Multi-Scale Detail Enhancement Network for Image Super-Resolution

Deep convolutional neural networks (CNNs) are widely used in single image super-resolution (SISR) and provide remarkable performance. However, most existing CNN-based super-resolution (SR) models focus mainly on designing deep or wide architecture and neglect intended detail enhancement, thereby hindering the CNN representational capacity. To resolve this problem, we propose a multi-scale detail enhancement network (MS-DEN) for SISR. Specifically, we introduce a multi-scale detail extraction module (MS-DEM), which first converts features into a 3-channel simulation image, and then, directly extracts detail information from the simulation image space. Furthermore, we concatenate the 3-channel image and extracted detail image to generate detail-guided features. Subsequently, we propose a multi-context channel attention module (MC-CAM) to relatively better fuse local and global features, and enhance features containing discontinuous detail information. With detail enhancement, MS-DEN can restore highly accurate details and lead to performance improvement. Numerous experiments show that our MS-DEN achieves competitive performance against the state-of-the-art methods.