Edge Guided Learning for Image Super-resolution with Realistic Textures

Super-resolution (SR) methods are used to reconstruct details in images to obtain an improved resolution. Recently, SR methods based on generative adversarial networks (GANs) have become seminal due to their effectiveness in generating textures. However, a common problem is the presence of unpleasant artifacts. In this paper, an edge-guided SR neural network (Edge-SRN) is proposed by introducing a plug-in edge detection module and incorporating a new edge loss, which increases the reconstruction accuracy and reduces artifacts. We also use the Edge-SRN as a teacher network to a knowledge distillation framework for training a lightweight student SR model. The student model learned from Edge-SRN outperforms its counterparts learned from GAN-based teachers or from the ground-truth HR images in both reconstruction accuracy and perceptual quality, which indicates the ability of reconstructing realistic textures can be transferred well from Edge-SRN to a small model. Extensive experiments on diverse criteria show the promising performance of our method compared with several state-of-the-art SR methods in the qualitative and quantitative evaluations. Our code is available at https://github.com/lizhangray/Edge-SRN