Image noise modeling is a long-standing problem with many applications in
computer vision. Early attempts that propose simple models, such as
signal-independent additive white Gaussian noise or the heteroscedastic
Gaussian noise model (a.k.a., camera noise level function) are not sufficient
to learn the complex behavior of the camera sensor noise. Recently, more
complex learning-based models have been proposed that yield better results in
noise synthesis and downstream tasks, such as denoising. However, their
dependence on supervised data (i.e., paired clean images) is a limiting factor
given the challenges in producing ground-truth images. This paper proposes a
framework for training a noise model and a denoiser simultaneously while
relying only on pairs of noisy images rather than noisy/clean paired image
data. We apply this framework to the training of the Noise Flow architecture.
The noise synthesis and density estimation results show that our framework
outperforms previous signal-processing-based noise models and is on par with
its supervised counterpart. The trained denoiser is also shown to significantly
improve upon both supervised and weakly supervised baseline denoising
approaches. The results indicate that the joint training of a denoiser and a
noise model yields significant improvements in the denoiser.

该论文提出了一个框架，通过仅依赖噪声图像对而不是噪声 / 清洁图像对，同时训练噪声模型和降噪器，应用于 Noise Flow 架构的训练。结果表明，该框架在噪声合成和密度估计方面优于先前的基于信号处理的噪声模型，并且与其受监督的对手相当，训练的降噪器也被证明显着改进了监督和无监督基线降噪方法，这表明降噪器和噪声模型的联合训练可以显著提高降噪器的性能。