Single image super-resolution based on nonlocal similarity and sparse representation

Images with high resolution (HR) are often desired and required in imaging applications. Super-resolution (SR) is a promising but still challenging task to obtain an HR image (or sequence) from the observed low-resolution (LR) image (or sequence). In this paper, we present an SR approach for single image, by combining the image observation model, image nonlocal similarity, and sparse representation of image patches. First, we regularize low rank of matrix that is composed of nonlocal similar patches, to obtain an initial estimate of HR image. Then, each patch of the estimated image is sparsely represented in terms of LR dictionary, and the sparse coding coefficients are used to generate an output HR patch in terms of HR dictionary, resulting in the enhancement of high-frequency components in the HR estimate. The two over-complete dictionaries are jointly learned from example patch pairs in advance. Finally, the refined HR estimate is further constrained by the observation model. Experimental results have shown that the proposed SR method has obvious visual improvement in preserving edges and structures, while achieving comparable overall objective quality to the state-of-the-art methods.