Region-based Markov random fields (MRF) is usually difficult to effectively describe the prior knowledge of complex natural images. To solve this problem, a local region consistency manifold constrained MRF(RCMC-MRF) model is proposed. Firstly, the proposed model uses low-dimensional manifold distribution of high-dimensional data to characterize complex geometry structure prior in local region of images, and builds a localized manifold prior constraints term for the image segmentation model. Then, the proposed model utilizes more local region information of images to construct a local spatial adaptive MRF based on the pairwise MRF. Finally, the complex geometry structure prior and local spatial adaptive statistical feature in the local region are incorporated according to the Bayesian theory. The Gibbs sample algorithm is used for optimization. Compared with the conventional region-based MRF model, experimental result shows that the proposed model can provide a better segmentation result. © 2019, Editorial Office of Control and Decision. All right reserved.
