ELD presents a novel physics-based noise formation model designed for extreme low-light photography. The project aims to address the challenges of accurate noise modeling in scenarios where traditional statistical models fall short. By capturing the characteristics of CMOS photosensors, ELD enables the synthesis of realistic noise patterns, facilitating research in advanced image denoising techniques. This work was presented at CVPR 2020 and published in TPAMI in 2021.
ELD introduces a highly accurate noise formation model based on CMOS sensor characteristics, enabling realistic synthetic samples. A new ELD dataset covering modern camera devices is introduced to evaluate the generalizability of neural network denoising schemes. Experiments demonstrate that convolutional neural networks trained with synthetic data can achieve performance comparable to or exceeding that of networks trained on paired real data. The codebase includes a customized rawpy library with extended functionalities and provides scripts for reproducing results and training models.
- Noise Model: Physics-based model accurately simulates noise characteristics for realistic noise synthesis.
- ELD Dataset: A dedicated dataset captures low-light images under various conditions for training and evaluation.
- Pretrained Models: Pretrained models are provided to facilitate rapid experimentation and research.
- Rawpy Integration: Leverages a customized
rawpylibrary for enhanced camera data handling. - Scripting: Includes scripts for data preparation, model training, and result reproduction.
- Calibration: Utilizes a modern radiometric calibration method (EMoR) for accurate camera response function modeling.
- Extensibility: The framework allows for adaptation and extension with new noise models and camera characteristics.
The project is relatively mature, with a published paper and release of datasets and pretrained models. Recent activity includes code releases and updates to training scripts and noise models, indicating ongoing maintenance. The availability of a comprehensive dataset and pretrained models enhances its reliability as a research tool. Active community engagement is suggested through the provided contact information.
This project benefits researchers working on extreme low-light image denoising, particularly those interested in physics-based approaches and synthetic data generation. It provides a comprehensive toolkit including a realistic noise model, a dedicated dataset, and pretrained models, offering a valuable resource for advancing image processing techniques. It allows researchers to evaluate and compare different denoising algorithms in a controlled environment.