Dynamic degradation learning for real-world image super-resolution

Deep learning-based image super-resolution (SR) methods typically require high-resolution (HR) and corresponding low-resolution (LR) images to train SR networks. Since it is difficult to capture paired LR–HR images of exactly same scene from real world, most methods construct the training dataset by simply employing bicubic downscaling on HR images. Unfortunately, SR models trained on such dataset often fail in real-world scenarios, because there is a domain gap between the generated and real LR images. To narrow the gap, we propose a novel down-sampling network which introduces degradation kernels explicitly from real LR images via a dynamic learning scheme. Specifically, our Dynamic Down-Sampling Network includes two paths that take the unpaired clean HR and real LR images as input separately. The HR-path is used for downscaling HR images to obtain their LR version; the LR-path learns the degradation information directly from real LR images. And the learned dynamic per-pixel degradation kernels are applied to the outputs of HR-path to generate the downscaled LR images with real blur kernels, which can further combine with some additive noises. The dual-paths are trained jointly, and the unpaired training images can work collaboratively. With the generated paired data, the existing SR networks can be trained in a supervised manner for real-world SR task. Extensive experiments on both synthetic and real images demonstrate the effectiveness of the proposed method, which provides more natural textures and reduces over-smoothing.