[CVPR 2020] conference version: a physics based noise formation model for extreme low light raw denoising

Abstract ：

Due to the lack of rich and real data , The learned single image denoising algorithm can not be well extended to the real original image which is not similar to the data used for training .

Although the heteroscedasticity Gaussian model of noise synthesis can alleviate this problem , However, the noise sources caused by digital camera electronic devices are still ignored to a large extent .

To solve this problem , The author puts forward a new idea based on CMOS A highly accurate noise formation model for optical sensor characteristics , So that we can synthesize real samples , Better match the physical characteristics of the image forming process .

The main contributions of this paper can be summarized as follows ：

• A noise model is established to synthesize real noisy images , It can match the quality of real data under extremely low light conditions ;
• A noise parameter calibration method is proposed , We can adapt our model to a given camera ;
• A data set using various camera devices is collected to verify the effectiveness and generality of our model ;

Concrete realization ：

From photons to electrons

In the process of exposure , Incident light in the form of photons hits light