MSSA-Net: Multi-Modal Structural and Semantic-Adaptive Network for Low-Light Image Enhancement

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Low-light image enhancement (LLIE) faces challenges with degraded structures and semantic ambiguity in extreme darkness.
• Existing methods struggle with structural inconsistency and semantic drift due to uniform strategies or static prompts.

Purpose of the Study:

• To propose a novel Multi-Modal Structural and Semantic-Adaptive Network (MSSA-Net) for improved LLIE.
• To address limitations of current LLIE methods in preserving structural integrity and semantic accuracy.

Main Methods:

• Developed a Multi-Scale Self-Refinement Block (MSRB) for progressive visible feature enhancement.
• Introduced a pseudo-infrared structural prior and Structure-Guided Cross-Attention (SGCA) for noise-insensitive geometric cues.
• Integrated a Large Multi-modal Model (LMM)-Driven Scene-Adaptive Attention mechanism for semantic embedding injection.

Main Results:

• MSSA-Net demonstrated superior performance across multiple benchmarks.
• The network achieved significant improvements in structural fidelity and brightness recovery.
• Enhanced semantic naturalness was observed in the enhanced low-light images.

Conclusions:

• MSSA-Net effectively balances structural restoration and semantic accuracy in LLIE.
• The proposed structure-anchored paradigm and multi-modal integration offer a robust solution for challenging low-light conditions.