Single image super-resolution with one-pass algorithm and local neighbor regression

Self-similar repetitive patterns inside natural images are widely exploited for various image recovery tasks such as image denoising, image deblurring, and single image super-resolution (SISR). In this paper, we present a two-stage self-similarity learning-based SISR method by gradually magnifying an input low-resolution (LR) image to the desired HR one. In the first stage, the one-pass algorithm is applied to improve the compatibilities between neighboring high-resolution (HR) patches and local neighbor regression (LNR) is used to establish the mapping relationship from the LR to HR image patches. In the second one, we further boost up the quality of the LNR-based result by incorporating a fast non-local means (NLM) based regularization term into the reconstruction-based SISR framework. Experiments indicate that the proposed method is able to yield state-of-art SR performance without relying on any external exemplars.