Normalization-based Residual Attention Network for Single Image Super-Resolution
Xiang Li, Juan Zhang · Research Square · 2022
Abstract Deep convolution-based single image super-resolution networks has achieved good performance in single image super-resolution(SISR). However, we observe that deeper networks for image super-resolution will improve the complexity for training networks and computational cost. On the other hand, as the layers of neural networks improve, it’s easy to lose the long-term information from preceding layers and ignore the latent relationship of features.To address these issues,this paper proposes a normalization-based residual attention network(NRAN) for single image super-resolution, which can improve the representational ability of network. The input low-resolution(LR) features are extracted via shallow feature extraction part, then we propose an normalization-based residual attention block(NRAB) in deep feature extraction part to improve the feature obtaining ability and feature correlations. Furthermore, in order to obtain better reconstructed image, we design a multi-scale feature reconstruction module(MFRM) to fuse multi-scale features. The experimental results show that our method is superior to other methods in evaluation metrics and visual results.