Single image super-resolution based on nonlocal sparse and low-rank regularization

Abstract

Image super resolution (SR) is an active research topic to obtain an high resolution (HR) image from the low resolution (LR) observation. Many results of existing methods may be corrupted by some artifacts. In this paper, we propose an SR reconstruction method for single image based on nonlocal sparse and low-rank regularization. We form a matrix for each patch with its vectorized similar patches to utilize the redundancy of similar patches in natural images. This matrix can be decomposed as the low rank component and sparse part, where the low rank component depictures the similarity and the sparse part depictures the fine differences and outliers. The SR result is achieved by the iterative method and corroborated by experimental results, showing that our method outperforms other prevalent methods.