Redescending Potential Function induced Two-edge Selection Strategy for Blind Image Deblurring

Blind Image deblurring problem is undoubtedly a fundamental task in computational imaging and computer vision. Its goal is to extract a clean image without knowledge of the blur kernel from a given blur image.Recently proposed RDP regularizations have shown their effectiveness in deblurring. However, under the prior framework, the physical intuition of RDP regularization cannot be utilized and the advantages of image collaborative smoothing-enhancing (CSE) cannot be achieved well. In this paper, we propose a combination of RDP regularization and edge selection methods to fill this gap. It is motivated by a simple idea that using edge selection allows RDP regularization to work directly on the salient edges of the image which makes it physically more intuitive and producing better results. Both conceptually and experimentally, this transformation has been demonstrated to be true.        A large number of experimental results show that the proposed method can obtain the most advanced results on both uniform and non-uniform data sets, even better than some deep learning methods. This method is also worth applying to the direction of deep learning.