Cross-Domain Fire Detection Across Indoor and Outdoor Scenes.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Vision-based fire detection models suffer from domain shifts between indoor and outdoor scenes.
• Supervised models trained on a single domain exhibit poor generalization in cross-domain scenarios.
• Existing methods fail to address appearance variations and diverse negative classes in fire detection.

Purpose of the Study:

• To develop a robust cross-domain framework for fire and smoke recognition.
• To improve the generalization of vision-based fire detection models across heterogeneous environments.
• To mitigate negative transfer and enhance domain-invariant feature learning.

Main Methods:

• Curated a large-scale Fire Detection Dataset from multiple public sources for cross-domain benchmarking.
• Employed a tailored cross-domain framework combining adversarial alignment and discrepancy-based statistical alignment.
• Utilized Domain-Adversarial Neural Networks (DANN) and Multi-Kernel Maximum Mean Discrepancy (MK-MMD) for domain adaptation.

Main Results:

• Domain adaptation significantly improved target-domain generalization compared to weak alignment baselines.
• DANN achieved 89.44% accuracy (Indoor → Outdoor) and 79.10% (Outdoor → Indoor).
• MK-MMD attained the best fire-class F1-score of 78.04% (Outdoor → Indoor).

Conclusions:

• Domain alignment is crucial for robust fire detection across diverse deployment environments.
• The proposed framework effectively addresses appearance variations and diverse negative classes.
• This research advances the reliability of vision-based fire detection systems in real-world scenarios.