Real-World super-resolution under the guidance of optimal transport

In the real world, lacking paired training data makes image super-resolution (SR) be a tricky unsupervised task. Existing methods are mainly train models on synthetic datasets and achieve the tradeoff between detail restoration and noise artifact suppression based on a priori knowledge, which indicate it cannot be optimal in both aspects. To solve this problem, we propose OTSR, a single image super-resolution method based on optimal transport theory. OTSR aims to find the optimal solution to the ill-posed SR problem, so that the model can restore high-frequency detail accurately and also suppress noise and artifacts well. Our method consists of three stages: real-world images degradation estimation, LR images generation and model optimization based on quadratic Wasserstein distance. Through the first two stages, the problem of no paired image is solved. In the third stage, under the guidance of optimal transport theory, the optimal mapping from LR to HR image space is learned. Extensive experiments show that our method outperforms the state-of-the-art methods in terms of both detail repair and noise artifact suppression. The source code is available at https://github.com/cognaclee/OTSR.