An Image Denoising Method Based on Swin Transformer V2 and U-Net Architecture

To address the issue of image degradation caused by noise during image acquisition and transmission, this paper proposes ST2CFNet, an image denoising model based on Swin Transformer V2. This model innovatively integrates Swin Transformer V2 into a U-Net architecture to fill the current research gap related to the application of Swin Transformer V2 in image denoising. The model's dual-branch mechanism of Transformer and local convolution efficiently extracts and combines the long-distance pixel correlations and local feature information of images, preserving more image details. A multi-feature fusion module is designed to learn more beneficial knowledge for image denoising from features across multiple layers of the network. To enhance convergence speed, adaptive adjustments were made to the Swin Transformer V2 block, effectively stabilizing and improving the training outcomes. Experimental comparisons with several popular image denoising methods demonstrate that this approach outperforms local convolution-based methods in terms of average peak signal-to-noise ratio and offers lower fioating-point operations compared to similarly effective Transformer-based methods, indicating that the model can significantly enhance inference efficiency while preserving more image details.