Relearning a Downsampled Low-Resolution Representation for Image Super-Resolution

Abstract

Super-Resolution (SR)-based image transmission systems typically employ a framework where the sender transmits the downsampled Low-Resolution (LR) image to save transmission bandwidth, and then the receiver runs an SR module to super-resolve it to its original resolution. While existing work primarily focuses on exploiting various SR modules for improving the super-resolved images, the performance of each SR module inevitably encounters its own upper limit imposed by the module itself. In this letter, we propose a novel relearning method to overcome this limitation. Specifically, we investigate an adversarial relearning network that, upon receiving a downsampled LR image, generates its new adversarial LR representation that is better adapted to the super-resolving pipeline of the SR module. This adaptation enhances the suitability of the LR representation for the given SR module, creating a perceptual effect of surpassing the SR module’s performance ceiling, ultimately leading to higher-quality super-resolved images. We further introduce a cycle-consistency loss to guide the adversarial relearning process based solely on the LR image itself, without requiring any ground-truth supervision, since the original-resolution image is unavailable in the receiver. Extensive experiments validate the performance of the proposed method in terms of quantitative PSNR/SSIM/LPIPS scores and visual effects of the super-resolved images.