A Lightweight Single-Image Super-Resolution Method Based on the Parallel Connection of Convolution and Swin Transformer Blocks

In recent years, with the development of deep learning technologies, Vision Transformers combined with Convolutional Neural Networks (CNNs) have made significant progress in the field of single-image super-resolution (SISR). However, existing methods still face issues such as incomplete high-frequency information reconstruction, training instability caused by residual connections, and insufficient cross-window information exchange. To address these problems and better leverage both local and global information, this paper proposes a super-resolution reconstruction network based on the Parallel Connection of Convolution and Swin Transformer Block (PCCSTB) to model the local and global features of an image. Specifically, through a parallel structure of channel feature-enhanced convolution and Swin Transformer, the network extracts, enhances, and fuses the local and global information. Additionally, this paper designs a fusion module to integrate the global and local information extracted by CNNs. The experimental results show that the proposed network effectively balances SR performance and network complexity, achieving good results in the lightweight SR domain. For instance, in the 4× super-resolution experiment on the Urban100 dataset, the network achieves an inference speed of 55 frames per second under the same device conditions, which is more than seven times as fast as the state-of-the-art network Shifted Window-based Image Restoration (SwinIR). Moreover, the network’s Peak Signal-to-Noise Ratio (PSNR) outperforms SwinIR by 0.29 dB at a 4× scale on the Set5 dataset, indicating that the network efficiently performs high-resolution image reconstruction.