摘要
针对目前火灾中烟雾和火焰的检测研究比较单一,且缺少烟雾火焰数据集、检测模型的抗干扰能力不强等导致检测精度低的问题,提出了一种改进YOLOv5的烟火目标检测算法。首先通过网络爬虫获取火灾图片,然后利用开源软件LabelImg在图片中标注烟雾和火焰目标,构建高质量烟雾火焰数据集,并且利用数据集增强方法扩充数据集,以此来提高模型的泛化能力,其次使用YOLOv5的自适应锚框技术对火焰烟雾数据集进行聚类分析,来获得最优的聚类结果,最后改进YOLOv5中的CSP1模块,对特征进行了融合处理,获取了表达能力更强的特征,提高了模型对火焰烟雾的检测效果。实验结果表明,改进的算法对火焰和烟雾的总体检测准确率和召回率为85.9%和88.4%,检测的平均精度(mean Average Precision,mAP)mAP@0.5为90%,相比原始YOLOv5提高了2%,检测速度为42.2帧/s(Frame Per Second, FPS),能够满足火灾检测的准确性和实时性,可以对火灾进行有效检测。
Aiming at the problems of low detection accuracy due to the lack of smoke and flame data sets and the weak anti-interference ability of the detection model,the current research on smoke and flame detection in fires is relatively single,and an improved YOLOv5 pyrotechnic target detection algorithm is proposed.Firstly,fire pictures are obtained through Web crawlers,then the open source software LabelImg is used to mark smoke and flame targets in the pictures,a high-quality smoke and flame dataset are built,and the dataset enhancement method is used to expand the dataset to improve the generalization ability of the model.Secondly,the adaptive anchor frame technology of YOLOv5 is used to cluster and analyze the flame smoke data set to obtain the optimal clustering result.Finally,the CSP1 module in YOLOv5 is improved,and the features are fused to obtain a more expressive image,which improves the detection effect of the model on flame smoke.The experimental results show that the improved algorithm has an overall detection accuracy and recall rate of 85.9%and 88.4%for flames and smoke,and the average detection precision(mean Average Precision,mAP)mAP@0.5 is 90%,which is improved by 2%compared with the original YOLOv5,and the detection speed is 42.2 frames/s(Frame Per Second,FPS),which can meet the accuracy and real-time performance of fire detection,and can effectively detect fire.
作者
王承茂
黄润才
顾磊欣
WANG Chengmao;HUANG Runcai;GU Leixin(School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201620,China)
出处
《智能计算机与应用》
2025年第5期163-172,共10页
Intelligent Computer and Applications
