Scale-arbitrary Infrared Super-resolution Network based on Channel Attention Mechanisms

In our engineering application, it is necessary to up-scale low-resolution infrared images to the size of high-resolution visible light images. We have studied the latest single image superresolution technology based on deep neural networks(CNNs) and found that there are two main practices to improve superresolution performance: increasing the depth of the network and applying attention mechanisms. Therefore, we design a channel attention interaction(CAI) module at the head of the network and channel attention feature groups(RCAFGS) based on the progressive feature fusion(PFF) strategy in the network backbone, which can ease the gradient vanishing problem caused by increasing the depth and treat channel features wisely. And we adopt the scale-aware feature adaptation module and scaleaware upsampling module for tackling the problem of the non-integer asymmetric size. Finally, we propose one scale-arbitrary infrared super-resolution network based on channel attention mechanisms.