Abstract: In response to the need for fast, accurate, and real-time fire detection of historical buildings, this paper builds a dataset specifically for historical building fire detection, which is used for deep learning in historical building fire detection for the first time. By fusing the CBAM attention mechanism and combining it with multi-scale feature fusion, we improve and propose the FireNet-AMF network based on the FireNet network. The fire detection capability of the FireNet-AMF network is verified on the FireNet dataset and the historical building fire detection dataset. The FireNet-AMF network achieves an accuracy of 95.08% for fire detection with the FireNet dataset, an improvement of 1.17% compared to the FireNet network, and an accuracy of 95.62% for experiments on the historical building fire detection dataset we built, which is an improvement of 1.62% compared to the FireNet network. The network ensures a light weight while guaranteeing a high level of historical building fire detection accuracy.

Key words: historical building, fire detection, image classification, FireNet, attention mechanism, multi-scale feature fusion