Blind Image Deblurring Based On Adaptive Priors

Blind image deblurring aims to recover a latent clear image given only a single blurry observation,which is one of the research focuses of image restoration.It is essentially a seriously ill-posed inverse problem.Existing studies have designed many efficient priors for the latent image based on the maximum a posteriori(MAP)estimation.These priors are integrated into the deblurring model to narrow down the solution space and reduce the ill-conditioned nature of the original problem.However,most effective priors are non-convex and non-smooth,they make the model optimization challenging.Meanwhile,using a single prior to deal with most real-word scenarios is obviously insufficient due to unknown image distribution,complex kernel structure,non-uniform noise and so on.To solve these problems,in this paper,we propose a prior navigated adaptive optimization scheme,and introduce different types of priors adaptively in the optimization process to perform sparse structure control(SSC)on the latent image,instead of the traditional single form of prior.Specifically,we design a lightweight MAP model,where only the necessary data term and optimization conditions are included,and all prior terms are set as a constant.Then,we modify the traditional projected gradient descent process,and use the result of SSC as the initialization of each projected gradient descent.Note that SSC as a separate module has good extensibility.In this paper,we implement the SSC module based on _p-normregularization and sparse structure learning respectively.Essentially,_p-normregularization corresponds to a series of shrinkage functions,while sparse structure learning corresponds to a set of denoising convolutional neural networks,both of which can control the sparse structure of image by adjusting hyperparameters.Compared to the former,the algorithm implemented by the latter is more robust to image noise,but its performance and generalization hinge on the training data.The proposed methods implemented in two ways are compared with existing works on both benchmark datasets and real-world images,and achieve state-of-the-art performance.Finally,we analyze the convergence of the algorithm with adaptive priors through experiments.