Research On Blind Image Deblurring With Discriminative Priors And Its Extension

As one of the important carriers of information acquisition,preservation and transmission,image plays an irreplaceable role in modern fields.However,during the imaging process,due to the influence of noise,defocus,relative displacement or other factors,some information of the image is missing,and finally a degraded image is formed.As we all know,blind image deblurring is a highly ill-posed problem,which requires reasonable prior knowledge to constrain the blind image deblurring model so as to obtain the optimal solution of clear image.In this paper,under the traditional model and deep learning framework respectively,a new blind image deblurring algorithm is proposed with discriminative prior.The specific work is as follows:1.This paper proposes a new re-descending potential(RDP)prior with discriminative and complementary smoothing characteristics,and constructs a new blind deblurring algorithm model named L0-RDP.Firstly,compared with the previous blind deblurring regularization terms,the RDP discriminative prior in this paper can easily distinguish clear images from blured images,and drive the deblurring minimization problem to prefer clear images.More importantly,gradient-based L0 regularization term damages some edge information of image,and RDP discriminative prior can pursue more accurate edge information as the core clue of blur kernel estimation under the condition of distinguishing image edge steepness,which improves the accuracy of blur kernel estimation.In this paper,a lot of experiments demonstrate that L0-RDP can effectively estimate the blur kernel,and the restored image has high definition,finer edge,and strong robustness of the L0-RDP algorithm.2.Currently,the popular end-to-end blind image deblurring methods are vulnerable to the limitation of training datasets.Although Self Deblur method solves the dependency problem of datasets with its advanced self-supervised learning feature,this method obviously ignores the excavating and utilization of blind deblurring prior knowledge.Therefore,this paper attempts to integrate the RDP discriminative prior in the traditional model into Self Deblur network architecture,and Self Deblur+ is proposed to realize the effective combination of self-supervised learning and traditional methods.In addition,in view of the characteristics of less information and small size of blur kernel,this paper further adjusts the fully connected network structure of Self Deblur to ensure the continuity of blur kernel.Experiments on synthetic and real blurred images demonstrate that the Self Deblur+ method proposed has stronger deblurring performance than Self Deblur method.3.At present,the majority of blind deblurring works are affected by the model itself or the training datasets,which is difficult to be directly applied to the blind super-resolution task.Based on the self-supervised learning mode and powerful deblurring performance of Self Deblur+,this paper attempts to directly apply Self Deblur+ model to the blind super-resolution problem without changing the network framework,so as to realize the efficient unification of the blind deblurring and blind super-resolution models.The experimental results on the simulation dataset demonstrate that Self Deblur+ has stronger blind overfraction reconstruction performance compared with the newly proposed DIP-FKP method.