Unsupervised Single Image Super-Resolution Using Cycle Generative Adversarial Network

The current state-of-the-art deep learning based Single Image Super-Resolution (SISR) techniques employ supervised learning in the training process. In this learning, the Low-Resolution (LR) images are prepared by applying known degradation such as bicubic downsampling to the High-Resolution (HR) images. Unfortunately, bicubic down-sampling eliminates the natural image characteristics such as sensor noise, degradation due to built-in hardware, etc., and generates smooth images and hence generated images are different from the real-world data. When deep learning model is trained using such artificially generated LR-HR pairs, they often are prone to generate better SR results for real-world images. To circumvent this problem, we propose an SR framework that can train in an unsupervised manner using Generative Adversarial Networks (GANs). It contains mainly couple of networks called SR network and degradation network which work on an unpaired data of LR-HR images. The SR network learns to eliminate noise present in the LR image and super-resolve it. While, degradation network performs inverse of SR network (i.e. down-sampling and adding degradation from real-world images). We demonstrate the effectiveness of the proposed method by conducting extensive experiments on NTIRE-2020 Real-world SR challenge dataset where it demonstrates the superior performance over state-of-the-art methods in terms of both quantitative and qualitative assessments.