Symbiotic Evolutionary Learning for Task-Adaptive Infrared and Visible Image Fusion

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Conventional Infrared and Visible Image Fusion (IVIF) methods often fail to guarantee utility for downstream vision tasks.
• Task-driven fusion approaches using fixed weights struggle with conflicting objectives in multi-task scenarios, leading to suboptimal generalization.

Purpose of the Study:

• To propose a symbiotic evolutionary learning framework (EvoFuse) for task-adaptive IVIF.
• To overcome limitations of static loss weights and heuristic criteria in image fusion.
• To enhance the correlation between image fusion and perception tasks.

Main Methods:

• Developed a re-parameterizable fusion architecture for multi-branch capacity during training and compact inference.
• Introduced an evolutionary search mechanism to dynamically evolve task-aware loss-weight configurations.
• Incorporated a saliency discriminative loss to emphasize semantically crucial regions.

Main Results:

• EvoFuse achieved competitive or superior results across eleven datasets for fusion and downstream perception tasks.
• The method demonstrated efficient inference capabilities.
• The evolutionary approach enabled mutual adaptation between fusion and task models.

Conclusions:

• The proposed EvoFuse framework offers a novel approach to task-adaptive IVIF.
• Dynamic weight evolution and Pareto-inspired criteria improve fusion performance and generalization.
• EvoFuse provides a more effective and efficient solution for integrating thermal and visible imagery.