Real-World Superresolution by Using Deep Degradation Learning

Most current deep convolutional neural networks can achieve excellent results on a single image superresolution and are trained using corresponding high-resolution (HR) images and low-resolution (LR) images. Conversely, their superresolution performance in real-world superresolution tests is reduced because these methods create paired LR images by simply interpolating and downsampling HR images, which is very different from natural degradation. In this article, we design a new unsupervised framework conditioned by degradation representations of real-world hyperresolution problems. The approach presented in this paper consists of three stages: we first learn the implicit degradation representation from real-world LR images and then acquire LR images by shrinking the network, which will share similar degradation with real-world images. Finally, we make paired data of the generated real LR images and HR images for training the SR network. Our approach can obtain better results than the recent SR approach on the NTIRE2020 real-world SR challenge Trackl dataset.