Optimized Dual Fire Attention Network and Medium-Scale Fire Classification Benchmark.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Fire Safety Engineering

Background:

• Deep learning models have advanced vision-based fire detection but suffer from high false alarm rates and slow inference speeds, limiting real-world use.
• Existing methods struggle to balance accuracy with computational efficiency for practical fire detection systems.

Purpose of the Study:

• To develop an effective and efficient dual fire attention network (DFAN) for improved vision-based fire detection.
• To address the trade-off between computational cost and accuracy in fire detection systems.
• To introduce a challenging, medium-scale fire classification dataset for advancing fire detection research.

Main Methods:

• Implemented a dual fire attention network (DFAN) incorporating channel and spatial attention mechanisms.
• Optimized the DFAN model using a meta-heuristic approach to reduce parameters and increase inference speed (FPS).
• Created and utilized a diverse fire classification dataset with complex scenarios and imbalanced classes.

Main Results:

• The DFAN model achieved superior performance compared to 21 state-of-the-art methods across four benchmark datasets.
• Optimization resulted in approximately 50% higher frames per second (FPS) values, enhancing efficiency.
• The proposed dataset includes diverse indoor and outdoor fire classes, addressing limitations of traditional datasets.

Conclusions:

• The DFAN offers a robust baseline for high-accuracy, efficient fire detection on edge devices.
• The novel dataset and optimized DFAN model significantly advance the field of vision-based fire detection.
• Publicly available code and dataset encourage further research in intelligent fire detection systems.