Self-Supervised Image Denoising for Real-World Images With Context-Aware Transformer

In recent years, the development of deep learning has been pushing image\ndenoising to a new level. Among them, self-supervised denoising is increasingly\npopular because it does not require any prior knowledge. Most of the existing\nself-supervised methods are based on convolutional neural networks (CNN), which\nare restricted by the locality of the receptive field and would cause color\nshifts or textures loss. In this paper, we propose a novel Denoise Transformer\nfor real-world image denoising, which is mainly constructed with Context-aware\nDenoise Transformer (CADT) units and Secondary Noise Extractor (SNE) block.\nCADT is designed as a dual-branch structure, where the global branch uses a\nwindow-based Transformer encoder to extract the global information, while the\nlocal branch focuses on the extraction of local features with small receptive\nfield. By incorporating CADT as basic components, we build a hierarchical\nnetwork to directly learn the noise distribution information through residual\nlearning and obtain the first stage denoised output. Then, we design SNE in low\ncomputation for secondary global noise extraction. Finally the blind spots are\ncollected from the Denoise Transformer output and reconstructed, forming the\nfinal denoised image. Extensive experiments on the real-world SIDD benchmark\nachieve 50.62/0.990 for PSNR/SSIM, which is competitive with the current\nstate-of-the-art method and only 0.17/0.001 lower. Visual comparisons on public\nsRGB, Raw-RGB and greyscale datasets prove that our proposed Denoise\nTransformer has a competitive performance, especially on blurred textures and\nlow-light images, without using additional knowledge, e.g., noise level or\nnoise type, regarding the underlying unknown noise.\n