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 ›› 2024, Vol. 43 ›› Issue (3): 378-383.

• • 上一篇    下一篇

融合注意力机制的轻量级火灾检测模型

曹康壮, 焦双健   

  1. (中国海洋大学 工程学院,山东 青岛 266400)
  • 出版日期:2024-03-15 发布日期:2024-03-15
  • 作者简介:曹康壮(1997- ),男,山东滨州人,中国海洋大学工程学院硕士研究生,主要从事计算机视觉、安全科学与灾害防治方面的研究,山东省青岛市黄岛区三沙路1299号,266400。

A lightweight fire detection model integrating attention mechanism

Cao Kangzhuang, Jiao Shuangjian   

  1. (College of Engineering, Ocean University of China, Shandong Qingdao 266400, China)
  • Online:2024-03-15 Published:2024-03-15

摘要: 基于视觉信息的火灾检测对消防工作具有重要意义,但现阶段相关研究提出的方法大多是基于高性能的硬件设备开展,这限制了相关成果的实际应用。在YOLOv5目标检测算法基础上使用ShuffleNetv2网络为主干构造轻量化模型,并引入SIoU损失函数提高模型目标框的定位精度,同时在模型中添加Shuffle Attention注意力机制,提高在复杂环境下对目标火焰的识别精度。试验证明,与YOLOv5原模型相比,改进后的模型在实现更好识别效果的同时,参数量减少了54.2%,检测速度提升了40.5%。将模型部署嵌入式设备验证其应用效率,结果显示,模型在实现32帧/s检测速度的同时维持了较好的识别效果。

关键词: 卷积神经网络, 火灾检测, YOLOv5, 注意力机制, JetsonNano

Abstract: Based on visual information, fire detection is of great significance to fire protection work. However, most of the methods proposed by relevant research institutions at this stage are based on high-performance hardware devices, which limits the practical deployment and application of relevant results. In response to this, this paper uses ShuffleNetv2 network as the main backbone to construct a lightweight model based on YOLOv5 target detection model, and introduces the SIoU loss function to improve the positioning accuracy of the model's target box. Additionally, the Shuffle Attention module is added to the model to improve its recognition accuracy of flame targets in complex environments. Experiments have shown that compared to the original YOLOv5 model, the improved model not only achieves better recognition results but also reduces the parameter count by 54.2% and improves detection speed by 40.5%. Finally, the model is deployed to embedded devices to verify its application efficiency, and the results show that while maintaining recognition performance, the model achieves a detection speed of 32 f/s.

Key words: convolutional neural networks, fire monitoring, Yolov5, attention module, Jetson Nano