ResCon: Residual consistency for real-world super-resolution

Real-world super-resolution is a highly challenging problem in the field of computer vision. Besides enhancing image resolution and improving visual details, information loss due to complex real-world degradations is desired to be restored. One of the primary hardness of this problem is finding sufficiently large paired datasets for training. Researchers have developed techniques that generate synthetic low-resolution pairs using high-resolution images with a generative adversarial network-based degradation generator to address this issue. In these approaches, the degradation generator is trained by utilizing real-world low-resolution images as the target domain, generating a degraded low-resolution counterpart of the high-resolution input. However, in general, one major drawback of these methods is that the real-world low-resolution images are only used to train the degradation generator. Therefore, they are not directly utilized for super-resolution training. We propose the ResCon, Residual Consistency, method to address this matter. Our approach enables the direct use of real-world low-resolution images in super-resolution training, in addition to degradation generator training. Our method is built on reconstructing real-world low-resolution images in the low-resolution image domain. We conducted extensive experiments on a large unpaired real-world face image dataset and compared the proposed method with two similar studies. We comprehensively evaluated super-resolved images considering image quality and usability, in addition to the traditional domain-based assessment methods. Moreover, we discussed the visual quality of the generated outputs in detail.