Research On Texture Images Segmentation Based On Markov Random Field

Image segmentation is the basis for many high level computer vision tasks,such as object identification,motion analysis and scene reconstruction.It also links low level vision system to high level vision system.Many images content a lot of texture information and texture image segmentation is widely used in remote sensing image analysis and medical image processing today.It is crucial to apply an appropriate model to describe the texture characteristics.Markov random field model(MRF)has a perfect theoretical foundation and good ability of space expression,which can describe complex texture information compatibly.So this paper mainly researches on textual images modeling and segmentation based on MRF model.The framework can be divided into three parts referring to different methods:Firstly,proposed a weighted image segmentation method based on Markov Random Field.On the basis of traditional potential function,the paper introduces the relationship of neighborhood pixels to describe the probability of pixels being partitioned into the same class.We convert image segmentation problem into maximum a posterior for extreme value by Bayesian theory.Meanwhile,we introduce the variable weight method to connect feature field and label field.Finally,the iterative conditional model is applied to acquire optimal segmentations.Compared with K-means and traditional MRF algorithm,the proposed method presents effectiveness and robustness in the experiments.Secondly,to solve the difficult problem of expectations caused by the interaction between hidden variables when using EM algorithm to estimate parameters for hierarchical MRF,the mean-field theory is introduced into Gaussian-Markov random field model(GMRF).Parameters can be estimated easily through a simple linear equation in case of without window function.An interactive potential function based on Bayesian belief propagation algorithm is proposed to change the situation that the fixed or variable weighted potential function cannot express the interaction of image regions.Experiments demonstrate that proposed method not only has good regional classification but also smoothly internal region.In addition,the mixed and confused phenomenon of traditional hierarchical MRF is improved in wavelet domain.Thirdly,proposed a hierarchical Gauss-Markov random field model in wavelet domain that was applied to the texture image segmentation.Multiscale random field model methods only consider inter-scale casual Markov random field model to describe local statistic information,which are not satisfactory as segmentation results.To address this deficiency,we used Gauss-Markov random field to modeling,meanwhile taking more intra-scale local spatial interaction information into consideration.We use supervised method for training image parameters and multi-objective solving technique to optimize sequential maximum a posteriori estimation.We demonstrate the performance of the proposed method with texture images and noised images.From the visual effect and classification correct ratio,the resulting WGMRF model are more accurate and robustness than the WMSRF method.