摘要
针对目前在火灾预警方面还存在火焰烟雾检测效果差、误报率高、检测速度慢等问题,在YOLOv7框架下提出改进的YOLOv7目标检测算法。首先在数据集建立方面采用的数据集包括Bilkent University公开的数据集和部分自建数据集,共计9 621张图片,并且通过对数据集采用Mosaic数据增强的方式,增加数据的多样性;其次对backbone部分,引入GhostNetV2模块,降低参数量的同时,增加模型检测的准确性,并且在neck部分,引入Slim范式,使用一次性聚合方法来设计跨级部分网络VoV-GSCSP模块。VoV-GSCSP模块降低了计算量和网络结构的复杂性,并保持了足够的精度;最后引入解耦头,增加模型的收敛速度,并且进一步提高网络模型的准确性。总体改进后的网络模型参数量降低了约3.4 MB,mAP@0.5提高了2.4%,并且检测速度也得到提升,满足火焰烟雾实时性的需求。
To address the current problems of poor flame smoke detection and high false alarm rate in fire warning,this paper proposes an improved YOLOv7 target detection algorithm in the framework of YOLOv7.First of all,the dataset used in this paper includes the public dataset of Bilkent University and some self-built datasets,totaling 9621 images,And by using Mosaic data augmentation on the dataset,the diversity of the data is increased.Secondly,for the backbone part,the GhostNetV2 module is introduced to reduce the number of parameters while increasing the accuracy of model detection,and in the head part,the Slim paradigm is introduced to use a one-time aggregation method to design the cross-level partial network The VoV-GSCSP module reduces the computational effort and complexity of the network structure,but maintains sufficient accuracy.Finally,the decouped head is introduced to increase the convergence speed of the model and to further improve the accuracy of the network model.The overall improved network model has reduced the number of parameters by about 3.4 MB,mAP@0.5 improved by 2.4%,and the detection speed has been improved to meet the demand of real-time flame smoke.
作者
谢康康
朱文忠
谢林森
肖顺兴
Xie Kangkang;Zhu Wenzhong;Xie Linsen;Xiao Shunxing(School of Computer Science&.Engineering,Sichuan University of Science&Engineering,Yibin 643000,China)
出处
《国外电子测量技术》
北大核心
2023年第7期41-49,共9页
Foreign Electronic Measurement Technology
基金
四川省科技研发重点项目(2019YFG0200)
四川省科技创新(苗子工程)培育项目(2022049)
四川轻化工大学研究生创新基金(Y2022134)项目资助。
作者简介
谢康康,硕士研究生,主要研究方向为深度学习。E-mail:1141450555@qq.com;朱文忠,教授,主要研究方向为物联网技术,工业缺陷检测,深度学习。E-mail:zwz@suse.edu.cn;谢林森,硕士研究生,主要研究方向为深度学习。E-mail:2420059124@qq.com;肖顺兴,硕士研究生,主要研究方向为深度学习。E-mail:450503661@qq.com。
