Deep Residual Attention U-Net for Cross-Dataset Image Denoising and Restoration

Abstract

Real-world images are often noisy. Noises may come from sensor limitations, low illuminations, and other factors. Regardless of the sources, noises degrade image's quality and often affect the accuracy of later analysis in computer vision. To solve this, various denoising methods have been proposed. However, many of these methods' effectiveness often varies across datasets that may contain different noise characteristics. In this study, we propose AttenResUNet, a UNet-based architecture to denoise images across diverse noise conditions. This architecture combines residual dense connections with attention mechanisms. The aim is to capture both global image context and fine structural details during image reconstruction. Multi-scale feature extraction is used to represent information at different spatial resolutions, while attention gates help the network emphasize relevant regions of the image. The proposed model is evaluated on four benchmark datasets containing real noisy images with various degradation characteristics. Experimental results indicate that AttenResUNet consistently achieves higher PSNR and SSIM values than several baseline models. Interestingly, the proposed model showed particularly strong improvements on low-light images, where several baseline methods tended to oversmooth textures. Ablation experiments also confirm the importance of both the attention modules and the residual dense connections in achieving these improvements.