Blind Image Deblurring via Local Maximum Difference Prior

Abstract

Blind image deblurring is a well-known conundrum in the digital image processing field. To get a solid and pleasing deblurred result, reasonable statistical prior of the true image and the blur kernel is required. In this work, a novel and efficient blind image deblurring method which utilizes the Local Maximum Difference Prior (LMD) is presented. We find that the maximum value of the sum of the differences between the intensity of one pixel and its surrounding 8 pixels in the local image patch becomes smaller with motion blur. This phenomenon is an intrinsic feature of the motion blur process, we demonstrate it theoretically in this paper. By introducing a linear operator to compute LMD and adopting the $L_{1} $ norm constrain to the LMD involved term, an effective optimization scheme which makes use of a half-quadratic splitting strategy is exploited. Experimental results show that the presented method is more robust and outperforms the most advanced deblurring methods on both composite images and ground-truth scenes. Besides, this algorithm is more general because it does not require any heuristic edge selection steps or need too many extreme value pixels in the input image.