Non-uniform motion deblurring with Kernel grid regularization

Camera shake during the exposure time often leads to spatially varying blurring effect on images. Existing work usually uses patch based methods that assume the blur in each patch is uniform to solve this problem. However, these kinds of methods do not consider the consistency between different patches and thus leading to inaccurate results with ringing artifacts. In this paper, we propose a Kernel mapping regularized method to solve the non-uniform deblurring problem, where the consistency between image patches is considered to improve blur Kernel estimation. We analyze the theoretical framework of blur Kernels which can be described as a motion path transference, and propose a robust Kernel estimation algorithm based on Earth mover's distance (Wasserstein metric) to preserve the properties of blur Kernels. In addition, we develop a new Kernel refinement method based on a proposed Ink Dot Diffusion that uses 8 directions of Kernel mapping flow where the erroneous Kernels are identified and corrected. Experimental results demonstrate that the proposed algorithm performs favorably against the state-of-the-art image deblurring methods.