Study On Image Segmentation Based On Markov Random Field

Image segmentation is one of key issues in computer vision. The image segmentation based on MRF model has received much appreciation, such as the ability to make use of prior knowledge, the ability to generate connected boundary, less parameter, can easily combine with other segmentation method. So the algorithm has widely used in the image segmentation field.Image segmentation algorithm based on MRF is studied in this thesis, and especially two problems are discussed, that is the parameter estimation in MRF and the solving of the MAP problem in MRF.Firstly, the MAP problem in MRF is discussed. A new SA algorithm based on vibrant points is presented to increase the speed of the traditional SA algorithm in solving the MAP problem. After the pre-segmentation of the image, image pixels are divided into two classes: the stable points and the vibrant points. The vibrant points are stored by a linked list. Only the vibrant points are dealt with in each iteration to reduce computation load. And then, the stop rule of the SA algorithm is also improved to avoid the computation of global energy. The experiment results indicate that the improved SA algorithm based on vibrant points can greatly improve the computational efficiency while maintaining the segmentation effect.Secondly, the parameter estimation method in MRF is studied. Two traditional parameter estimation algorithms are introduced: the sample training algorithm and the EM algorithm, and then numerical experiments are done to compare with the two algorithms. Afterward, a novel estimate method based on quad-tree decomposing is proposed to estimate the isomorphic coefficients in inhomogeneous markov random field. The experiment results show that the isomorphic coefficient estimated by this algorithm can significantly improve the effect of the segmentation and the adaptability of the image segmentation algorithm.