Multi-feature fusion image super-resolution network based on Swin Transformer

The Swin Transformer-based approach shows astounding performance in lower-level visual tasks such as image resolution enhancement. However, through analysis, we found that some networks have the problem of insufficient ability to model the local information of low-resolution images, resulting in inadequate feature extraction and poor quality of reconstructed images, which implies that we need to do some work to explore the upper limits of the transformer's capabilities as much as possible. In response to these challenges, we propose a novel transformer network called Channel-Space Hybrid Transformer (CSHT), which integrates the attention mechanism while facilitating efficient spatial information extraction. Specifically, the designed algorithm redesigns the Swin Transformer layers, uses the window self-attention in parallel with the channel attention module, and redesigns the Feed-Forward Network (FFN). This fusion strategy enhances the network's representational capabilities, leading to more comprehensive image feature extraction. The redesigned feedforward network we call Spatial Feed-Forward Network (SFFN), a module that mitigates the degradation of feature representation caused by redundant channel information, while effectively capturing spatial information. The proposed method's effectiveness is validated through extensive experiments on public datasets, demonstrating a favorable balance between performance and computational overhead.