当前位置>主页 > 期刊在线 > 智能制造 >

智能制造21年23期

基于 WRLS-YOLOv4 的大型工程机械检测
余长生¹ ,秦伦明¹ ,王悉²,陈鹏¹
(1. 上海电力大学 电子与信息工程学院,上海 200090;2. 北京交通大学 电子信息工程学院,北京 100044)

摘  要:为解决大型工程机械不规范施工导致输电线路出现故障的问题,提出一种改进 YOLOv4 的大型工程机械设备检测方法。当前 YOLOv4 算法存在识别准确率低、漏检率高的缺点,文章借鉴热重启机制思想,设置学习率按余弦函数周期性衰减来减小 loss 值,提高识别准确率。引入标签平滑对正负样本的标签值进行微调整,避免网络过拟合,降低漏检率。实验结果表明,改进后的 WRLS-YOLOv4 算法识别大型工程机械较 Faster-RCNN、SSD、YOLOv3、YOLOv4 效果更好,能够为输电线路进行工程机械监测提供参考依据。


关键词:输电线路;目标检测;深度学习;YOLOv4



DOI:10.19850/j.cnki.2096-4706.2021.23.040


基金项目:国家自然科学基金面上项目 (62073024)


中图分类号:TP391                                          文献标识码: A                                   文章编号:2096-4706(2021)23-0159-04


Detection of Large Construction Machinery Based on WRLS-YOLOv4

YU Changsheng1 , QIN Lunming1 , WANG Xi 2 , CHEN Peng1

(1.College of Electronics and Information Engineering, Shanghai University of Electric Power, Shanghai 200090, China; 2.School of Electronic and Information Engineering, Beijing Jiaotong University, Beijing 100044, China)

Abstract: In order to solve the problem of transmission line failure caused by non-standard construction of large construction machinery, a detection method for large construction machinery and equipment based on improved YOLOv4 is proposed. The current YOLOv4 algorithm has the disadvantages of low recognition accuracy and high missed detection rate. This paper draws on the idea of hot restart mechanism and set the learning rate to decrease the loss value by periodically decaying the cosine function, improve the recognition accuracy. The label smoothing is introduced to fine-tune the label values of positive and negative samples to avoid network overfitting and reduce the missed detection rate. The experimental results show that the improved WRLS-YOLOv4 algorithm recognizes large construction machinery better than Faster-RCNN, SSD, YOLOv3 and YOLOv4, and can provide a reference basis for construction machinery monitoring on transmission lines.

Keywords: transmission line; object detection; deep learning; YOLOv4


参考文献:

[1] 刘建伟,周娅,黄祖钦,等.高压输电线路除冰技术综述 [J]. 机械设计与制造,2012(5):285-287.

[2] LOSHCHIOV I, HUTTER F. SGDR:Stochastic Gradient Descent with Warm Restarts [J/OL].arXiv:608.03983 [cs.LG].[2021- 11-03].http://arxiv.org/abs/1608.03983,2016.

[3] 邱志斌,朱轩,廖才波,等 . 基于目标检测的电网涉鸟故障相关鸟种智能识别 [J]. 电网技术,2022,46(1):369-377

[4] 朱松豪,赵云斌 . 基于半监督生成式对抗网络的异常行为检测 [J]. 南京邮电大学学报,2020,40(4):50-56.

[5] 谢斌红,袁帅,龚大立 . 基于 RDB-YOLOv4 的煤矿 井下有遮挡行人检测 [J/OL]. 计算机工程与应用,(2021-04- 20).https://kns.cnki.net/kcms/detail/detail.aspx?dbcode= CAPJ&dbname=CAPJLAST&filename=JSGG2021041800B&u niplatform=NZKPT&v=3Qinehu7XHJ1uiLWDnHwCMeat_k_ uuMqdTuQF5FxoOrvXKBp-3BWElhF90pYeA29.


作者简介:余长生(1996—),男,汉族,四川内江人,硕士在读,主要研究方向:异物检测;通讯作者:秦伦明(1983—),男, 汉族,江苏靖江人,讲师,博士,主要研究方向:电力设备在线监测、 多源图像视频信息处理等;王悉(1980—),男,汉族,河北邯郸人, 副教授,博士,主要研究方向:列车智能驾驶,模型预测控制;陈鹏,(1997—),男,汉族,山东青岛人,硕士在读,主要研究方向: 深度学习与目标检测。