Font Size: a A A

Research On Regularization Constraint Based Blind Image Deblurring Methods

Posted on:2024-05-23Degree:DoctorType:Dissertation
Country:ChinaCandidate:D Y YangFull Text:PDF
GTID:1528307124993679Subject:Control Science and Engineering
Abstract/Summary:PDF Full Text Request
Blind image deblurring is an important research topic in the field of artificial intelligence and pattern recognition,and also is an important technology in low-level computer visual processing.It is widely used in various artificial intelligence scenes.The main task of blind image deblurring is to estimate the blur kernel from the blur observation and recover the latent sharp image.During the shooting process,the image will encounter many irresistible factors,such as camera shake,out of focus,noise,low light,overexposed,foreground motion and complex weather conditions,which will seriously degrade the quality of the obtained image and affect the intuitive visual observation of humans and subsequent computer vision processing tasks.Therefore,recovering clear and meaningful latent image from blurred observation has important research significance.In recent years,blind image deblurring algorithm has achieved excellent performance in edge selection,data item fitting,sparse constraint,low rank constraint,and channel prior.Specifically,edge selection emphasizes the role of texture contours in restoration tasks.Data item fitting is an important manner to eliminate image noise.Sparse and low rank are mainly constraints on the statistical distribution of image pixels and gradients.Channel prior is to reduce the effect of blur on the extreme values of the image through designing a selection strategy.Under the framework of maximum a posterior(MAP),this dissertation analyzes the factors that cause image quality degradation in the blur process,and designs different priors to overcome the effects of blur,and finally constructs new blind deblurring methods.The main work of this dissertation are summarized as follows:(1)A blind deblurring method based on an enhanced sparse prior is proposed.Aiming at the situation that blurring damages the high-frequency part of the image,the log function is used to compensate the degradation of the high-frequency components while suppressing the low-frequency components,thus stabilizing the estimation process of the intermediate image.The classic (?) norm is a naturally sparse measure.Since its optimization is an non-deterministic polynomial(NP)-hard problem,the sparse l1 norm is usually used to approximate the (?) norm.Nonetheless,the penalty level of l1 norm on each data element depends on the intensity of the image pixel,while the penalty level of the (?) norm is the same.The proposed log function is formulated as a weighted l1 norm using the first-order Taylor expansion.The proposed prior is closer to the (?) norm than the traditional l1 norm.and eliminates the dependence of data elements.Due to the introduction of the weighted l1 norm,the optimization of the objective function becomes very difficult.We employ alternate direction minimization(ADM),iterative threshold shrinkage algorithm(ITSA)and fast fourier transform(FFT)to improve the efficiency of the algorithm.Experimental results show that,compared with the existing blind image deblurring methods,this method can effectively improve the performance of blind deblurring.(3)A sparse and low-rank joint prior for the blind image deblurring method has been developed.The method involves comparing and analyzing the statistical feature distributions of natural and blurred images.It utilizes the re-weighted nuclear norm and (?) norm to capture the low-rank and sparse characteristics of natural images,respectively,and combines them to form a mixed image prior.The application of the weighted nuclear norm eliminates small artifact edges and fine texture structures while preserving the main structure of the image,playing a crucial role in estimating the blur kernel.The (?) norm of the gradient is typically employed to smooth image contours,and suppress ringing artifacts.Additionally,the weighted nuclear norm offers increased flexibility and robustness in the presence of noise.The problem is solved using the alternate direction optimization strategy and the weighted nuclear norm minimization(WNNM)algorithm.Experimental results demonstrate that this method effectively suppresses noise,restores sharp latent images,and competes favorably with state-of-the-art methods.(3)A blind deblurring method based on dual-channel contrast prior(Dual-CP)is designed.Aiming at the problem that the contrast of an image is significantly degenerated after blurring process,a prior to enhance the contrast of intermediate image was studied,and the feasibility of using this prior to estimate the blur kernel was explored.Utilizing the difference error between the dark channel and bright channel to model the contrast of the image local patches,and combining with the sparse constraint of the image gradient,a reliable intermediate latent image with high contrast to estimate a more accurate blur kernel and realize the purpose of blind deblurring is restored.At the same time,the optimization of this prior is a challenging problem.The half-quadratic splitting(HQS)method is used to solve this non-convex (?) minimization problem.Extensive experiments of this method on different datasets show high performance,especially in low light images.Meanwhile,the exploration of the prior constraint norm and the validity experiments have consistently verified the reasonable validity of the proposed method.Furthermore,the proposed dual-channel contrast prior can be directly extended to non-uniform images deblurring.(4)A blind deblurring method based on sparse channel prior is presented.The classic dark channel prior(DCP)and bright channel prior(BCP)models the image local patches extreme values and embeds constraints on the extreme values into the blur kernel estimation model.However,in areas where bright pixels and dark pixels are mixed,a single dark channel or bright channel is unlikely to help kernel estimation.By studying the ratio of the dark channel prior to the bright channel prior,the pixel changes in the confounding areas are modeled as sparse channel prior to solve the blind deblurring problem.We prove that the prior is favor clear image than corresponding blurred one both empirically and mathematically.The proposed sparse channel prior not only enhances the sparsity of the classic DCP,but also reveals the confrontational relationship between DCP and BCP.By utilizing auxiliary variable technology to integrate the proposed sparse channel information into the iterative recovery process,a reliable intermediate image can be obtained for kernel estimation.A large number of experiments on the real and synthetic blur sets show that the sparse channel prior is effective for blind deblurring.Compared with existing methods,this method exhibits a certain degree of competitiveness.(5)A blind image deblurring method based on fast local extreme intensity prior(LEP)is proposed.Aiming at the problem that the extreme intensity value is damaged after image blurring,a local extreme intensity value collection function regularized by (?) norm was studied,and finally embedded into the blind deblurring framework as an image prior.The prior is used to deal with the degradation problem through the restoration strategy of extreme intensity values on local non-overlapping patches.Additionally,due to the inherent sparsity of LEP,its sparsity change serve as an important indicator to distinguish between blurry and clear images.The optimization of this prior is challenging,the sparse representation of latent images is obtained by sparse inducing the local extreme intensity prior,and use the ideas of threshold shrinkage and alternate optimization to improve the efficiency of the algorithm.Experiments on multiple datasets show that this method can effectively restore images and has higher computational efficiency.More importantly,experiments have shown that this method can handle various challenging blurry scenarios.
Keywords/Search Tags:blind image deblurring, image prior, sparse, channel
PDF Full Text Request
Related items