Self-Supervised multi-image Super-Resolution Network for Space Image Restoration

Abstract

Space-based space surveillance (SBSS) has the advantage of being flexible and highly interpretable, allowing it to play a strategic role in monitoring space targets. During actual missions, however, bursts of push-frame imaging methods, dynamic noise during observation, large motions and other factors can result in visible degradation, and it isn’t easy to have the corresponding high-resolution (HR) images for the low-resolution (LR) images acquired at the high-speed intersection, which seriously limits the existing multi-image super-resolution (MISR) methods for space-based applications. To overcome these problems, this paper proposes a self-supervised video super-resolution deep learning method that can be trained end-to-end in space-based observations without LR/HR image pairs. We generate extra training pairs using different sizes of sampling factors for LR videos and train the video hyper-resolution network from coarse to fine using a pyramid format. Extensive experiments on the public satellite dataset BUAA-SID-share1.0 show that our approach outperforms traditional video hyper-resolution methods.