The Application Research Of Wavelet Transform And Markov Random Field To Image Denoising And Segmentation

Image corrupted by noise will have a disadvantageous effect on high level image processing and it is necessary for us to carry on the research of image denoising. Image segmentation is a main task in image processing, but to produce an accurate result of image segmentation is not an easy matter for us, thus carrying on image segmentation research is essential. The primary contents are as following:1 , A new image denoising method based on the wavelet coefficient modulus maxima and Neyman-Pearson principle threshold is presented, which overcomes the dilemma of image denoising and edge preserving. Firstly, the edge detection property of wavelet transform is analysed and a theorem that the wavelet transform of adding Gaussian noise still obeys Gaussian distribution is put forward. Then the image denoising threshold is determined according to the noise distribution and Neyman-Pearson principle, the initial estimated wavelet coefficients are determined by the threshold, and the expected wavelet coefficients are got by the estimated coefficients. Based on the assumption that the observed image is the sum of the expected image and white noise, the qualitative and quantitative performance of the proposed denoising method is compared with others. Simulation results show that the proposed method can efficiently denoise, such as lowing mean square error, increasing signal-to-noise ratio and the correlative coefficients, while preserving the details of the original image.2, An image denoising method based on wavelet domain spatial context modeling is put forward. This method can capture the spatial clustering property of the intrascale wavelet coefficients and the non-gaussian distribution and persistence of the interscale wavelet coefficients. Firstly, the deficiencies of wavelet domain hidden markov model (HMM) are analyzed, the wavelet coefficient significant measures are determined according to the likelihood ratio of two state distrbution of coefficients, then the wavelet coefficients label is determined according to the modulus of wavelet coefficients. Finally, the expected wavelet coefficients are given according to the local spatial neighborhood wavelet coefficient labels and the initial shrinkage factor got by HMM. Simulation results show that the proposed denoising method can efficiently capture the clustering property of wavelet coefficients, while increases peak signal-to-noise ratio and the correlative coefficients.3> After analyzing the shortcomings of spatial markov model, a new image segmentation method based on hierarchical MRF model is proposed. In order to accurately describe the region, the low-level character field distribution of segmentated image is modeled by finite general mixture model (FGMM) and the finite Gaussian mixture model is only one of the FGMM. For high-level label image, the interior region is modeled by Gibbs distribution and the boundary is modeled by anisotropic MRF. An improved expectation maximization algorithm is used to estimate the parameters of gray field. The local approximation idea is used to determine label field model parameters. The label is determined by the minimum descriptive length principle of information theory. According to the posterior distribution of label image conditioned on the gray image corresponding to the conditional probability of FGMM-MRF model, the Bayesian formulation and the local iterated conditional modes optimization algorithm are adopted, and based on the maximum a posterior criterion the segmentation result is obtained. Numerical simulations demonstrate that the whole property and boundary of image areas show better vision effect with a test to synthetic image and real MR brain image.4> The hierarchical structure of image, the clustering property of wavelet coefficients and the deficiencies of spatial fuzzy cluster method (FCM) are analyzed. Combinding the wavelet transform hierarchical structure with FCM, a new image segmentation algorithm based on wavelet domain FCM is proposed. The comparision experiments between classical FCM segmentation algorithm and the wavelet domain FCM algorithm are carried out. Simulation results show that the proposed method is excellent. Based on the hierarchical structure of wavelet transform and the shortcoming of the independence between wavelet coefficient labels in hidden markov tree model, a new image hierarchical model segmentation algorithm is presented. Distribution of wavelet coefficients are described by Gaussian mixture model in character model. In order to embody the clustering of wavelet coefficients, the wavlelet coefficient label is determined by its father and brother wavelet coefficient labels. Using Bayesian principle, the corresponding segmentation causal algorithm is given. Numerical simulation with a test to the real image shows that the proposed algorithm is effective and excellent.