Calibration-Free Raw Image Denoising via Fine-Grained Noise Estimation

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Deep learning has advanced image denoising, but real-world applications require better noise modeling and estimation.
• Current approaches focus on realistic noise models for training data or noise level estimation for denoisers.

Purpose of the Study:

• To develop an innovative pipeline for noise estimation and realistic noise synthesis.
• To bridge noise modeling and noise estimation strategies for improved denoising.

Main Methods:

• Integration of a fine-grained statistical noise model and contrastive learning.
• Utilizing unique data augmentation techniques to enhance learning.
• Developing a calibration-free pipeline for estimating noise parameters and synthesizing realistic noise.

Main Results:

• The pipeline effectively estimates noise parameters for unknown cameras without calibration frames or paired data.
• Demonstrated effectiveness in both normal and low-light scenarios.
• Successful synthesis of camera-specific noise distributions for realistic training data.

Conclusions:

• The proposed pipeline offers an adaptable and effective solution for real-world noise synthesis and denoising.
• Estimating parameters for fine-grained noise models enhances applicability in challenging low-light conditions.
• The calibration-free approach simplifies noise modeling for diverse imaging sensors.