Factory Fire Detection Using TRA-YOLO Network

Sen Xiang,Shengyong Yin,Gang Yu,Xiaolu Xu,Li Yu

DOI: https://doi.org/10.1109/ccdc62350.2024.10588005

2024-01-01

Abstract:Surveillance analysis based on deep learning is an important approach for fire prevention and control in factories. However, the detection of flames and smoke still faces problems such as false alarms due to the diverse forms of flames and smoke, as well as the background interference. To address these issues, in this paper, we propose a novel fire detection model TRA-YOLO. On the one hand, the network combines the Transformer encoder with CNN to improve the performance in extracting both the global and the local features. On the other hand, an improved ACK-Res2Net module is proposed, which uses multi-scale convolution kernels to adaptively learn features and enhances the network’s feature extraction capability in different scales, improving the detection accuracy of flames and smoke with different deformations. Additionally, the adaptive spatial feature fusion (ASFF) structure is used to suppress interference and enhance the important information, which further strengthen the anti-interference performance. Experiments show that TRA-YOLO achieves an accuracy gain of 4.1% over YOLOv5, and the average precision of flame and smoke is 2.9% higher than that of YOLOv5. Furthermore, the detection speed reaches 50fps, meeting the demand of real-time fire and flame alarming in factories.