Deep Generate Residual Similar Feature Networks for Image Super-Resolution

Abstract

In this paper, Driven by advanced convolutional neural networks, we present a deep generate residual similar feature networks to improve super-resolution performance. Many researches have found that many CNN network training of SR requires skill and computing equipment. The input low-resolution features and CNN intermediate features contain rich similar feature maps. The different upscaling factor is used for needing the different model, which increase computational complexity. For the problem raised above, we proposed the generate residual feature (GRF) module to generate more high-frequency information in the residual feature. Each generated residual structure contains the short skip connection and long skip connection. Furthermore, we extract the similar feature in the residual feature by considering the interrelation among residual feature. Qualitative and quantitative assessments on benchmark datasets shows that we use different methods to achieve the same effect as best results of SR, while we make the network light weighted. Meanwhile, Our experiments shows that the pedestrian detection in the monitoring scene has achieved good results.