Robust Super-resolution Based on Selective Samples

We propose a robust super-resolution algorithm based on internal and external sample selection, which can effectively alleviate the false information caused by ill-conditioned external training set. Our algorithm introduces an "arbitrator" to select more reliable samples for the reconstruction model learning, and integrates the internal learning and external learning into a unified framework. Experiments demonstrate that our algorithm can obtain stable and high quality results with no requirement for the relevancy of external samples. Further, we propose an accelerated version, which is a very practical reconstruction algorithm with high efficiency and strong robustness.