Pyramidical Based Image Deblurring via Kernel Continuity Prior

Blind image deblurring aims to reconstruct a sharp version of a given blurred image by estimating the blur kernel. This is generally a highly ill-posed problem because both the blur kernel and the sharp image are unknown. To solve this problem, different methods use sophisticated image priors to construct an intermediate sharp image and then estimate the blur kernel. In this paper, we utilize the advantage of the pyramid scheme that is usually used in blind deblurring literature, and instead of using complicated constraints to reconstruct the intermediate latent image, we use a super-resolution approach in a scale-by-scale manner. This scheme is performed quickly, although we realized that multi-scale prior leads to noisy kernels, and the estimated blur kernels are not sufficiently continuous. We focus on kernel regularization to encourage sparsity and continuity, thus we add a kernel continuity constraint to the optimization function. Finally, by conducting experiments and providing quantitative and visual results, we demonstrate that the proposed deblurring method is fast and provides acceptable results for blurred cases where state-of-the-art methods fail.