Semantic-Guided Multi-Level Collaborative Fusion Network for Visible and Infrared Images

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Image fusion enhances downstream tasks but current methods lack semantic richness.
• Semantic deficiency in fusion representations limits compatibility with subsequent applications.

Purpose of the Study:

• To propose DSIFuse, a semantic-guided multi-level collaborative fusion network.
• To improve the semantic quality of fused images for better downstream task performance.

Main Methods:

• Leveraging semantic priors and global context via auxiliary segmentation branches.
• Employing a cross-modal feature correction mechanism for semantic alignment.
• Utilizing a three-level interaction strategy and semantic compensation block for enriched representations.

Main Results:

• DSIFuse generates clear fusion images with improved structural consistency and reduced artifacts.
• Fused representations from DSIFuse enhance performance in downstream object detection tasks.
• Experimental validation on public datasets confirms the effectiveness of the proposed network.

Conclusions:

• DSIFuse effectively addresses the semantic deficiency in image fusion.
• The proposed network significantly improves downstream task performance through semantically enriched fusion.
• DSIFuse offers a promising approach for visible-infrared image fusion applications.