摘要
针对我国发生多起雾炮车违规喷雾空气自动监测站事件,雾炮车数据集空白,雾炮车所喷的雾的形状各异难以标注,监测实时要求高,准确率高等问题。本研究建立了雾炮车喷雾数据集,设计了一种雾的标注方式,并在YOLOv5网络基础上,提出了一个嵌入注意力机制的轻量级雾炮车检测网络。首先,利用K-means++计算出更适合任务的锚框;其次,嵌入注意力机制(CA)模块,用于提升网络特征提取能力;然后将Neck处Conv修改GSConv,并将C3模块更换为GSC3模块,以降低模型参数;最后,将NMS替换为Soft NMS,用于减少漏检增强检测稳定性。实验结果表明,所提标注方式较其他标注方式提升了总体mAP13%;所提网络的参数量仅为YOLOv5s的83%并达到67.8%的mAP。与主流目标检测网络相比,所提网络在保持精度提升的同时降低了参数量,并保持了检测速度。
In response to the multiple incidents of fog cannon vehicles spraying in violation of regulations at automated air monitoring stations in China,the lack of a data set for fog cannon vehicles,the difficulty in labeling the various shapes of fog sprayed by fog cannon vehicles,the high real-time monitoring demands,and the need for high accuracy,this study established a spray data set for fog cannon vehicles.A method for annotating fog was designed,and a lightweight fog cannon vehicle detection network embedded with an attention mechanism was proposed,based on the YOLOv5 network.Firstly,the anchor box that was most suitable for the task was calculated using K-means++.Secondly,an attention mechanism(CA)module was embedded to enhance the feature extraction capability of the network.The Conv at the Neck was then modified to GSConv,and the C3 module was replaced with the GSC3 module,reducing the model parameters.Finally,NMS was replaced with Soft NMS to reduce the miss rate and enhance the stability of detection.The experimental results showed that compared to other annotation methods,the proposed annotation method increased the overall mAP by 13%.The parameter volume of the proposed network was only 83%of YOLOv5s and achieved an mAP of 67.8%.Compared with the mainstream target detection network,the proposed network reduced the volume of parameters while maintaining an increase in accuracy and the speed of detection.
作者
钟逸伦
刘黎志
李赢杰
ZHONG Yilun;LIU Lizhi;LI Yingjie(Hubei Key Laboratory of Intelligent Robotics,Wuhan Institute of Technology,Wuhan 430205,China;School of Computer Science and Engineering,Wuhan Institute of Technology,Wuhan 430205,China;Wuhan Rate Technology Co.,Ltd.,Wuhan 430223,China)
出处
《环境工程学报》
CAS
CSCD
北大核心
2024年第5期1434-1441,共8页
Chinese Journal of Environmental Engineering
基金
智能机器人湖北省重点实验室创新基金资助项目(HBIRL202207)
湖北省教育厅科学研究计划指导性项目(B2017051)。
