A CW-attention transformer-based method for image super-resolution reconstruction

Although Transformer-based image super-resolution reconstruction algorithms typically possess strong global modeling capabilities, they still suffer from insufficient extraction of local features in low-resolution images. The inadequate local modeling ability often adversely affects the reconstructed images’ quality, while the inherent self-attention mechanism introduces high computational costs. To address these problems, the paper puts forward a CW-Attention Transformerbased method for image super-resolution reconstruction. Built upon the SwinIR architecture, the proposed model first introduces a cross-attention mechanism formed by paralleling a channel attention mechanism module with a window attention mechanism, which effectively enhances global information extraction and improves detail processing capabilities. Second, a spatial attention module is integrated with Swin Transformer to construct a Fusion Transformer Layer, thereby strengthening the access of latent information. Finally, dilated convolution layers combined with ReLU activation functions are incorporated to reduce computational complexity while expanding the global receptive field for more effective utilization of contextual information. The results on four test data sets show that compared with other algorithms, the model achieves better results in PSNR and SSIM, and significantly improves the reconstruction quality.