Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

A Physics-based Noise Formation Model for Extreme Low-light Raw Denoising (2003.12751v2)

Published 28 Mar 2020 in eess.IV and cs.CV

Abstract: Lacking rich and realistic data, learned single image denoising algorithms generalize poorly to real raw images that do not resemble the data used for training. Although the problem can be alleviated by the heteroscedastic Gaussian model for noise synthesis, the noise sources caused by digital camera electronics are still largely overlooked, despite their significant effect on raw measurement, especially under extremely low-light condition. To address this issue, we present a highly accurate noise formation model based on the characteristics of CMOS photosensors, thereby enabling us to synthesize realistic samples that better match the physics of image formation process. Given the proposed noise model, we additionally propose a method to calibrate the noise parameters for available modern digital cameras, which is simple and reproducible for any new device. We systematically study the generalizability of a neural network trained with existing schemes, by introducing a new low-light denoising dataset that covers many modern digital cameras from diverse brands. Extensive empirical results collectively show that by utilizing our proposed noise formation model, a network can reach the capability as if it had been trained with rich real data, which demonstrates the effectiveness of our noise formation model.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (4)
  1. Kaixuan Wei (13 papers)
  2. Ying Fu (98 papers)
  3. Jiaolong Yang (47 papers)
  4. Hua Huang (70 papers)
Citations (187)

Summary

Overview of a Physics-Based Noise Formation Model for Extreme Low-Light Raw Denoising

This paper addresses a critical challenge in computational photography: denoising raw images under extreme low-light conditions. The focus is on creating a physics-based noise formation model that closely mimics the actual noise characteristics encountered during the image formation process in modern CMOS sensors. This approach aims to improve the generalizability of neural network-based denoising algorithms that often suffer due to the discrepancy between synthetic training data and real-world images.

Motivation and Contributions

Conventional denoising models often rely on synthetic data generated using additive Gaussian noise models. While these models are efficient for moderate noise levels, they fail to capture the noise intricacies present under extreme low-light conditions, such as photon shot noise and circuit-induced banding artifacts, leading to suboptimal performance in real applications. This paper proposes a sophisticated noise formation model accounting for different noise sources inherent to electronic image sensors, thereby advancing the fidelity and applicability of synthetic data.

Key contributions include:

  1. Physics-Based Noise Model: The model encapsulates the noise characteristics native to CMOS sensors, including photon shot noise, pixel circuit noise, and quantization noise, diverging from the simplistic Gaussian assumptions.
  2. Noise Parameter Calibration: A method to calibrate noise parameters across different camera devices is developed, utilizing bias and flat-field frames for realistic parameter estimation.
  3. Extreme Low-Light Denoising Dataset (ELD): A comprehensive dataset is curated, featuring images captured by various modern cameras under controlled low-light conditions, facilitating a robust evaluation of denoising models.

Results and Analysis

The paper presents extensive empirical analysis revealing that denoising networks trained using the proposed noise model can achieve performance levels comparable to networks trained on real, paired datasets. Specifically, when tested on the SID Sony dataset and the newly introduced ELD dataset, the proposed model demonstrated substantial improvements in PSNR and SSIM scores, suggesting its effectiveness in characterizing complex noise patterns in low-light imagery.

The results also indicated that integrating realistic noise parameters into synthetic data generation significantly closes the domain gap, thereby enhancing model generalization. Visually, the denoised images showed less banding and color artifacts, further validating the model's robustness. Additionally, the paper emphasizes the flexibility of the noise formation model; it can be swiftly adapted to new camera models by recalibrating the noise parameters, showcasing its practical applicability in variably configured imaging systems.

Implications and Future Directions

The implications of this research stretch beyond the immediate problem of low-light denoising, paving the way for broader advancements in the simulation of realistic training data for machine learning tasks in computer vision. This enhanced realism in synthetic datasets could mitigate the need for exhaustive and costly real-world data collection, thus accelerating progress in various applications from augmented reality to autonomous vehicles operating in low-light environments.

Future work may focus on further refining the noise model to encapsulate more complex interactions within the sensor circuitry and adapting the model for video denoising challenges where temporal noise patterns need consideration. Additionally, leveraging emerging sensor technologies and integrating spectral sensitivity variability could lead to another leap in photorealistic simulation capabilities.

Overall, this work solidifies the foundation for future explorations into physics-based noise modeling, contributing a significant stepping stone towards truly autonomous learning systems that thrive in challenging lighting conditions.