Research On Blur Kernel Estimation And Image Restoration Methods For Blind Deconvolution

With the rapid development of mobile imaging devices,especially the popularized mobile phones with camera,image has played the most important role in recording daily life and sharing with friends.Since the imaging environment is very complex,such as noise,low light condition,object motion and camera motion due to its instability,it is very likely to introduce kinds of degradations in captured images,in which blur is one of the most key degradation factor,severely limiting the visual perception and high level computer vision analysis.On one hand,in consumer level mobile cameras,some techniques such as auto-focus and optical anti-vibration have been equipped,which relieves the blur caused by camera motion and defocus.However,these techniques can only provide mild solutions,and cannot handle more complex degradation cases.On the other hand,in computational photography,some researchers attempt to tackle these issues by modifying camera hardwares.The common strategy is to design new camera architecture,coded aperture and coded exposure configurations for collecting such more information that high quality image can be recovered in a higher probability.However,all these hardware-based solutions suffer from the problems of high cost,poor portability,complex configuration and so on,making them far from replacing the conventional cameras in consumer market.Therefore,it is in dire need to develop effective and efficient image restoration methods for conventional cameras.The degradation of blur is commonly formulated as the convolution of blur kernel and clean image,and thus image deblurring is also named as deconvolution.Meanwhile,deconvolution is an ill-posed inverse problem,due to the complex noises in capture,transportation and storage.Especially since blur procedure is not known in practice,blind deconvlution is a more challenging task.The existing blind deconvolution methods usually include estimating blur kernel and recovering clean image,both facing bottlenecks in computational efficiency,robustness and restoration quality.To address these issues,we in this dissertation will develop restoration models along with efficient and effective optimization methods for blind deconvolution.The contributions can be summarized as:(1)Blind deconvolution iteratively performs two stages,i.e.,blur kernel estimation and non-blind deconvolution for recovering image,in which efficient non-blind deconvolution plays the key role in guaranteeing high efficiency of blind deconvolution.The existing non-blind restoration methods are modeled in image domain,which directly recovers clean image from its blurry observation.We propose a novel image restoration model in gradient domain.Since gradient is more sparse than image,it is expected to achieve high convergence rate.By incorporating with Total Variation model,we further develop two fast optimization algorithms based on alternating direction method of multipliers(ADMM),resulting in derivative ADMM(D-ADMM)algorithms.D-ADMM algorithms can both converge to the global optimal solution,and is more computationally efficient than that modeled in image domain.Moreover,blur kernel estimation can be more well conditioned in gradient domain than that in image domain,and thus the derivative restoration model provides a solution for designing efficient blind deconvolution methods.(2)As for blur kernel estimation,the existing methods usually borrow natural image priors,and by carefully tuning parameters for each restoration stage,trivial solution can be avoided.Based on our proposed derivative restoration model,we propose an iterationwise prior framework for blur kernel estimation,in which image prior is modeled as hyperLaplacian distribution and each iteration has its own parameters.In this way,iterationwise priors can naturally provide dynamic salient edge selection for blur kernel estimation.Furthermore,to avoid hand-crafted parameter tuning for each iteration,we propose a discriminative learning framework,by which iteration-wise priors can be learned from training samples.Compared with natural image priors,iteration-wise priors are learned supervised by accurately estimating blur kernel,and thus can improve the robustness of blur kernel estimation.Also the iteration-wise priors can be successfully applied to other synthetic and real world blurry images.(3)Blur kernel estimation error is inevitable in blind deconvolution,and the existing non-blind restoration methods are all developed with kernel error free assumption,making it very likely to introduce artifacts such as ringing effects and color distortions.To model blur kernel estimation error,we propose a novel partial deconvolution model for recovering high quality images.We first estimate reference Fourier spectrum from blurry image that is more close to groundtruth one,we introduce a binary partial map to indicate the reliability of estimated blur kernel in Fourier domain.Therefore,during deconvolution only reliable Fourier entries are used,while adverse effect of blur kernel estimation error can be suppressed.To jointly estimate partial map and update clean image,we propose an E-M framework.Moreover,zero points in Fourier spectrum of estimated blur kernel can be naturally included in partial map without increasing computational cost,and thus de-ringing capability can be possed in partial deconvolution.By incorporating partial deconvolution into wavelet-based and learning-based restoration methods,high quality images can be recovered.(4)Furthermore,blur kernel estimation error by a certain blind deconvolution method has specific property It is very difficult to model blur kernel estimation error using a universal physical model for different blind deconvolution methods.We propose a simultaneous fidelity and regularization learning(SFARL)model,in which regularization term models natural image priors,while fidelity term models residual caused by blur kernel estimation error.We employ a set of large size filters to extract spatial context in residual image,and then a set of non-linear functions are used to model the distribution of filters responses.As for each blind deconvolution method,we can establish training dataset with abundant samples,and then SFARL can be effectively trained for recovering high quality images.Moreover,SFARL has a strong modeling capability,and can be applied to many other computer vision tasks,e.g.,image denoising,removing rain streaks etc.