Consensual Contrastive Regularization Acting as Preserving Force for Self-Supervised Generalization of Real Image Denoising

Real image denoising methods excel in indistribution performance but struggle to generalize diverse realworld noises generated by different cameras. This limitation comes from different processing pipelines and camera settings, resulting in unique characteristics of real-world noises. However, out-of-distribution performance is crucial, given the impracticality of using identical cameras. To mitigate this generalization problem, our method utilizes a self-supervised framework to not only push away from training noise distribution but also preserve it. Firstly, we propose pixel-shuffle downsampling factor-based multi-teacher distillation as a pushing force. Secondly, consensual contrastive regularization is introduced as a preserving force for original training noise distribution by providing a constrained representation space. Finally, our method reaches the best average out-of-distribution performance, and is extremely lightweight so it's easy to deploy in low-computation devices.