主管：中华人民共和国应急管理部
主办：应急管理部天津消防研究所
ISSN 1009-0029  CN 12-1311/TU
消防科学与技术

消防科学与技术 ›› 2023, Vol. 42 ›› Issue (8): 1113-1116.

• • 上一篇    下一篇

基于CA-Res注意力机制的YOLOv5图像火灾检测算法

李晓旭1,2,3, 李泊宁1,2,3, 张曦1,2,3, 于春雨1,2,3   

  1. （1. 应急管理部沈阳消防研究所，辽宁 沈阳 110034；2. 辽宁省火灾防治技术重点实验室，辽宁 沈阳 110034；3. 消防与应急救援国家工程研究中心，辽宁 沈阳 110034）
  • 出版日期:2023-08-15 发布日期:2023-08-15
  • 作者简介:作者简介：李晓旭(1995- )，女，应急管理部沈阳消防研究所研究第二研究室实习员，主要从事火灾探测方面的研究，辽宁省沈阳市皇姑区文大路218-20号，110034。
  • 基金资助:
    基金项目：国家重点研发计划项目（2021YFC3001605）

Image fire detection algorithm based on YOLOv5 with CA-Res attention mechanism

Li Xiaoxu1,2,3, Li Boning1,2,3, Zhang Xi1,2,3, Yu Chunyu1,2,3   

  1. (1. Shenyang Fire Science and Technology Research Institute of MEM, Liaoning Shenyang 110034, China;2. Liaoning Key Laboratory of Fire Prevention Technology, Liaoning Shenyang 110034, China;3. National Engineering Research Center of Fire and Emergency Rescue, Liaoning Shenyang 110034,China)
  • Online:2023-08-15 Published:2023-08-15

摘要： 为解决人工和传感器火灾检测方法中存在的精度低、速度慢等问题，在分析火灾图像特征的基础上，设计了一种融入可调节坐标残差注意力的改进YOLOv5多路火灾检测算法。该算法在YOLOv5网络模型上进行改进，可以自动提取和学习图像的特征。首先，通过坐标注意力机制，将位置信息嵌入到通道注意力中，使网络能够获得更大范围的信息，提高了对火灾图像的检测精度。其次，通过残差连接保留火焰的初始特征，将初始特征与坐标注意力特征进行自适应结合，实现更好的识别效果。最后，在多路摄像头捕获的实时视频数据上进行火灾的检测与识别，具有较高的准确率。该火灾检测方法可以有效识别和检测火灾早期产生的火焰信息，减少在火灾早期错过最佳扑救时间造成的损失。

关键词: 火灾检测, 注意力机制, 目标检测, 深度学习

Abstract: Abstract: To solve the problems of low accuracy and slow speed in manual and sensor fire detection methods, based on the analysis of fire features, this paper designs an improved YOLOv5 multiplex fire detection algorithm incorporating adjustable coordinate residual attention. The YOLOv5 network adopted in our algorithm can automatically extract and learn the features of the images. First, the location information is embedded into the channel attention through the coordinate attention mechanism, which enables the network to obtain a larger range of information and improves the detection accuracy of fire images. Secondly, the initial features of the flame are retained through residual connections, and the initial features are adaptively combined with the coordinate attention features to achieve better performance. Finally, the detection and recognition with high accuracy are performed on real?time video streams captured by multiple cameras. The method proposed in this paper can identify and detect not only the flame information generated by fire, but also the smoke generated in the early stage of fire to reduce the loss of missing the best remediation time in the early stage of fire.

Key words: fire detection, attention mechanism, object detection, deep learning