Semi-LLIE: Semi-supervised contrastive learning with Mamba-based low-light enhancement

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Low-light image enhancement (LLIE) has advanced significantly.
• A major challenge in LLIE is the scarcity of paired low- and normal-light training data.
• Existing methods struggle with realistic illumination transfer and fine detail restoration in dark regions.

Purpose of the Study:

• To propose a novel semi-supervised framework, Semi-LLIE, for low-light image enhancement.
• To address the limitations of existing methods in utilizing unpaired data and restoring details.
• To improve the naturalness of colors and enhance textures in low-light images.

Main Methods:

• A semi-supervised framework (Semi-LLIE) utilizing unpaired low- and normal-light images via the mean-teacher paradigm.
• A semantic-aware contrastive loss leveraging vision-language representations for illumination alignment.
• A Mamba-based backbone with multi-scale feature learning for enhanced detail restoration.
• A RAM-based perceptive loss for semantic-level texture enhancement.

Main Results:

• Semi-LLIE demonstrates superior performance over existing methods in low-light image enhancement.
• The proposed methods achieve improved quantitative and qualitative metrics.
• Enhanced naturalness in color and restored fine-grained details in dark regions.

Conclusions:

• The Semi-LLIE framework effectively addresses the challenge of limited paired data in low-light image enhancement.
• The novel loss functions and backbone architecture significantly improve illumination transfer and detail restoration.
• Semi-LLIE offers a promising solution for high-quality low-light image enhancement.