Research On Solving Blur Kernel In Blind Image Deblurring

Image deblurring is an important issue in task of image restoration. It aims to recover the clear image from a blurred observation. According to the known blur kernel and the unknown blur kernel, the task can be classified as non- blind image deblurring and blind image deblurring. On the condition of unknown blur kernel, we not only need to estimate blur kernel but also need to solve for clear latent image. So, the problem reflect as a seriously ill-posed problem. Doubtless, the estimation of blur kernel play an important role in task of blind image deblurring. In this paper, we analyze the existed methods of non- blind image deblurring and blind image deblurring. Then, three different algorithms for blind image deblurring were proposed.The algorithms put more emphasis on blur kernel estimation with reasonable use of image prior knowledge.1. In our work, we explore a novel approach to estimate a robust blur kernel using low-rank approximation. With the obtained blur kernel, a more faithful latent image can be recovered. In first step, the proposed approach attempts to estimate two blur kernels for a blurry image rather than to estimate a blur kernel directly as previous methods do. Both of the two kernels are obtained by two modified edge-based methods which benefit from autoregressive model and heuristic image filters. In next step, through an advanced strategy of low rank approximation, a more accurate kernel can be obtained from the two obtained blur kernels. Finally, a popular non-blind deconvolution method is employed to reconstruct a clear image. Through this method, the helpful difference which lies in two estimated blur kernels can be combined for a more accurate and robust blur kernel. Meanwhile, noises and singular values existed in the process of solving blur kernel can be reduced. The experiment results demonstrate that the proposed algorithm can reduce the inaccuracy of blur kernel estimation. Meanwhile, better deblurring result under both visual and quantitative assessments can be achieved2. In our work, we explore an effective strategy of multi-scale nonlocal regularization. The strategy tailored towards the estimated blur kernel to a real blur kernel in the process of coarse to fine fashion. In detail, at last iteration of different scales, we introduce the nonlocal prior to regularize the solution of blur kernel. Experimental results demonstrate that the strategy can make the approach which based on the patch priors to work better for blur kernel estimation. Finally, with a non-blind deblurring algorithm, a faithful latent image can be obtained. With a comprehensive evaluation criterion, it shows that our approach performs better than some state-of-the-art results.3. Based the work of(2), we proposed a specific shock filters which can enhance salient edges reasonably. Meanwhile, we perform the specific shock filters in the phase of blind image deblurring. We also combine the specific shock filters with advanced patched prior information. Through the way, we can use the edge information of image in hidden and distinct way. Thus, we can reduce the negative effect caused by naive shock filters. With the strategy, a more accurate blur kernel can be obtained. Finally, a more faithful recovery result can be obtained.