Scale adaptive region selection for deblurring

Abstract

A scale adaptive region selection method for deblurring based on sparse representation and gradient priors is proposed. Using a pre-trained blurred dictionary, sparse representation can sparsely reconstruct the blur examples in an input blurred image. The initial location of the selected region is a patch which is reconstructed by the minimum number of blurred atoms by solving a minimisation problem, making the best example to represent a smooth-like component. With the initial location, the gradient priors based on authors’ observation is used by searching a series of extended discrete scales. The proposed method makes the region selection for deblurring highly effective and efficient by utilising sparse representation and informative features. Experimental results on synthetic blurred images demonstrate that the proposed method behaves favourably in blur kernel estimation and deblurring quality against state-of-the-art approaches.