Robust Self-Supervised Real-Image Denoising via Consensual Contrastive Regularization as Preserving Force

Different processing pipelines and camera settings exist, resulting in unique characteristics of real-world noises. However, out-of-distribution performance is essential, as using the same cameras as those in the training dataset is uncommon. Despite a few methods having been proposed to mitigate this problem, they all overlook self-supervised real-image denoising methods, which possess two characteristics particularly beneficial for out-of-distribution generalization. In this paper, we build a self-supervised denoiser that prioritizes out-of-distribution to achieve robust real-image denoising. In our denoiser, we use pixel-shuffle downsampling factor-based teacher distillation, in which denoised outputs of noisy images without pixel-shuffle downsampling serves as a pushing force to deviate away from pre-trained noise distribution learned by a student network. Additionally, consensual contrastive regularization is introduced in the representation space as a preserving force for the distribution. Trained on the SIDD dataset, our denoiser achieves state-of-the-art out-of-distribution performance when evaluated across multiple datasets: including the PolyU, CC, HighISO, Huawei, and iPhone datasets. It achieves an average PSNR of 38.77 dB and SSIM of 0.9609, outperforming all existing methods. Furthermore, when combined with a fine-tuning method, our denoiser maintains the best average PSNR after 5, 10, or 20 iterations under consistent conditions.