Improving GAN for Image Super-Resolution by Using Attention Mechanism and Dense Module
Junhan Zhao, Zhijie Zhou, Yu Liu · 2022 7th International Conference on Intelligent Computing and Signal Processing (ICSP) · 2022
With the gradual maturity of the convolutional neural networks (CNN) in image recognition, deep-learning based singe image superresolution (SISR) has shown great promise but also poses a challenge to current researches. Specifically, most SR methods have problems such as failing to reconstruct finer texture details, difficulty in determining the interdependence between each channel of the feature map, to name a few. To solve these problems, we propose a multi-level dense skip network (MDSN) and adapt it into SRGAN. This is a powerful model designed for SISR tasks, which combines confrontation training and extracts abundant local features via the contiguous memory mechanism. We achieve a trade-off between model performance and complexity by introducing the ECA module, which ensures the appropriate cross-channel interaction. The testing results on the public dataset demonstrate richer texture details and better visual effects of superresolution images restored by our model.