Improved edge-guided network for single image super-resolution

In recent years, deep learning has been successfully applied to image super-resolution. It is still a challenge to reconstruct high-frequency details from low-resolution images. However, many works lack attention to the high-frequency part. We find that edge prior information can be used to extract high-frequency parts and applying soft edges to image reconstruction has achieved great results. Inspired by this, we focus on how to make full use of edge information to generate high-frequency details. We propose an improved edge-guided neural network for single image super-resolution (IEGSR), which makes full use of the edge prior information to reconstruct images with more abundant high-frequency information. For high-frequency information, we propose an edge-net to generate image edges better. For low-frequency information, we propose a global and local feature extraction module (GLM) to reconstruct the texture details. For the fusion of high-frequency information and low-frequency information, we propose a progressive fusion method, which can greatly reduce the number of parameters. Extensive experimental results demonstrate that our method can obtain images with sharper details. Applying our model to the Manga109 test set, the PSNR value of 4 times image super-resolution is as high as 39.02.