Fast And Robust Nonparametric Blind Image Restoration

Image has been widely used in all areas of modern society,and because of the limitations of the imaging equipment and the influence of external environment to imaging,degraded images are usually obtained with information loss to a great extent.Among them,the image degradation caused by blurring is one of the most basic and most fundamental research problems in image restoration.Since the image blurring process is a non-inversible process,it is usually necessary to obtain a clear image by some image prior or regularization model.In this thesis,three MAP-based non-parametric blind deblurring methods are proposed using different kinds of developed image priors.The main research results are provided as follows:Firstly,in the blind deblurring task the unnatural L0 sparse image prior usually causes staircase artifacts in the flat regions which natually leads to less accurate blur kernel estimation.Considering the block method of 3-dimension(BM3D)has good denoising and smoothing performance,this thesis proposes a new blind image deblurring model based on fused L0 and BM3 D.The new model can improve the smoothness of the edge and flat area by using the non-local similarity of the image while extracting salient edge information of the image,which thus helps a lot to improve the kernel accuracy.The experimental results on benchmark dataset demonstrate that the proposed method can effectively generate more accurate blur-kernels than the L0-based method,thus obtaining sharper images.Secondly,unlike the natural image prior,the normalized sparsity image prior is claimed help to produce sharp images rather than blurred ones.However,the normalized sparsity prior itsefl has problems of kernel estimation inaccuracy,high computational cost as well as parameter sensitivity.To deal with these problems,this thesis presents another blind deblurring method based on the so-called generalized normalized sparsity prior,and derive a fast numerical algorithm by use of the ADMM.Under the same conditions,the experimental results show that the presented new method not only improves the accuracy of estimated blur kernels but also improves the computational efficiency.Besides,the robustness to parameter choice is achieved to some extent by the new method.Finally,this thesis combines the generalized total variation and generalized normalized sparsity prior model,advocating a new enhanced normalized sparsity prior for blind image deblurring.The generalized total variation provides an adaptive mechanism for the flexible processing of edge information and flat region.By use of high order gradient information in the image,the staircase artifacts caused by the normalized sparsity prior can be greatly suppressed by the new TGV-based blind deblurring model,and hence more visual pleasant images can be recovered.With experiments on both benchmark dataset and real-world color blurred images,the new TGV-based boosted normalized sparsity prior not only has achieved much better blind deblurring visual performance,but also manefisted a great more robustnes to the size or the complexity of blur kernels.