Less Is More: Infrared and Visible Images Fusion via Semantic-Guided Mixture of Multi-Feature Experts.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Infrared and visible image fusion (IVIF) is crucial for scene understanding, but current methods struggle with redundancy and diverse scenarios.
• Static deep learning models often fail to effectively handle modality redundancy, leading to noise and artifacts in fused images.

Purpose of the Study:

• To develop an advanced IVIF method that dynamically integrates complementary features while eliminating redundancy.
• To align image fusion with downstream semantic segmentation tasks for improved performance.

Main Methods:

• Proposed a semantic-guided mixture of multi-feature experts (SD-MoMFE) for IVIF.
• Employed an expert routing mechanism for dynamic selection of specialized feature processing networks.
• Integrated a segmentation head to semantically guide feature selection for improved fusion and segmentation alignment.

Main Results:

• Demonstrated the effectiveness of the proposed method on five benchmark datasets.
• Achieved superior performance in both image fusion quality and subsequent semantic segmentation accuracy.
• The method successfully handles redundancy and preserves significant features from both modalities.

Conclusions:

• The semantic-guided mixture of multi-feature experts offers a robust solution for IVIF.
• Dynamic feature selection and semantic guidance enhance fused image quality and downstream task performance.
• This approach advances the capabilities of image fusion for complex real-world applications.