改进YOLOv5的森林烟雾及烟雾源检测方法  

Improved forest smoke and smoke source detection for YOLOv5

作　　者：刘晓聪 耿丽清[1,2] 杨耿煌[1,2] 杨亚东 施莹 LIU Xiaocong;GENG Liqing;YANG Genghuang;YANG Yadong;SHI Ying(School of Automation and Electrical Engineering,Tianjin University of Technology and Education,Tianjin 300222,China;Tianjin Key Laboratory of Information Sensing&Intelligent Control,Tianjin University of Technology and Education,Tianjin 300222,China;Tianjin Sino-German University of Applied Sciences,Tianjin 300350,China)

机构地区：[1]天津职业技术师范大学自动化与电气工程学院,天津300222 [2]天津职业技术师范大学天津市信息传感与智能控制重点实验室,天津300222 [3]天津中德应用技术大学,天津300350

出　　处：《天津职业技术师范大学学报》2024年第4期18-24,共7页Journal of Tianjin University of Technology and Education

基　　金：天津市科技计划项目(23YDTPJC00320);天津市高等学校科技发展基金计划项目(2022ZD010);天津市信息传感与智能控制重点实验室基金项目(2023KFJJ04)。

摘　　要：针对森林火灾发生初期烟雾检测模型存在精度差、错检率高以及缺乏烟雾源检测的问题,提出一种基于改进YOLOv5的森林烟雾及烟雾源检测算法。该算法在特征融合区域加入CA注意力机制,以增强模型对输入数据的空间结构理解;通过在骨干网络加入上下文转换器网络(contextual transformer networks,CoTNet),提高网络相邻键的上下文感知能力;后更换损失函数为Shape-IoU提高算法定位精度。结果表明,使用自建数据集训练改进的YOLOv5模型,森林烟雾及烟雾源检测精度达到95.1%。与YOLOv7、YOLOv7-tiny模型相比,改进的YOLOv5模型的准确率、召回率、平均精度均值(mean average precision,mAP)分别提高4.4%~22.7%、9.2%~26.3%、4.6%~26.9%。Aiming at the problems of poor accuracy,high misdetection rate and the lack of smoke source detection in smoke detection model at the early stage of forest fires,this paper proposes a forest smoke and smoke source detection algorithm based on improved YOLOv5.The algorithm integrates a CA attention mechanism in the feature fusion region to enhance the model′s understanding of the spatial structure of the input data.Additionally,a contextual transformer network(CoTNet) is incorporated into the backbone network to improve the context-awareness of network neighboring keys.The loss function is replaced with Shape-IoU to improve the localization accuracy of the algorithm.Experiment results demonstrate that the improved YOLOv5 model,trained on a self-built dataset,achieves an accuracy of 95.1% in forest smoke and smoke source detection.Compared with YOLOv7 and YOLOv7-tiny models,the improved YOLOv5 model exhibits improvements in accuracy,recall,and mean average precision(mAP) values,increasing from 4.4% to 22.7%,from 9.2% to 26.3%,and from 4.6% to26.9%,respectively.

关 键 词：烟雾检测 YOLOv5 CoTNet Shape-IoU CA注意力机制 

分 类 号：S762[农业科学—森林保护学] TP18[农业科学—林学]