LR3M: Robust Low-Light Enhancement via Low-Rank Regularized Retinex Model

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Low-light image and video enhancement is crucial for visual quality.
• Existing methods struggle with heavy noise, leading to undesirable visual artifacts.
• Noise suppression is often a separate step, not integrated into the enhancement process.

Purpose of the Study:

• To develop a robust low-light enhancement approach that simultaneously addresses noise suppression.
• To introduce a novel model, the Low-Rank Regularized Retinex Model (LR3M), for integrated enhancement and denoising.
• To extend the LR3M method for effective video low-light enhancement utilizing temporal coherence.

Main Methods:

• Proposed the Low-Rank Regularized Retinex Model (LR3M), incorporating low-rank priors into Retinex decomposition.
• Estimated piecewise-smoothed illumination and noise-suppressed reflectance sequentially.
• Applied LR3M to videos by leveraging inter-frame illumination coherence and reflectance patch similarity for low-rank priors.

Main Results:

• The LR3M method effectively enhances low-light images and videos while suppressing intensive noise.
• Sequential estimation of illumination and reflectance maps avoids residual noise.
• Video enhancement benefits from temporal correspondence, improving overall quality.

Conclusions:

• The proposed LR3M offers a superior approach to low-light enhancement by integrating noise suppression.
• The method demonstrates robust performance across diverse images and videos.
• Achieves state-of-the-art results in both enhancement and denoising quality.