Dual-Streams Edge Driven Encoder-Decoder Network for Image Super-Resolution

Single image super-resolution (SR) aims at reconstructing high-resolution (HR) images from low-resolution (LR) ones. One of the most key issues is to recover finer image details of LR images. In this paper, considering the importance of edge prior for image SR, we propose a dual-streams edge driven encoder-decoder network, which combines edge stream-based encoder-decoder network (edge-EDN) and color stream based encoder-decoder network (color-EDN) to reconstruct HR images with more image details. Instead of utilizing two sub-networks to learn edge information and color image contents respectively, a multitask learning framework is developed to jointly train edge-EDN and color-EDN. Therefore, as the structure prior, the reconstructed HR edge maps are fused with learned features of color stream to refine the HR color images. To reconstruct HR images with better visual quality, a total loss function combining edge loss and color loss is designed to make an optimal tradeoff between the image fidelity and texture details. Our extensive benchmark evaluations demonstrate that our method for SR performs better both on high objective quality and human visual perception compared with several state-of-the-art SR methods.