EESR: Edge Enhanced Super-Resolution

In this paper, an edge enhanced super-resolution network (EESR) is proposed for a better generation of high-frequency structures in blind super resolution. In EESR, an edge detection network is used to extract high-frequency pixels. A generative adversarial network (GAN) is used to fine-tune the EESR network. And a more applicative criterion "perceptive texture score" (PTS) is defined and used to evaluate the quality of generated high-resolution image. Experiments show that the new EESR is able to recover textures with 4 times up-sampling, and gained PTS of 0.6210 on DIV2K test set, which is much better than the state-of-the-art methods.