Single image super-interpolation using adjusted self-exemplars

Abstract

Super-resolution (SR) is an elegant technique that can reconstruct high-resolution (HR) videos/images from their low-resolution (LR) counterparts. Most of the conventional SR methods utilize linear mappings to learn complex LR-to-HR relationships, where these linear mappings are often learned from training. Inspired by our previous linear mapping based SR method [1], we propose a novel super-interpolation based SR method that utilizes adjusted self-exemplars. That is, in order to find sufficient amounts of LR-HR patch pairs in self-exemplars, we iteratively augment self-exemplars from an LR input image to create additional self-exemplars. In doing so, our proposed SR method is able to find well-learned linear mappings on-line from self-exemplars without using external training images, and outperforms other conventional SR methods.