Learning Physics-Informed Noise Models from Dark Frames for Low-Light Raw Image Denoising.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Current low-light raw image denoising relies heavily on synthetic data.
• Existing noise modeling methods (physics-based and learning-based) have limitations in accuracy and data dependency.
• Effective noise modeling is crucial for the practical application of denoising algorithms.

Purpose of the Study:

• To develop a novel strategy for training low-light denoising methods by learning noise models from dark frames, reducing reliance on paired real data.
• To introduce an efficient physics-informed noise neural proxy (PNNP) for accurate real-world sensor noise modeling.
• To improve the effectiveness and practicality of synthetic data for low-light raw image denoising.

Main Methods:

• Proposed a strategy to learn noise models from dark frames, eliminating the need for paired real data.
• Introduced the physics-informed noise neural proxy (PNNP) integrating physical priors into neural networks.
• Developed three key techniques: physics-guided noise decoupling (PND), physics-aware proxy model (PPM), and differentiable distribution loss (DDL).

Main Results:

• PNNP effectively characterizes real-world sensor noise distributions.
• PND flexibly handles varying noise levels, reducing modeling complexity.
• PPM and DDL enhance the accuracy and precision of synthetic noise modeling by incorporating physical constraints and explicit distribution supervision.
• Demonstrated superior performance in practical low-light raw image denoising tasks on public datasets.

Conclusions:

• The proposed dark frame-based noise modeling strategy significantly breaks data dependency for training denoising methods.
• PNNP offers a powerful and efficient approach to approximating real-world sensor noise.
• The method shows significant potential for advancing practical low-light image denoising applications.