Single Image Super-Resolution Using Feedback Attention Networks

Image super-resolution (SR) is a low-level vision task to recover HR images from their matching LR images. Recently, deep learning has provided a great boost to image SR tasks. In super-resolution tasks, obtaining better visual effects requires more accurate acquisition of high-frequency information. However, the input of low-resolution (LR) images contains a large amount of low-frequency information, which hinders the representational capability of networks and makes it more difficult to train. To address these issues, we propose a feedback attention network (FBAN) to better correct the features from LR input. Specifically, we use a constrained recursive neural network (RNN) to implement this feedback manner. Attention mechanism is added to the feedback block to make main network better extract high-frequency information. The attention mechanism consists of channel attention (CA) and spatial attention (SA). Extensive experimental results demonstrate the superiority of our FBAN.