Group Feature Information Distillation Network For Single Image Super-resolution
Ming Zhuo Chen, Jun Wu · 2021 7th International Conference on Computer and Communications (ICCC) · 2021
Single image super-resolution has achieved great success in various methods based on convolutional neural network (CNN). For a long period of time, the deep neural network (DNN) has a strong ability to map from low-resolution (LR) images to high-resolution (HR) images, and increasing the network depth can quickly improve performance, so the network is getting deeper and deeper. This also leads to a sharp increase in the computation cost, which restricts the application in many real scenarios. Therefore, research on fast and lightweight models is becoming more and more significant. One of the most advanced methods is information distillation algorithm (IDA). In this paper, based on the IDA, we apply the group convolution strategy to construct a group feature information distillation network (GFIDN). Specifically, the middle features are expanded and then grouped, and the IDA is applied in each group feature multiple times to extract useful features gradually. On the one hand, the distillation result of the previous group feature is fed to the next information distillation module. On the other hand, all distillation results are merged, then the candidate features are summarized according to their importance. We find that the GFIDN can not only enjoy the advantages of lightweight of group convolution and high efficiency of jump connection, but also maintain high model accuracy due to the application of IDA. Extensive experimental results suggest that the parameter scale and forward inference time are greatly reduced.