Image Segmentation Basedon Bayes Belief Propagation

Much important information can be obtained from SAR images that play an important role in military affairs and civilian activity. Therefore, the study for SAR images'processing has practical significance and application prospect. Texture images are common in practical applications and have important studying values. Segmentation is a very important task in the analysis and understanding of SAR image and texture image, so an efficient and accurate method for image segmentation is crucial. Some image segmentation methods for SAR images and texture images are studied in the dissertation, and validate them by experiment. The main contributions can be summarized as follows:1. An approach which combines nonsubsampled contourlet transform (NSCT) and MRF is proposed for texture images segmentation. The approach is supervised, which model the distribution of NSCT's high frequency coefficients by generalized Gaussian density and use the parameters of distribution as texture feature. MRF is used to model the spatial information of image, the segmentation problem is replaced by a MAP estimate and the estimation is realized by Bayes belief propagation to acquire the finial segmentation result.2. Approachs for image segmentation based MRF model are studied and two approachs are proposed for SAR segmentation, which is based MRF model and Bayes belief propagation. The first segmentation approach is based on regional Markov random field model and Bayes belief propagation, through the analysis of the existing method based MRF model for SAR images segmentation, introduce the two important definitions for association potential and interaction potential function, which is key for the segmentation method based regional MRF model and Bayes belief propagation. A detailed analysis of the new potential function'advantages and the whole process of the proposed approach are given lately. Hierarchical approachs for SAR images segmentation is studied, a hierarchical energy-based model is proposed firstly, which can precisely describe the local and global characteristics of image content at different hierarchies. And then hierarchical Bayes belief propagation is used to minimize the energy function and segment SAR images. Compared to the single hierarchical segmentation method, the method proposed in this dissertation can promote the information interaction in a large range and utilize the segmentation result in many hierarchies. From the segmentation results, a conclusion can be drawn that the two proposed methods are effective for SAR image segmentation.3. Three statistical models for image segmentation are studied, including finite mixture model, hidden Markov random field model (HMRF) and hidden Markov random measure field mode. After discuss the three models and point out the disadvantages of them, a new approach is proposed, which combines the advantages of HMRF model and Bayes belief propagation. The method provides a new good way for segmentation of SAR images. It is effective in parameters estimation, convergence rate, and segmentation quality. Supported by the National Natural Science Foundation of China under Grant Nos.60673097...