Edge Information-based Blind Image Deblurring Algorithms

In the areas of astronomical observation,remote sensing imaging and public security,image quality is an important problem which has been paid much attention to.But the blurry effect would be introduced by various factors in image acqusition process,causes the degra-dation of an image.Because of the complicated degraded reasons of real blurred images,it is difficult to model the blur process.Additionally,blind deblurring is a challenging task due to its highly ill-posedness.Most of the existing blind image deblurring methods have only some theoretical value,which means they have not provided effective ways to solve pratical problems.Thus the research on blind image deblurring has a wide application prospect,and is one of research hotspots and difficulties in the field of image processing.This thesis follows the main thread of the processing flow in blind image deblurring.Focusing on the key problems,kernel estimation and non-blind image deblurring,the thesis designs corresonpding strategies and effective algorithms.The main achivements and innovations are listed as follows:(1)In the framework of heuristic edge-enhancing based blind deblurring,this thesis proposes a blind image deblurring method based on image decomposition and steerable gra-dient.Firstly,a fast image decomposition method has been introduced to obtain the cartoon part.Then the enhanced cartoon image is regarded as the initial latent image of following blur kernel estimation.Further,to accurately capture the edge features,steerable gradient bases on image local structure has been developed as an edge descriptor.Finally,with the linearity of convolution operator,the frequency response of steerable derivative filters are approximated by the linear combination of the frequency response of two standard deriva-tive filters,such that the model can be quickly solved by FFTs.The simulation and real data experimental results demonstrate that the model can accurately estimate the blur kernel and sharp image.(2)In the framework of MAP estimation based blind deblurring,this thesis proposes an optimal reweighted l1 norm based blind image deblurring method.To promote the ap-proximating ability to l0 norm,weighted l1 norm has been introduced into the model as an enhanced sparsity measure.Additionally,the weights can be computed according to the lo-cal structures to eliminate the negative effects of the tiny textures.For further reduce the influence of noise and isolated points,the weights are optimized in the iterative scheme.Finanlly,the half quadratic splitting technique is applied to efficiently solve the kernel esti-mation model.The experimental results demonstrate the superiority to other methods.(3)Considering the sparsity of image gradient assumes that the image is piece-wise smooth,which would lead stair-casing effects in restored image.To address the limitation,high-order TV model has been introduced into the blur kernel estimation model.Thus this thesis proposes a fused l0-l1 regularized blind image deblurring method.The new sparsity-inducing model can not only suppress the stair-casing effects,but also preserve the continuity of edges.To solve the proposed model,splitting augumented Lagrangian method has been developed.Finally,experimental results show that the proposed method achieves improve-ments both quantitatively and qualitatively.(4)For non-blind image deblurring,incorporating the l0 norm based sparsity in image gradient domain and nonlocal similarity as the implementary prior information,the thesis proposes a two-stage non-blind image deblurring method.At the first stage,salient edges and large scale textures are produced by minimizing the l0 norm of gradients.Then at the sec-ond stage,a nonlocal similarity regularization term and an edge similarity constraint are em-ployed to develop the nonlocal regression model based non-blind image deblurring method.For the optimization of the two separated models,corresponding numerical algorithms are designed specifically.The experimental results indicate that the proposed method can ef-fectively restore the sharp image,especially in aspects of preserving edges and textures,and also suppress Gibbs effects around image boundaries.