Blind Image Deblurring using Kernel-guided Nonlocal Patches and Nonlocal Low-rank Images

Abstract

Blind image deblurring aims to restore a high-quality image from a blurry image. Blind image deblurring has gained considerable attention in recent years because it involves many challenges in problem formulation, regularization, and optimization. In optimization perspective, blind image deblurring is a severely ill-posed inverse problem

therefore, effective regularizations are required in order to obtain a high-quality latent image from a single blurred one. In this paper, we propose nonlocal regularizations to improve blind image deblurring. First, we propose to use the nonlocal patches selected by similarity weighted by the kernel for the next blur-kernel estimation. Using these kernel-guided nonlocal patches, we impose regularization that nonlocal patches would produce the similar values by convolution. Imposing this regularization improves the kernel estimation. Second, we propose to use a nonlocal low-rank image obtained from the composition of nonlocal similar patches. Using this nonlocal low-rank image, we impose regularization that the latent image is similar to this nonlocal low-rank image. A nonlocal low-rank image contains less noise by its intrinsic property. Imposing this regularization improves the estimation of the latent image with less noise. We evaluated our method quantitatively and qualitatively by comparing several conventional blind deblurring methods. For the quantitative evaluation, we computed the sum of squared error, peak signal-to-noise ratio, and structural similarity index. For blurry images without noise, our method was generally superior to the other methods. Especially, the results of ours were sharper on structures and smoother on flat regions. For blurry and noisy images, our method highly outperformed the conventional methods. Most of other methods could not successfully estimate the blur-kernel, and the image blur was not removed. On the other hand, our method successfully estimate the blur-kernel by overcoming the noise and restored a high-quality of deblurred image with less noise.