摘要
针对火焰烟雾检测速度慢、难以在嵌入式设备上部署等问题,设计了一种GSN-YOLOv5s的火焰烟雾检测方法。以YOLOv5s为基础,采用网络参数更小、检测速度更快的Ghost Bottleneck结构来替换原网络中的BottleneckCsp结构。考虑到减少网络参数可能会影响检测精度,在各网络层间增加SENet来降低减少参数带来的影响。通过对比实验结果表明,GSN-YOLOv5s的平均精度均值比YOLOv5s高2.4%,检测时间比YOLOv5s快20%。
A GSN-YOLOv5s target detection algorithm is designed to address the problems of slow flame smoke detection and difficulty in deployment on embedded devices.Based on YOLOv5s,the Ghost Bottleneck structure with smaller network parameters and faster detection speed replaces the BottleneckCsp structure of the original network.Considering that the reduction of network parameters may affect the detection accuracy of the whole network,the SENet is added between the network layers to reduce the impact caused by the reduced parameters.A comparison test shows that the GSN-YOLOv5s network improves the detection accuracy by 2.4%compared to the YOLOv5s network,and the GSN-YOLOv5s network reduces the detection time by 20%compared to the YOLOv5s network.
作者
何亚平
苏盈盈
周能扬
张气皓
阎垒
HE Yaping;SU Yingying;ZHOU Nengyang;ZHANG Qihao;YAN Lei(College of Electrical Engineering,Chongqing University of Science and Technology,Chongqing 401331,China)
出处
《重庆科技学院学报（自然科学版）》
CAS
2022年第6期55-59,83,共6页
Journal of Chongqing University of Science and Technology：Natural Sciences Edition
基金
重庆市教育委员会科学技术研究项目“面向智能化工厂转型的通用型指针式仪表识别方法及实现”(KJQN202101510)
重庆科技学院创新项目“基于改进YOLOv7的带钢缺陷检测系统设计及其实现”(YKJCX2220408),”基于Inception V3模型的智能垃圾分类方法及其实现”(YKJCX2120410)。
