Unsupervised Image Super-Resolution with an Indirect Supervised Path

Abstract

The task of single image super-resolution (SISR) aims at reconstructing a high-resolution (HR) image from a low- resolution (LR) image. Although significant progress has been made with deep learning models, they are trained on synthetic paired data in a supervised way and do not perform well on real cases. There are several attempts that directly apply unsupervised image translation models to address such a problem. However, unsupervised image translation models need to be modified to adapt to unsupervised low-level vision task which poses higher requirement on the accuracy of translation. In this work, we propose a novel framework which is composed of two stages: 1) unsupervised image translation between real LR and synthetic LR images; 2) supervised super-resolution from approximated real LR images to the paired HR images. It takes the synthetic LR images as a bridge and creates an indirect supervised path. We show that our framework is so flexible that any unsupervised translation model and deep learning based super-resolution model can be integrated into it. Besides, a collaborative training strategy is proposed to encourage the two stages collaborate with each other for better degradation learning and super-resolution performance. The proposed method achieves very good performance on datasets of NTIRE 2017, NTIRE 2018 and NTIRE 2020, even comparable with supervised methods.