Research On Binary Image Blind Restoration Algorithm Based On Sparse Prior

Image restoration,aiming to restore the ground truth image from the degraded image,is a hot area in digital image processing and machine learning.Generally,image restoration can be categorized into two types:blind image restoration and non-blind image restoration.Blind image restoration is the more applicable one which assumes the point spread function?PSF?is unknown,thus we need to not only estimate the PSF but also restore the degraded image.Blind image restoration is an ill-posed problem,to which previous work usually adds prior knowledge via regularization.However,we need to choose the prior information according to the specific image,and different prior can result in different complexity.This paper focuses on one special problem of blind image restoration——binary image blind restoration.There are only two specific values for each pixel in the binary image.Analysis of binary image shows that the pixel values of binary image are sparse,the performance of using the traditional₁L/₂L norm prior is not appropriate,hence we use theL₀-regularizer instead.Particularly,the main contributions of this paper are:Firstly,this thesis discusses the related work of image restoration problem,including the classic image deconvolution theory,the traditional image degradation/restoration methods and the frequently-used total variation image restoration method.Then we summarize the existing blind image restoration algorithms and analyze the shortcomings of these algorithms.This thesis analyzes the reason of image blurring,and the formulation of the three common blur types:defocus blur,motion blur and Gaussian blur.For more complicated spatial invariant deconvolution,we analyze the advantage and disadvantage of estimating the PSF using different regularizations in both image and image gradient space.We choose the most suitable prior for binary degraded image kernel estimating through the experiments.Secondly,this thesis proposes an alternating minimization algorithm for binary image blind restoration based on theL₀-norm minimization framework considering the binary property.Common image restoration methods regard binary image as gray-scale image to restore with an optional threshold for binarization,while considering the special property of the binary image will get better recovery results for this particular type of image.The proposed blind image restoration algorithm is based on the framework ofL₀ minimization program,while restraining the latent restored image to be binary.Finally,this thesis designs a scheme to solve the variables alternately and expounds the optimization process of the objective function of our binary image restoration model.We also design several contrast experiments to analyze the effect and robustness of our algorithm,the experimental results show that the proposed method consistently outperforms other blind deconvolution algorithms.