Infrared image super-resolution via transformed self-similarity

Single image super-resolution is of great importance in computer vision. Various methods (e.g. learning methods) have been successfully developed in recent years. Despite the demonstrated success in the natural images, less research focuses on the infrared images. In this paper, we present a transformed self-similarity based super-resolution method without any learning priors, restore high-resolution infrared images from low-resolution ones. We exploit appearance similarity, dense error, and region covariances, and use the detected cues to guide the patch search process. We also add scale cue to consider local scale variations. We then present a compositional framework to simultaneously accommodate the four different cues. Experimental results demonstrate that our method performs better than previous methods, restores pleasant results, and high evaluate scores further show the effectiveness and robustness of our method for the infrared images.