Segmentation by MRF with Transformation-Invariant Shape Priors

Shape priors has greatly enhanced low-level driven image segmentations, however existing graph cut based segmentation methods still restrict to pre-aligned shape priors. The major contribution of this paper is to incorporate transformation-invariant shape priors into the graph cut algorithm for automatic image segmentations. The expectation of shape transformation and image knowledge are encoded into energy functions that is optimized in a MRF maximum likelihood framework using the expectation-maximization. The iteratively updated expectation process improves the segmentation robustness. In turn, the maximum likelihood segmentation is realized integrally by casting the lower-bound of energy function in a graph structure that can be effectively optimized by graph-cuts algorithm in order to achieve a global solution and also increase the accuracy of the probabilities measurement. Finally, experimental results demonstrate the potentials of our method under conditions of noises, clutters, and incomplete occlusions.