Compressed image super-resolution based on invertible degradation and restoration

The past decade has witnessed the increasing popularity of images on the Internet. To save the storage space, Internet images are usually downscaled and compressed. However, traditional image super-resolution (SR) methods struggle to effectively handle the complex degradation introduced by compression. This paper introduces a novel compressed image SR (CISR) method based on invertible degradation and restoration, namely InvCISR, aiming to restore high-resolution images from their low-resolution counterparts via explicitly learning from the degradation process. Different from the existing CISR works, the proposed invertible architecture enables complete recovery of degradation information, facilitating degradation-aware restoration. Additionally, a new codec simulation module and dual feedback training protocol are developed to further improve the invertible CISR pipeline, by minimizing the gap between degradation and restoration processes. Extensive experiments demonstrate that our InvCISR outperforms 12 state-of-the-art methods on multiple scenarios, including non-blind and blind CISR for four image codecs and compressed image enhancement, achieving at least 0.3dB PSNR improvement on 5 public datasets.