SwinCTC: Efficient Network for Superresolution Reconstruction of Remote Sensing Images Based on Nonlocal Feature Enhancement by Sliding Window Mechanism

In Transformer-based super-resolution reconstruction tasks for remote sensing images, the window attention mechanism (WMSA) has become a key method for reducing the secondary complexity of traditional self-attention. However, the window self-attention mechanism still requires a significant amount of computational resources, especially when processing large remote sensing images. To address this issue, we propose a convolutional block structure (SwinCTC) based on the sliding window mechanism, which replaces the traditional window/sliding self-attention and drastically reduces computational complexity. It comprises a Residual Channel Enhanced Attention (RCEA) module and group convolution, which enhances the efficiency of group convolution by dynamically refining the channel weights through RCEA. Additionally, the CTC-Block further refines the window feature representation by introducing a spatial attention enhancement module that focuses on key spatial details and selectively emphasizes the information regions within each window. Finally, a convolution-based feedforward network is introduced to bolster the network's capacity to model high-frequency information in images. The experimental results demonstrate that the proposed method outperforms other classical remote sensing image super-resolution reconstruction models in terms of PSNR and SSIM evaluation metrics on the NWPU-RESISC45 and NWPU-VHR datasets. Compared with the baseline model SwinIR, the number of parameters is reduced by 39.4%, the number of floating-point operations is reduced by 42.4%, and the average inference speed reaches 22.86 ms while maintaining performance.