Research On Prior Constrain Based Image Blind Restoration Method

Image blind restoration is a serious ill-posed problem, generally needed to combine some priors about image and blur kernel to impose regularization constraints on the whole process to trun it into a well state. Image blind restoration contains two basic frameworks. One is prior identification, which first estimates the blur kernel, and then restores the blurred image using the estimated blur kernel. The other is joint identification, which estimates the blur kernel and latent image simultaneously. Due to divide the whole image blind restoration into two separate processes, namely blur kernel estimation and latent image restoration, prior identification need less computation than joint identification. This thesis does the image blind restoration research work under prior identification. First, it had given an in-depth analysis on the problem of image and blur kernel constrains, and then proposed a new image blind method based on prior constrains according to the analytic results. In the blur kernel estimation process, the proposed method construct a cost function by the L0 sparse prior of image and Gaussian prior of blur kernel, and then in the latent image restoration process restorate the blurred image using an non-blind image restoration method based on Hyper-Laplacian prior constrain. The main works summarize as follow:1) Made a survey on the development status of domestic and foreign about image blind restoration method, and studied the relevant theory of image blind restoration in-depth, then the main research difficulties of image blind restoration was summarized.2) Given a research on the problem of prior constrains with respect to images and blur kernel in the blur kernel estimation and image restoration process, respectively. In the blur kernel estimation, mainly analyzed the image prior of L1 prior, L1/L2 normalized prior, L0-Sparse prior. In the latent image restoration, mainly analyzed the image prior of Gaussian prior, Laplacian prior, Mix- Gaussian prior, Hyper-Laplacian prior.The analyze result show that in the blur kernel estimation process use the L0-sparse prior of image is helpful to estimate an accurate blur kernel, and in the latent image restoration process use the Hyper-Laplacian prior is helpful to restore the clear image details.3) Proposed a L0-sparse constrain based blur kernel estimation method, which used the L0-sparse prior of image and Gussian prior of blur kernel to construct the cost function, and at the same time introduced an adaptive adjustment factor into the image regularization parameter. This factor can adjust the image regularization parameter according the different areas of image automatically, thus the presented method can still estimate an accurate blur kernel when the blurred image contains rich details or degraded by large blur kernel.4) Designed an iterative algorithm based on half-quadratic penalty technique to solve it. To suppress noise interference with blur kernel and at the same time not destroys the intrinsic characteristics of blur kernel, impose normalization and dynamic threshold constrains on the estimated blur kernel after each iteration. After the accurate blur kernel was estimated, the blind method based on Hyper-Laplacian prior constrain was used to do the blur image restoration.5) To verify the effectiveness of the proposed method, this article has do some experiments on the artificial blurred images and some challenged real-world blurred images, and compared with the existing state-of-art methods, the results illustrated that the proposed method has some improvement on the subjective evaluation and objective evaluation of PSNR.