Markov Random Field Model-Based Image Segmentation Using Local And Global Methods

Markov Random Filed (MRF) theory had been widely used in the fields of computer vision and digital image processing. MRF provides a method which describes the correlated relation of images conveniently and directly.The proof of the equivalence of Gibbs distribution and MRF made rapid development of the application with MRF method.Image segmentation is a vital phase of image analysis.MRF-based segmentation method is based-on the MRF model. In this method,due to various uncertainties,an optimal solution is sought.A popular optimality criterion is the Maximum a Posteriori (MAP) probability principle in which both the prior distribution of the true image class and the conditional(likelihood) distribution of the data are taken into account.The MAP principle and MRF together form the MRF-MAP framework.In this paper,some useful and impotant MRF models are investigated in summary,such as Auto Model,Multilevel Logistic Model,Mutiscale Random Field Model,FRAME Model.More attention is paid to the local and global methods for solving the image segmentation problems based on the MRF-MAP framework.1.Among local method, ICM (Iterated Conditional Modes) algorithm,RL (Relaxation Labeling) algorithm,HCF (Highest Confidence First) algorithm are investigated and experiments show their respective characteristics.And also,an improved HCF algorithm is proposed.Through experiments on three groups of images,we conclude that ,The deterministic algorithms of ICM and HCF quickly converge to a local energy minimum but are dependent largely on the initial configuration,the result of the ICM algorithm sometimes is not continuous,and RL algorithm lacks the detailed regions' recognization ability.The improved HCF outperforms the ICM and RL.The performance arrangement of these algorithms are discussed in the end.2.The traditional global method, SA(Simulated annealing) algorithm's anlysis and its experiment result are presented.An improved SA algorithm using Dynamic Parameter Method to changE the coefficient parameter of the final energy.Improved SA algorithm modifies the parameter to speed up the convergence and shows good segmentation results for noisy images.Besides,the algorithms combining the annealing and the local optimal solutions are also investigated.3.We also conclude that,using the MRF-MAP framework to solve the image segmentation problem can not only achieve robust and good segmentation results but also can absorb other useful algorithms' thoughts and principles to extend and develop itself to satisfy specific needs for imge segmentation.Besides,some realistic problems and disadvantages are discussed.