Research on Enhancing Fire Detection Performance in Ancient Architecture Under Occlusion Scenarios Based on YOLO-AR

Area of Science:

• Computer Vision
• Artificial Intelligence
• Cultural Heritage Preservation

Background:

• Fire detection in ancient architecture is challenging due to complex scenes and structural peculiarities.
• Traditional methods struggle with occlusions and unique building features.

Purpose of the Study:

• To propose YOLO-AR, an improved YOLOv8-based algorithm for accurate fire detection in ancient structures.
• To enhance flame and smoke feature capture and improve localization in occluded scenarios.

Main Methods:

• Utilized an improved YOLOv8 framework incorporating the Convolutional Block Attention Module (CBAM).
• Introduced Repulsion Loss to optimize bounding box accuracy in dense and occluded conditions.
• Trained and evaluated on a custom dataset of 15,847 ancient architecture images.

Main Results:

• YOLO-AR demonstrated superior performance over mainstream algorithms in Precision, Recall, and mAP.
• Achieved a detection precision of 90.7% and a recall rate of 89.7%.
• Outperformed comparative methods on a specialized ancient architecture dataset.

Conclusions:

• YOLO-AR offers an efficient and reliable visual detection solution for early fire warning systems in ancient buildings.
• The algorithm contributes significantly to the preservation of cultural heritage through advanced fire detection.
• The study highlights the effectiveness of attention mechanisms and specialized loss functions in challenging visual detection tasks.