关键词: 古建筑, 可燃危险品, YOLOv11, 自集成注意力机制(SEAM), 霍奇斯-莱曼(HL)估计

Abstract:

In order to achieve the early detection and warning of fire hazards, an improved fire source detection model, HSSP-YOLOv11, for identifying flammable hazardous materials was proposed. Firstly, a dedicated dataset for fire hazards consisting of 14 032 images was constructed, covering typical risk categories such as cigarettes and lighters, which provided a solid data foundation for model training and evaluation. Secondly, multiple module optimizations were performed based on YOLOv11 backbone network: the Strip Pooling(SP) module was introduced to expand the feature receptive field. HL Pooling method was designed based on HL estimation to enhance the feature extraction capability and information transmission efficiency of the pooling layer, and the HLSEAM module was developed by replacing the original average pooling layer in the SEAM module with HL Pooling, to improve the robustness of feature aggregation and strengthen the model's ability to focus on features of small targets. Experimental results demonstrate that mean average precision(mAP) of the HSSP-YOLOv11 model reaches 86.2%, which is 1.2% higher than that of the original YOLOv11 model. After embedding the proposed HLSEAM module into YOLOv11, the mAP is increased by 0.3% compared with the original model and by 0.5% compared with the YOLOv11 model embedded with the SEAM module. These results verify that the improved model exhibits superior detection stability and generalization performance for fire hazards in ancient building scenarios. The models designed in this study have achieved more accurate and stable detection of fire hazards in ancient building scenarios.

Key words: ancient architecture, fire hazard materials, YOLOv11, self-ensemble attention mechanisms(SEAM), Hodges-Lehmann(HL) estimation