A hybrid multi-scale indirect vision detector algorithm for heritage building fire detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Fire Safety Engineering

Background:

• Indirect vision technology faces challenges in heritage building fire detection, specifically with small target detection and vision imbalance.
• Existing methods struggle to effectively capture flame features across various scales and semantic levels.

Purpose of the Study:

• To propose a novel hybrid multi-scale indirect vision detection algorithm (HMIV-DET) to address the limitations in heritage building fire detection.
• To enhance the detection performance for small targets and improve vision balance in indirect vision systems.

Main Methods:

• Developed the HMIV-DET algorithm featuring Adaptive Indirect Vision Enhancement (AIVE) for vision imbalance, Adaptive Multi-kernel Feature Orchestration Block (AMFOBlock) for multi-scale feature extraction, and Hierarchical Cross-Scale Feature Fusion Network (HCSFPN) for feature interaction.
• Utilized a self-constructed Heritage Building Indirect Vision Fire Dataset for experimental validation.

Main Results:

• The HMIV-DET algorithm demonstrated significant improvements over the YOLOv11 baseline.
• Achieved a 2.8% increase in mAP50 and a 3.3% increase in mAP50-95.
• The algorithm maintains lightweight characteristics while enhancing detection performance.

Conclusions:

• The proposed HMIV-DET algorithm effectively addresses challenges in small target detection and vision imbalance for heritage building fire detection.
• HMIV-DET offers a promising lightweight solution for improved fire detection using indirect vision technology.