Single image super-resolution via internal gradient similarity

Image super-resolution aims to reconstruct a high-resolution image from one or multiple low-resolution images which is an essential operation in a variety of applications. Due to the inherent ambiguity for superresolution, it is a challenging task to reconstruct clear, artifacts-free edges while still preserving rich and natural textures. In this paper, we propose a novel, straightforward, and effective single image superresolution method based on internal across-scale gradient similarity. The low-resolution gradients are first upsampled and then fed into an optimization framework to construct the final high-resolution output. The proposed approach is able to synthesize natural high-frequency texture details and maintain clean edges even under large scaling factors. Experimental results demonstrate that out method outperforms existing single image super-resolution techniques. We further evaluate the super-resolution performance when both internal statistics and external statistics are adopted. It is demonstrated that generally, internal statistics are sufficient for single image super-resolution.