Image Super-Resolution Based on Non-local Convolutional Neural Network

Super-resolution (SR) is a widely used image processing technology. Recently, deep learning theory has been introduced into image SR task and achieved a significant improvement. However, feature extraction by convolutional mask is often operated in image local blocks, few attentions are given to non-local. Our aim is to insert a non-local network module into deep learning framework to further improve the performance of SR network. We choose ResNet as backbone and SRGAN as benchmark. The non-local network module is added into ResNet and a newgenerator network based on SRGANis designed. To our knowledge, it is the first time to introduce non-local module into SR convolutional neural network. With a full discussion about the new generator network structure, our experiment shows that the introduction of non-local network for high resolution image generation has a superior performance to local or global convolutional neural network in image SR tasks.