| In order to process and analyze images effectively, many people give the enormous attentions to statistical image modeling technology all the time. In the image segmentation domain, the statistical image modeling catches the images' key information through some simple mathematical models, and achieves good application effect; its correlative technique's research becomes the hot spot in the domestic and foreign scholars'study. In this dissertation, statistical image modeling on spatial domain and the wavelet domain is studied in terms of the image segmentation application, and some new models and image segmentation methods are proposed. The main contents of the thesis can be summarized as follows:(1) An Enhanced Hidden Markov Model-Hidden Markov Tree (EHMM-HMT) model and an image Segmentation method based on EHMM-HMT are proposed. In order to improve the HMM'descriptive power to image macrostructure, the EHMM model is designed on the spatial domain. Then, an EHMM-HMT model is proposed by modeling the dependency of the inter-image blocks and the characters of the intra-image block through EHMM and HMT models respectively. The experimental results show that EHMM-HMT model improves the model's description of the image macroscopic characters.(2) An Enhanced Hidden Markov Model-Hidden Markov Tree-3S (EHMM-HMT3S) model and an image Segmentation method based on EHMM-HMT3S are proposed. In order to further enhance EHMM-HMT's modeling precision to the characters of intra-block, we analyze the descriptive power of the HMT and the HMT-3S model, find out the HMT-3S model can capture the subbands statistical dependence of the Discrete Wavelet Transform (DWT) compared to the HMT model. An EHMM-HMT3S model is developed through combining EHMM model with HMT-3S, the experimental results show that EHMM-HMT3S provides the better results for image segmentation.(3) In order to improve the EHMM-HMT model's boundary detection ability and decrease its computational complexity, a multi-scale image segmentation method based on EHMM-HMT and Multi-States Weighted Hidden Markov Tree (MSWHMT) models is proposed. For the deficiencies of higher computational complexity and weak border perceiving of the image segmentation method based on EHMM-HMT model, a multi-scale image segmentation method based on MSWHMT model is designed. The MSWHMT describes the texture image's statistical relations of the different texture substructures in emphatically, and discards description of spatial structure relations between the texture substructures, integrates the macroscopic estimate with microcosmic description to the texture substructure. The MSWHMT strengthens model's distinction ability to the different textures compared to HMT, eliminates the interaction between the blocks and enhances its boundary sensation ability compared to EHMM-HMT. The experimental results show that the newly proposed MSWHMT reach the goal with reducing texture misclassification on the region boundaries on the finer scales in the raw segmentation and decrease the model's computational complexity.(4) Multi-scale Bayesian fusing strategy combining with the boundaries is proposed. There are two reasons for the proposed fusing strategy. One is that raw segmentation results on the coarsest scale based on the EHMM model are reliable on the regions, the other is that raw segmentation results on the finer scales based on the MSWHMT model are accurate on the boundary regions. A new boundary is defined through the raw segmentation results and the likelihoods in the process of fusion, and the boundary nodes and the non-boundary nodes are treated differently. The experimental results show that the newly proposed fusing strategy integrates the region uniformity with the boundary maintaining and gets good fusion results.This dissertation is supported by the National Natural Science Foundation of China under Grant No.60673097, and the National High Technology Development 863 Program of China under Grant No.2007AA12Z136. |
