Blind Image Deconvolution Based On Group Sparsity And Self-similarity

During the acquisition of digital images,the obtained images often suffer from a certain degree of blurriness due to many uncertain factors,which brings great inconvenience to the practical application of multimedia information.Therefore,how to recover images has become a crucial topic in the research field of images.According to whether the blur kernel of the image is known or not,the problem of image restoration can be divided into blind deconvolution and non-blind deconvolution.In this paper,we aim at the blind deconvolution of images,that is,estimating the blur kernel and clear image when only a blurred image is known.By focusing on group-based sparsity and structural self-similarity,the research on single image deblurring is carried out.The main achievements and innovations of this thesis are as follows:1.This paper proposes a blind deconvolution method based on group sparsity and structural self-similarity,which combines group sparsity and self-similarity as priors for regularization constraints.By fusing image self-similarity characteristics,this method introduces the similar image block group formed from similar image blocks in down-sampled images as basic unit of sparse representation.The blur kernel is iteratively estimated layer by layer in the image pyramid,and a regularization term is constructed to utilize the sparsity and multi-scale self-similarity features of images.Experimental results demonstrate that by using the properties of images,this method can well estimate blur kernels corresponding to clear images and bring good visual effects.2.This paper proposes a blind deconvolution method based on salient edges and group sparsity,by applying the regularization of group sparsity on the edge region of the image,extracting similar image blocks from the blurred image to form similar image block groups as the basic unit of sparse representation,and obtaining the blur kernel iteratively layer by layer in the image pyramid.This method takes blocks only from the edge region of the previous estimated sharp image as training samples for the dictionary of Group Sparse Representation in the current layer,which enhances edges of the image during the iterative solving process in the image pyramid and ensures theaccuracy of edge information during the blur kernel estimation.Experimental results demonstrate that this method has achieved good visual effect in the estimation of blur kernels and clear images and made more significant effect in the case of large blur.