Image Denoising Based On Statistical Models In The Wavelet Domain

Image denoising usually imposes a compromise between reducing noise and preserving significant image details. To achieve a good performance in this respect, the denoising methods based on statistical models in the wavelet domain are well studied in this thesis. The statistical models of wavelet coefficients are mainly divided into three categories: the statistical model of interscale dependencies, the statistical model of intrascale dependencies and the marginal probability model of wavelet coefficients. Based on these statistical models, two different classes of denoising methods are developed, which are Bayesian denoising methods based on hidden Markov tree (HMT) models and based on Markov random field (MRF) models. The latter are more easily extended, so this thesis is devoted to this class of denoising methods.Geometrical Bayesian method is a representative of Bayesian denoising methods based on MRF models, and it provides a framework in which statistical models of wavelet coefficients can be exploited. To improve the noise suppression performance, the statistical models of wavelet coefficients are studied and an improved geometrical Bayesian method called anisotropic MRF model based geometrical Bayesian (AMRFMGB) method is proposed. The statistical model of intrascale dependencies, the marginal probability model of wavelet coefficients and the statistical model of interscale dependencies are the core of AMRFMGB method. Based on the oriented clustering property of wavelet coefficients, an anisotropic MRF model is constructed, which can capture the intrascale dependencies. Based on the non-Gaussian property of the wavelet...