Blur kernel estimation via salient edges and nonlocal regularization

Blind image deblurring is a severely ill-posed inverse problem. To obtain a high quality latent image from a single blurred one, effective regularizations are required. In this paper, we propose a nonlocal regularization to improve blur kernel estimation. Under convolution operation, even similar patches could result in the quite different values. However, if the estimated kernel is correct, the nonlocal similar patches weighted by that kernel may result in the similar value by convolution. Therefore, the weighted nonlocal patches can improve the kernel estimation. We extract the nonlocal patches in terms of the weighted similarity by the kernel and then use them for regularization of the kernel estimation. Since the nonlocal regularization is a data-authentic prior, our approach not only mitigates the ill-posedness but also imposes the effective prior to kernel estimation. Experimental results show that our approach outperforms conventional blind deblurring algorithms.