FG-SRGAN: A Feature-Guided Super-Resolution Generative Adversarial Network for Unpaired Image Super-Resolution

Recently, the performance of single image super-resolution has been significantly improved by convolution neural networks (CNN). However, most of these networks are trained with paired images and take the bicubic-downsampled images as inputs. It's impractical if we want to super-resolve low-resolution images in the real world, since there is no ground truth high-resolution images corresponding to the low-resolution images. To tackle this challenge, a Feature-Guided Super-Resolution Generative Adversarial Network (FG-SRGAN) for unpaired image super-resolution is proposed in this paper. A guidance module is introduced in FG-SRGAN, which is utilized to reduce the space of possible mapping functions and help to learn the correct mapping function from low-resolution domain to high-resolution domain. Furthermore, we treat the outputs of guidance module as fake examples, which can be leveraged using another adversarial loss. This is beneficial for the main task as it forces FG-SRGAN to learn valid representations for super-resolution. When applied to super-resolve low-resolution face images in the real world, FG-SRGAN is able to achieve satisfactory performance both qualitatively and quantitatively.