GCSTormer: Gated Swin Transformer with Channel weights for Image Denoising

In recent years, deep learning models typified by CNN and Transformer have achieved remarkable success in computer vision. However, CNN struggles to capture global information and lacks effective cross-region interaction. Transformer excels in global feature modeling but comes with high computational complexity and a large number of parameters.In this paper, we propose GCSTormer, a novel image denoising model that integrates shallow and deep feature extraction. For shallow feature extraction, we employ a three-layer atrous convolution with dense skip connections. This design enhances CNN’s ability to capture global information. In the main architecture of the model, we utilize the U-Net structure to learn multi-scale image information and adopt the improved Swin Transformer block as the basic layer. This block has been improved in two aspects. First, we use gated convolution to optimize information flow. Second, we apply transposed attention to compute attention in the channel domain. To evaluate the model performance, we conduct denoising experiments on grayscale and color images with synthetic Gaussian noise. The experimental results show our model has good results in both quantitative metrics and visual quality while maintaining a certain number of parameters and complexity.