DASR+: Training Domain Distance Aware Network for Unsupervised Image Super-Resolution

Recently, unsupervised super-resolution has attracted increasing attention due to its promising potential in real-world applications. Early studies proposed generating synthetic low-resolution (LR) images that share similar degradation characteristics with real LR images while preserving the content of high-resolution (HR) images, and utilizing the generated pseudo pairs for supervised training of super-resolution networks. Nevertheless, the SR performance of these approaches is limited by the inevitable domain gap between synthetic and real LR images. In this paper, we propose a novel domain distance aware super-resolution (DASR+) approach with newly designed domain gap aware training strategy, domain distance weighted supervision strategy, and a domain distance adaptive network for unsupervised image super-resolution. Particularly, domain gap aware training takes additional benefit from real LR images and performs feature-level as well as image-level adversarial learning to leverage real-world LR images in the target domain. In addition, domain distance weighted supervision enables a more rational use of generated LR-HR pairs using domain distance information, and the domain distance adaptive SR network injects domain distance information into the network architecture to adaptively adjust the SR mapping. As a result, the SR model trained on synthetic data can be generalized to super-resolve real-world LR images without corresponding ground-truth images. We evaluate DASR+ on both synthetic and real datasets, our approach consistently outperforms state-of-the-art unsupervised SR methods in generating results with more realistic textures.