Research On Restoration Methods For Non-uniform Blurry Images

Image blurring caused by de-focus or camera shaking is a very common problem.Image deblurring is the most fascinating and the most important part of image processing no matter from the theory point or from the view of application.Although such research can date back to 70 years ago,the deblurred images always are not sharp enough.Most of current deblurring approaches,including maximum a posteriori based method,variational Bayesian based method and edge prediction based methods,are always composed of two important terms: the first one is the data fitting term,which models the process of the image blurring;The other one is the regularization term,which models the natural image priors for settling the ill-posedness of the deblurring problems.The data fitting term models forward blurring process.Most of the existing model suppose the blur image as the convolution of latent sharp image and the blur kernel.Nevertheless,recent researches has revealed that real-world blurring is non-uniform,and there has been a number of non-uniform blurring models.However,few of them can be used in real applications because of heavy memory burden or high computational cost.Image deblurring problem belongs to inverse problem,and the solution is not unique,in consequence,is a ill-posed problem.An efficient solution for ill-posed problem is to adopt image prior.In general,most representative model for image prior is the regularization term.Nevertheless,most of the popular regularizations prone to over-smoothing the deblurred image,and thus always result in unnatural deblurred images with lots of texture loss.In addition,there will be a mass of regularization parameters to be manually specified and deblurring results are always parameter sensitive,setting the parameters effectively for each blurring image could be a heavy workload.Blind Image deblurring method,can be classified into two parts: blur operator estimation and sharp image restoration given the blur operator.The second part,is the socalled non-blind deblurring method.Most of the image deblurring research start out with non-blind deblurring method.In this paper,we will use this strategy.To address these issues,we propose a series of image deblurring methods from the perspective of appropriate blur model and regularization term,which mainly includes:(1)Additive convolution model for non-uniform blur.We propose an additive convolution model(ACM)for improving the generalization ability of non-uniform blur model.The proposed ACM has three merits: firstly,it can model several kinds of real world nonuniform blur,and can easily embed into many deblurring methods.Secondly,filters in ACM can be substituted by series of basic filters,number of which is proportional to the computational cost of ACM.We propose a PCA-based method for learning the basic filters,leading to low computational complexity of ACM.Lastly,memory consumption is reduced significantly and thus avoided the application platform constraints faced by other non-uniform deblurring methods.Finally,we propose a ACM-based non-blind deblurring approach for non-uniform blurring,verified the effect of ACM.(2)An image delburring method based on gradient histogram preservation model(PGHD).The proposed ACM-based non-blind deblurring approach for non-uniform blurring utilize the TV regularziation,which unavoidably over-smoothing the textures and bringing in artifacts while promoting convergency speed.Therefore,we propose a PGHD-based approach.We utilize the ability of gradient histogram preservation model on enriching textures in deblurred images,and incorporate with the non-local centralized sparse representation model.PGHD can efficiently reduce the ringing artifact and noise derived during image deblurring.Experiment results show that the visual perception of the deblurred images have been obviously promoted.(3)Generalized additive convolution model(GAC)for non-uniform blur Blind deblurring is more practical than the above non-blind deblurring problems.However,in blind deblurring,the proposed ACM need to constantly updating basic filters and the coefficients,disabled its utilization.And thus based on the ACM scheme,we propose a GAC model to represent the non-uniform blurring.GAC inherits the generalization,computing complexity and memory footprint from ACM,and can easily be embed into blind non-uniform blurred image restoration algorithms.Furthermore,we proved that disassembling of camera shake trajectory into minimal numbers of slices and fibers can further reduce the computation complexity of the GAC model.So we propose a greedy algorithm to disassemble the camera shake trajectory.Finally,we constructed a GAC-based blind deblurring scheme for non-uniform blur images,verified the effect of GAC.(4)Blind deblurring of non-uniform blurred image with adaptive regularization term.Non-uniform deblurring method always need manually adjusting of the regularization parameters.But natural images follow various priors,corresponding to different parameters.Manually adjust the parameters is a tedious task and may be inefficient.Therefore,we propose a gradient image prior based regularization with adaptive parameters,and thenconstructed a blind deblurring method for non-uniform blur images based on maximum a posteriori frame.Experiments show that this method can significantly improve the efficiency of the blind deblurring method and visual perception of the deblurred images.