Image Super-Resolution Using a Enhanced Swin Transformer Network

Abstract

Image Super-Resolution is a technique in the field of image processing that involves enhancing low-resolution images to generate high-resolution images. This technique aims to improve the clarity and details of images, thereby enhancing their quality and usability. While state-of-the-art image restoration methods are based on convolutional neural networks, they still face challenges such as high demand for training data, computational resource requirements, and difficulty in handling fine details. In this paper, we propose ASTSR, a super-resolution reconstruction model based on data augmentation and Swin Transformer. ASTSR consists of four components: data augmentation, shallow feature extraction, deep feature extraction, and image reconstruction. The data augmentation layer generates new training samples by randomly cropping and blurring different regions of images, thereby expanding the training dataset and improving the model's generalization ability and robustness. The deep feature extraction module is composed of multiple Swin Transformer residual blocks (STRBs). We conduct experiments on different datasets, and the results demonstrate that ASTSR achieves superior performance compared to other state-of-the-art methods, with a performance gain ranging from 0.04 to 0.36 dB, while reducing the total number of parameters by 24%.