Ancient wooden architecture, as a vital carrier of historical and cultural heritage, is highly susceptible to fire hazards due to its material characteristics and the long-term practice of incense burning. Such fire sources are typically small in size and exhibit strong concealment during combustion, making early detection extremely challenging. To address the problem of real-time detection of ultra-small fire sources under complex backgrounds, this study proposes a lightweight fire hazard detection model based on an improved YOLOv13, named SAW-YOLO (Scale-Aware Wavelet YOLO). The model achieves high detection accuracy while maintaining computational efficiency and structural compactness. The main improvements include: (1) introducing an Adaptive Wavelet Downsampling (AWD) module into the backbone network to reduce feature information loss through frequency-domain decomposition, thereby enhancing the extraction of low-level detail features; and (2) embedding a Scale-Aware Cross-Frequency Fusion (SACF) module in the neck network to achieve adaptive multi-scale feature fusion, improving the recognition of small targets.Experiments conducted on the self-constructed Ancient Building Incense Fire Hazard Dataset demonstrate that SAW-YOLOv13 outperforms the original YOLOv13 model, achieving increases of 2.4% in Precision, 1.7% in Recall, 2.9% in mAP50, and 3.2% in mAP50–95, while reducing model parameters by 18.5%. These results confirm that the proposed approach significantly enhances real-time performance and lightweight efficiency without compromising accuracy, providing a reliable technical solution for intelligent fire hazard monitoring in ancient architectural environments.
Hui et al. (2026) studied this question.