Breaking Low-Light Fusion Barrier: Unsupervised Darkness and Noise-Aware Visible and Infrared Image Fusion Network

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Infrared and visible image fusion integrates complementary data but struggles with low-light residual degradations.
• Existing methods often focus on illumination enhancement or noise suppression separately, or require paired data for supervised fusion.
• Multi-task optimization in supervised methods can lead to enhancement-denoising imbalance.

Purpose of the Study:

• To propose BLFusion, an unsupervised framework for darkness- and noise-aware infrared and visible image fusion.
• To address limitations of existing methods by performing illumination enhancement and noise suppression in dedicated stages.
• To enable fusion of complementary information without reliance on high-quality reference images.

Main Methods:

• A Retinex-guided state space model-based decomposition network enhances illumination in dark visible images.
• An unsupervised denoising fusion network jointly performs fusion and denoising using noise correlation disruption (shuffling) and a blind-spot network.
• Noise-free features from both infrared and visible modalities are fused to generate the final output image.

Main Results:

• BLFusion demonstrates superior performance compared to state-of-the-art methods in low-light and noisy conditions.
• The proposed framework shows robust generalization across diverse challenging scenarios.
• A new dataset, MRLL, with 500 infrared-visible pairs for nighttime scenarios was created and utilized.

Conclusions:

• BLFusion effectively overcomes limitations of existing image fusion techniques under low-light and noisy conditions.
• The unsupervised, two-stage approach achieves robust fusion by independently handling illumination enhancement and noise suppression.
• The MRLL dataset facilitates further research in real-world nighttime image fusion.