Image Denoising Via Prior Models

With the development of information technology,images become the most commonly used information media in human activities,and closely related to human life.Images captured in the real world are usually corrupted by noise due to the errors generated in the acquisition and transmission processing.Such noise corrupts important details of images and impairs their quality,even affect subsequent applications in computer vision.Thus,recovering a clean image from its noisy version is one of the most basic tasks in image processing and low-level computer vision.As image denoising is a highly ill-posed inverse problem,there are infinite solutions.To regularize the solution of image denoising,various priors of natural images have been exploited.According to their exploting priors,the existing image denoising methods can be divided into three categories: sparse prior based methods,low-rank prior based method and learning prior based method.To address the issues in these methods,we make the following main effort:(1)To solve the problem that over-dependence on parameters selection and tuning in the sparse representation,we study the nonparametric model and find that it still lacks exploiting image structure to improve the performance.Therefore we first propose the nonlocal structured beta process,then propose a sparse Bayesian dictionary learning framework with nonlocal structured beta process factor analysis and applicaiton to image denoising.Our proposed method is not dependent on parameters selection and tuning,especially our proposed method does not need to know noise variance in advance.Moreover,our method can infer the noise variance via Gibbs sampling and remove noise effectively.(2)To address the over-smoothing problem aroused by low rank model,we propose combine non-local similarity prior and local gradient prior.We introduce TGV to-low-rank approximation such that both of texture with regular patterns and irregular patterns can be preserved.A low-rank approximation with local structure preserving model is proposed.To optimize this model,we develop a split Bregman based optimization algorithm.A re-weighted strategy is developed not only to adaptively update the weight parameters of TGV regularization,but flexibly utilize local gradient information.Experimental results on widely used dataset demonstrate that the effectiveness of our proposed model and reveal its significant improvement over other state-of-the-arts.(3)To solve the problem that weighted model in mixed noise removal cannot estimate the weights robust,we propose a robust weights estimation based method.We assume that weights follow the Pareto distribution and infer the regularizer of weights.Then the low-rank approximation prior is adopted as the image prior.A regularization term is introduced to regularize the updating process of the weights and an adaptive strategy for parameter setting is developed,both of which make our estimated weights robust.The proposed model can be solved by alternatively optimization directly,which can estimate the weight matrix and remove mixed noise simultaneously.Experimental results on widely used dataset demonstrate the effectiveness of our proposed model.