Example-guided image prior for blind image deblurring

This paper proposes a patch-based deblurring method to leverage the unregistered sharp example image which shares global or local contents with the blurred image in a variant view. Firstly, we propose a coarse-to-fine scheme to achieve the accurate image patches matching and solve the mismatch problem caused by the blur ambiguity. Secondly, we use sharp image patches to form a patch prior which outperforms the generic prior in kernel estimation as it retain more image details which are usually filtered out using generic priors. The proposed method using unregistered sharp example image make it more practical to find examples from variant ways, e.g. the Internet, the succession capturing using hand-held camera, and multi-frame in videos. Experiments on real-world images show the patch prior can achieve more accurate and robust kernel estimation and outperforms state-of-the-art methods.