Super-resolution Reconstruction of Remote Sensing Images with Improved SwinIR Transformer

Aiming at the problem of large differences in feature information between different targets in remote sensing images and the complexity of the scene, an improved SwinIR Transformer remote sensing image super-resolution reconstruction algorithm is proposed, combining residual hybrid attention and spatial-gate feed-forward network. First, a new residual hybrid attention module (RHAB) is constructed, which is serially combined with multiple self-attention (MSA) to enhance feature extraction by mixing channel attention (CA), spatial attention (SA), and MSA, and to improve the model's ability in modeling locally and globally relevant information. Second, a spatial-gate feed-forward network (SGFN) is used instead of the multilayer perceptron (MLP) in the original network to effectively reduce redundant channel information, overcome the limitations of the MLP in dealing with local spatial features, and enhance the feature expression efficiency. The experiments use NWPU-RESISC45 and UCMerced_LandUse as the training set, RSOD-Dataset, RSSCN7 Data Set, and WHU-RS19 Data Set as the test set, and the model performance is evaluated by PSNR and SSIM metrics. The experimental results confirm that the model in this paper shows excellent performance in both objective evaluation and subjective visualization.