Based On Markov Random Field Image Segmentation

Theories of Markov Random Field (MRF) have been used in computer vision andimage processing widely, as a kind of image processing theory flame combining withBaeyes theory. MRF can provide restrictions in image processing field with more priorinformation, and provide some convenient and direct methods to describe spatialcorrelations of pixels with probability. Applications of MRF in the field of imageprocessing have been extended with confirmation of the equivalence between MRF andGibbs distribution.MRF theory is applied to two representative questions in this paper.1. Binary image de-noising is widely used in many fields such as finger imagepreprocessing, text image restoration. With the flexible cliques and effective priormodels, Gibbs Random Field (GRF) has gained more and more attentions in imageprocessing. However, in those GRF-based image de-noising algorithms, Gibbsdistribution binary potential clique parameter,β, can't be changed adaptively withdifferent area features when we adopt fuzzy Gibbs random field for imagede-noising. The article shows an adaptive algorithm to alter the value ofβ. Theapproach can automatically decreaseβto keep details near the object edges andincreaseβto suppress noises in smooth areas. Based on several simulation cases, theproposed adaptive algorithm is compared with the standard GRF algorithm, and theresults show that the new algorithm behaves better in identifying and resolvingcapability.2. Image segmentation is an important approach in medical imaging. The methodbased on MRF takes the parameter of Gibbs as a representation of the conjunctionof different pixels. So it is not affected by the noises. The paper provided a newmethod named Restructuring Elitist Strategy Genetic Algorithm (REGA) which isdifferent from some traditional methods, such as Iterated Conditional Mode (ICM),Simulated Annealing (SA) and Genetic Algorithm (GA). This algorithm designs ahybrid GA based on MRF for image segmentation through restructuring elitistindividuals. A series of experiments proves this algorithm can provide better results than ICM, SA and GA in the conditions of same parameters and computing time.3. Enlightened by the information processing theory of visual perception, the authorpresented a hybrid pyramid model based on Gaussian-Markov random field inwavelets space (W-GMHP) according to the model in gray space. Comparing withprocessing effects only in gray space, the model can take better processing effectswith more information given by wavelet transforming in low frequency (LL),middle frequency (LH,HL) and high frequency (HH). Then, the author primarilydiscussed the similiarity between the W-GMHP model and the procedure of visualperception information processing.