A Study On Image Restoration Based On Group Sparse Representation

Image restoration refers to the degraded image processing to get clear,high quality images.It is one of the most popular topics in many fields,such as Images processing,computer vision and pattern recognition,etc.Image restoration is possessed of important position in the fields of studying astronomy,remote sensing imaging,military and medical treatment which mainly includes image de-noising and de-blurring,single image super-resolution and compressed sensing image reconstruction.The requirement is removing the noise,fuzzy and kept the image edges,details.Thus,a pair of contradictions was appeared.Sparse representation based image restoration is one of the representative applications of sparse coding theory.It assumes that an exemplar image patch can be sparsely represented by the adaptive sub-dictionary,and the sparse representation coefficient can be solved by the convex optimization technique.The final restoration is accomplished by reconstruct the image by image patches.It provides a new research perspective for the study of image restoration problem,and has attracted wide attention around the world.Derived from image restoration based on sparse representation,how to improve the accuracy and efficiency of existing algorithms was discussed,and the effective image restoration algorithms were proposed.The main research works were shown as follows:1.An image restoration algorithm based on non-locally centralized simultaneous sparse codding was proposed.In order to restore natural images effectively,the structured sparse coding noise was introduced to exploit spatial correlations and to explore the nonlocal constraint of the local structure.The nonlocal self-similarity constraint was incorporated in the proposed method.Different from the previous image restoration algorithms,the method takes a low-rank approach toward simultaneous sparse codding and provide a conceptually simple interpretation from a bilateral variance estimation perspective,namely that singular-value decomposition of similar packed patches can be viewed as pooling both local nonlocal information for estimating signal variances.An effective algorithm is then proposed using alternative direction method.The experimental results illustrated that the proposed algorithm outperformed the existing methods in term of accuracy and efficiency.2.In order to improve the capability and flexibility of low-rank approximation in dealing with image restoration,an image restoration based on non-convex low-rank approximation was proposed.A general low-rank optimization framework to simultaneously learn the adaptive redundant matrix and the sparse coefficients was studied.The novel framework extends basic low-rank approach to a non-convex relaxation.The Bayesian interpretation approach was provided to estimate an adaptive regularization parameter.Extensive experimental results on natural image demonstrate the improvement of the proposed method over recent image restoration methods.3.Aiming at improving the correlation among the groups of image patches,a tensor Robust principle component analysis based image restoration algorithm was proposed.Compared with the matrix representation of group that represents each image patch as a vector,this tensor-based image representation directly takes a image patch as a matrix slice in a tensor,which preserves the nonlocal and local structure of the image.Efficient algorithms using proximal gradient method and alternative linearizing method are developed to solve the proposed model.The experimental results exhibited that the proposed algorithm can achieve the more satisfactory robustness and efficiency that its competitors.