Image Segmentation Based on High-Order MRF Model with Robust Local Spatial Information

Traditional low-order regional Markov Random Fields (MRF)model is difficult to accurately describe the global connectivity of complex natural images and often leads to the over-smoothing of the segmentation results. To solve this problem, a high-order MRF image segmentation model with robust local spatial information is proposed. Firstly, the proposed model introduces the local spatial relationship of the image by using the Hamming distance between the neighborhood pixels in the local region, then establishes a weighted Gaussian mixture likelihood feature between the label space and the pixel intensity field, which provides the local spatial consistency constraint; Secondly, the spatial global constraint relationship of the far away distance is introduced based on the Robust
$\mathrm{P}^{n}$
model, and the regional label consistency constraint of the MRF image segmentation model is established. Finally, based on Bayesian theory, a high-order MRF energy model with robust local spatial information for image segmentation is proposed, and the proposed model is optimized by Gibbs sampling algorithm. Compared with Traditional low-order regional MRF model, experimental result shows that the proposed model can provide a better segmentation.