From Synthetic to Real: A Calibration-Free Pipeline for Few-Shot Raw Image Denoising

Abstract

Calibration-based and paired data-based methods have achieved significant developments in the RAW image denoising field. However, the former requires accurate noise modeling to synthesize training data, which is laborious owing to the specificity across different camera sensors. Meanwhile, the latter relies on the large quantity and high quality of real paired datasets, which are difficult to collect in real-world scenarios. To overcome these limitations, we propose a simple pipeline termed as S2R to efficiently adapt Synthetic noise to Real noise. S2R contains i) a calibration-free synthetic pre-training stage to teach the network to recognize a variety of noise patterns and intensities and ii) a few-shot real fine-tuning stage for quickly adapting to target camera sensors. Moreover, a multi-perspective feature ensemble strategy is applied to enhance the network with stronger generalization ability and further boost the performance. We achieve a competitive score of 30.97 with PSNR 31.23dB and SSIM 0.95 on MultiRAW test set, ranking 1st place in the MIPI2024 Few-shot RAW Image Denoising Challenge.