FIRSTFirst none-learning-based denoising was proposed in 2013.
Learning-basedEnd-to-end optimizing.
Dual-domainRaw denoiser adapts to noise variations and sRGB denoiser adapts to ISP variations.
Denoising Paradigm.Trained with synthetic images, DualDn can successfully generalize to real-captured images with high-quality denoising.

DualDn: Dual-domain Denoising via Differentiable ISP

Ruikang Li1,2, Yujin Wang1,†, Shiqi Chen3, Fan Zhang1, Jinwei Gu2, Tianfan Xue2,
1Shanghai AI Laboratory, 2The Chinese University of Hong Kong, 3Zhejiang University
Indicates Corresponding Author

Camera ISPs Our DualDn
[Notice that DualDn is only trained on synthetic images, without using any images from these cameras or ISPs during training.]

Overview

Methods

There are two typical ways to inject a denoiser into the Image Signal Processing (ISP) pipeline: applying a denoiser directly to captured raw frames (raw domain) or to the ISP's output sRGB images (sRGB domain).
However, both approaches have their limitations. Residual noise from raw-domain denoising can be amplified by the subsequent ISP processing, and the sRGB domain struggles to handle spatially varying noise since it only sees noise distorted by the ISP. Consequently, most raw or sRGB domain denoising works only for specific noise distributions and ISP configurations.
Unlike previous single-domain denoising, DualDn consists of two denoising networks: one in the raw domain and one in the sRGB domain. The raw domain denoising adapts to sensor-specific noise as well as spatially varying noise levels, while the sRGB domain denoising adapts to ISP variations and removes residual noise amplified by the ISP. Both denoising networks are connected with a differentiable ISP, which is trained end-to-end and discarded during the inference stage.

BibTeX

@article{li2024dualdn,
      title={DualDn: Dual-domain Denoising via Differentiable ISP}, 
      author={Ruikang Li and Yujin Wang and Shiqi Chen and Fan Zhang and Jinwei Gu and Tianfan Xue},
      booktitle={European conference on computer vision},
      year={2024}
}