CERL: A Unified Optimization Framework for Light Enhancement With Realistic Noise

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Real-world low-light images suffer from spatially variant sensor noise, challenging conventional denoising methods.
• Existing low-light enhancement techniques often ignore or inadequately address real-world noise during the enhancement process.

Purpose of the Study:

• To develop a unified framework that seamlessly integrates low-light enhancement and noise suppression for real-world images.
• To address the limitations of current methods that treat noise removal as a separate step.

Main Methods:

• Developed Coordinated Enhancement for Real-world Low-light Noisy Images (CERL), a physics-grounded optimization framework.
• Customized a self-supervised denoising model adaptable without clean ground-truth images.
• Improved a state-of-the-art backbone for light enhancement and integrated it into a plug-and-play optimization.

Main Results:

• CERL consistently produced visually pleasing and artifact-free results compared to state-of-the-art methods.
• Evaluated on standard benchmarks and a new realistic low-light mobile photography dataset (RLMP).
• Demonstrated superior performance in handling heavy, realistic noise present in mobile-captured images.

Conclusions:

• The proposed unified framework effectively addresses the dual challenges of low-light enhancement and real-world noise suppression.
• CERL offers a robust and adaptable solution for improving image quality in challenging low-light conditions.
• The developed RLMP dataset and open-source code facilitate further research in this domain.