Research On Blind Image Deblurring Method Based On Structure Prior And Sparse Representation

With the rapid development of optical imaging equipment,especially the popularity of digital cameras,smart phones and other digital devices,images have become an important way for people to perceive the world and communicate with others.However,in the process of imaging,inevitably due to the interference of geometric aberrations,camera shake,changes in depth of field,object motion and other factors,the captured images have a blur effect,which affects people's perception and understanding of image information.Therefore,the restoration of clear images from blurred images has important research significance.Because the image deblurring problem is highly ill-posed,most of the current deblurring methods mainly design special image priors for specific types of images,thereby constraining the solution space and obtain good deblurring results.These methods cannot effectively process various types of images,their universality is relatively poor,and the computational cost is high.In addition,affected by the complex imaging environment,real images usually contain a lot of noise and outliers,making the deblurring of such images extremely challenging.In view of the above-mentioned problems,this paper develops and designs efficient deblurring models and algorithms from the perspective of restoration models and optimization algorithms,based on image structure priors and sparse representation theories.The main contributions are as follows:(1)Aiming at the poor universality of previous image priors,a robust blind image deblurring method based on a low-rank prior is proposed.To effectively implement low-rank regularization,a more flexible weighted Schatten p-norm minimization prior is adopted.This low-rank prior can better preserve the salient edges while eliminating harmful details and noise,thereby more accurately representing the sparsity and self-similarity of the image structure.In addition,in order to extract significant edges more efficiently,the L₀regularization gradient is introduced into the restoration model.To solve the non-convex model,a numerical iterative algorithm based on the half-quadratic splitting strategy and the generalized soft-thresholding algorithm is developed.We also extend the proposed method to non-uniform deblurring effectively.The proposed method shows excellent deblurring performance on all types of images,thus verifying the universality of the method.(2)Analyzing the correlation between low-rank minimization and group sparse representation,a novel blind image deblurring method based on group sparse representation is proposed.By observing the sparsity of adjacent similar patches of the clear image and the ground-truth blur kernel,a group sparse representation of the latent image and the blur kernel is proposed to further ensure the non-local similarity and local sparsity of the intermediate results.To effectively implement sparse representation for groups of similar patches,a non-convex weight L_p-norm minimization constraint is adopted.In addition,an efficient adaptive dictionary learning method is also designed to reduce the computational complexity of dictionary learning.To solve the severely non-convex model,a numerical optimization algorithm is designed for the alternate iteration of the latent image and the blur kernel.The proposed method has achieved excellent deblurring effects on blurred images of various types of scenes.(3)Because the image priors of most of the current methods are designed to be more complicated,the computational cost of the algorithm is relatively high.Based on this,a deeper sparse patch-wise maximum gradient prior is designed,and an efficient and fast blind deconvolution algorithm is developed.Compared with other image priors,the proposed prior is simpler and sparser.The experimental statistics and theory verify the relationship between this prior and the blur process,that is,the maximum gradient value of non-overlapping local patches is significantly reduced after the blurring.To recover more efficiently,the prior sparse constraints,iteration strategy and other aspects of the algorithm have been greatly improved.To optimize the non-convex model,an efficient numerical optimization scheme based on the half-quadratic splitting strategy is designed.In terms of computational efficiency and restoration quality,the proposed method is superior to state-of-the-art methods.(4)Because outliers(impulse noise,saturated pixels,etc.)can seriously damage the linear convolution relationship of the image,traditional deblurring methods are difficult to restore the image containing outliers.Based on this,a robust image deblurring method for outliers is proposed.We analyze the reasons for the serious impact of outliers on traditional methods,and propose a robust Welsch loss function for blind image deblurring.Due to the robustness of the Welsch function to outliers and the characteristics of effective edge perception,the data fidelity and image prior terms of the model are uniformly described using the Welsch function,which greatly simplifies the model.In addition,a more flexible and effective weight function is derived based on the Welsch function,and is applied to the designed iterative optimization algorithm.The proposed method has achieved advanced deblurring performance on images with or without outliers.