A novel approach based on differential evolution for blind deconvolution

Abstract

Blind deconvolution refers to a class of problems of recovering a sharp version of a blurred image without any information about the blur kernel. In this paper, we propose a novel approach for blind deconvolution based on differential evolution (DE) algorithm, which is arguably one of the most powerful stochastic real-parameter optimization algorithms. Thanks to DE algorithm, various non-conjugate kernel priors, which can be used to effectively restrain the estimated kernel from unexpected situations such as delta kernel, are prone to be introduced to the proposed approach. In order to accelerate the computation speed, we relax the image prior, utilizing the Gaussian prior instead of the well-known sparse prior. Then the optimization problem turns to be convex, what's more, the optimal solution can be effectively solved in frequency domain. In addition, we use the kernel prior cost to propose candidate solutions to speed up the computation further. Finally, given the estimated kernel, we estimate the sharp image by sparse prior. Experimental results and comparisons demonstrate the effectiveness of our method.