Single Infrared Super-Resolution via a Shifted Full-Scale Non-Local Network

Image super-resolution technology can reduce the cost and complexity of acquiring high-resolution infrared thermal images. In this study, we propose a novel end-to-end single infrared super-resolution network based on shifted full-scale non-local residual block. A full-scale non-local mechanism is first proposed to enhance the channel and spatial dependencies of the intra-frame feature map. Based on this mechanism, a shifted full-scale non-local residual block is constructed. By confining full-scale non-local to local windows while allowing for shifted-window connectivity, full-scale non-local residual block solves the problem that existed non-local structures are difficult to be reused and serves as backbone for super-resolution network. Qualitative and quantitative evaluation results show that our method has better performance on benchmark datasets of Set5, Set14, BSD100, and the self-built infrared dataset under the up-sample factor of $\times 2$ , $\times 3$ , and $\times 4$ .