SwinFusion: attention fusion super-resolution network based on SwinIR

Image super-resolution reconstruction is a significant research direction in the field of computer vision. In recent years, methods based on the Transformer architecture, particularly SwinIR, have demonstrated impressive results in the domain of image super-resolution. However, experimental findings indicate that SwinIR reconstructs images using only limited input information, suggesting that the model has not fully realized its potential. In the present work, we introduce a pioneering SwinIR-based Attention Fusion Super-Resolution Network, denoted as SwinFusion. This model combines Convolutional Block Attention Module (CBAM) and Neighborhood Attention (NA) mechanisms, leveraging spatial and channel features, as well as adjacent pixel information, to activate more informative pixels. Additionally, we introduce the MixUp data augmentation technique to enhance the model's generalization capabilities. Experimental results demonstrate that SwinFusion outperforms the original SwinIR on multiple standard datasets, affirming the effectiveness of the proposed approach and offering new insights for research in image super-resolution reconstruction