摘要
研发了一套基于深度学习的汽车线束缺陷检测系统。基于YOLOv3改进的Pr-YOLOv3算法来检测线束端子接插件缺陷,将主干提取网络替换成ResNet50,提高特征提取能力,减少参数量,吸收多尺度预测方式和特征融合方面的优势,将主干提取网络与FPN特征金字塔进行对接,丰富了特征的表达能力。用改进的YOLOv3模型进行训练,准确率可达98.61%,Recall指数可达98.6%。
A deep-learning based automotive wiring harnesses defect detection system is developed.The Pr-YOLOv3 algorithm based on improved YOLOv3 is used to detect defects in wiring harness terminal connectors,and the backbone extraction network is replaced with ResNet50,which improves the feature extraction capability and reduces the number of parameters.Drawing on the advantages in multi-scale prediction methods and feature fusion,the backbone extraction network is interfaced with the FPN feature pyramid,which enriches the feature expression ability.Trained with the improved YOLOv3 model,the accuracy can reach 98.61%and the Recall index can reach 98.6%.
作者
程晓颖
李海生
吕旭波
Cheng Xiaoying;Li Haisheng;Lv Xubo(School of Mechanical Engineering,University of Electronic Science and Technology of China,Chengdu,Sichuan 611731,China;School of Mechanical Engineering,Zhejiang Sci-tech University;Zhejiang SIMITEK Auto-Electronics CO.,LTD)
出处
《计算机时代》
2023年第12期29-33,共5页
Computer Era
关键词
射线无损检测
线束端子缺陷检测
卷积神经网络
YOLOv3
radiographic non-destructive testing
wire harness terminal defect detection
convolution neural network
YOLOv3