Study On Multi Priors Hybrid Constraints For Blind Image Deblurring

Blind image deblurring achieves blurred image restoration under circumstance of unknown blur kernel.Actually,blur kernels of real life blurred images are usually unknown.So researches of this thesis have great practical significance and academic value.In the process of blind image deblurring,blur kernel estimation plays an important role,which contains modeling and solution.In this thesis,regularization constraints based on prior knowledge are also called prior constraints.Therefore,adopting the two-step framework,this paper analyzes prior knowledge of image and blur kernel,and proposes a multi priors hybrid constraints for blind image deblurring method.The key of this method is applying multi priors hybrid constraints to the blur kernel estimation model.And estimate the accurate blur kernel with the designed solution.Then combined with the estimated blur kernel,blurred images can be restored by the hyper-laplacian non-blind image deblurring model.To validate the effectiveness of the proposed method,this paper conduct experiment on both synthetic and real life blurred images.And results revealed that the proposed method is superior to state-of-art methods subjectively and objectively.The paper's main works are presented as follows:?1?Made an investigation on research status of domestic and overseas blind image deblurring methods,and studied theoretical bases of blind image deblurring,especially the regularization method based on prior knowledge.Then summary the main difficult problems of blind image deblurring research.?2?Further introduced the regularization method based on prior knowledge for blind image deblurring.As for image prior constraint,analyzed L₁ prior?L₁/L₂ prior?L₀prior and dark channel prior.As for blur kernel prior constraint,analyzed L₁ prior?L₂prior and mixed prior as well.And then this thesis carried out basic experiments,whose results showed that L₀ prior on image and hybrid order L₂ prior on blur kernel could estimate blur kernel more accurate.?3?Proposed a blur kernel estimation model with image and blur kernel combined constraints.Making full use of prior knowledge,L₀ norm on image gradient and its dark channel was used to constrain image prior globally and locally.And adopted L₂ norm on blur kernel and its gradient to protect blur kernel's sparsity and continuity as well.Due to the introduction of image global and local prior constraints,the model could still work well on various sceneries blurred images,such as natural,face and text images,and estimated accurate blur kernels.?4?Because of the adoption of L₀ norm and dark channel constraint,the model solution faced with non-convex and non-linear problems.To solve them,half-quadratic splitting algorithm and linear approximation was introduced,respectively.During the solving process,normalization and dynamic threshold constraints were imposed on blur kernel after each iteration to protect its physical property.And then this thesis adopted classic non-blind image deblurring approach with the estimated blur kernel to achieve blurred image restoration.?5?To verify the validity of proposed method,experiments on synthetic and real life blurred images were carried out,which contained natural,face and text images.The proposed method and several representative methods were all involved.The results revealed that the proposed method was superior to state-of-art methods subjectively and objectively.