A Dual-U Structure for Image Denoising Based on Attention Mechanism
Junhui Yin, Kangjia Xu, Jing Kan, Fangyan Dong, Kewei Chen · 2023
The SWA-U2former with a nested dual U-shaped structure for removing complex noise from real images is proposed, which is an effective image denoising network framework. Each U-shaped structure includes an encoding and decoding structure based on the shifted-window core. The following are some benefits of the design: Firstly, it can better capture the dependencies between pixels over long distances and can resolve image information at different scales. In Shifted Window Attention Residual U-blocks (SWA-RSU), due to the use of a Shifted-window Transformer and the stacking of multiple encoders and decoders, it not only solves the problem of high computational complexity caused by the attention mechanism but also fuses feature maps at different scales, making the restoration effect more robust and ensuring the stability and accuracy of image denoising. Secondly, the Attention mechanism is combined with image convolution in the SWA-RSU module. the attention mechanism is mainly used to analyze the image. Since the entire network has a nested dual U-shaped structure, collecting the outputs of each layer of the SWA-RSU decoder, fusing their features, and finally using convolution to obtain the final result. The attention mechanism can help convolutional neural networks better deal with image noise and distortion. By applying different attention mechanisms to different regions, details and information in the image can be better preserved, thus improving the effectiveness of image denoising and restoration. Under these two designs, SWA-U2former can capture the deviation between degraded and original images at multiple scales, effectively completing the task of image denoising. Finally, conducted experiments on benchmark datasets, and SWA-U2former achieved outstanding performance compared to the best algorithms.