基于深度学习的电解电容表面视觉检测  

Visual inspection of electrolytic capacitor surface based on deep learning

作　　者：刘宏利[1] 于斌 LIU Hongli;YU Bin(Tianjin Key Laboratory of Complex System Control Theory and Application,Tianjin University of Technology,Tianjin 300384,China;School of Electrical Engineering and Automation,Tianjin University of technology,Tianjin 300384,China)

机构地区：[1]天津理工大学天津市复杂系统控制理论与应用重点实验室,天津300384 [2]天津理工大学电气工程与自动化学院,天津300384

出　　处：《天津理工大学学报》2024年第1期77-83,共7页Journal of Tianjin University of Technology

基　　金：天津市自然科学基金项目(17JCTPJC53100)。

摘　　要：在工业环境下的电解电容检测中,由于光照不均和其他噪声干扰,单步骤多盒探测器(single shot multibox detector,SSD)算法易出现检测效果不佳、漏检率高等问题。为提高电解电容检测的准确性,在原SSD算法的基础上提出一种基于深度学习的第3版轻量级网络的单步骤多盒探测器(mobile networks version 3-single shot multibox detector,MobileNetV3-SSD)算法。采用改进的视觉几何群网络(visual geometry group networks 16,VGG16)模型提取输入图像的特征信息,并利用网络中4个有效的特征图生成目标的众多先验框,经非极大值抑制(non maximum suppression,NMS)进行筛选确定边框位置,从而完成电容的表面识别。本算法还引入了迁移学习提高了算法模型本身的稳定性和泛化能力。相比于SSD和快速区域卷积神经网络(faster region convolutional neural networks,Faster R-CNN)算法,文中的算法识别精度和运行速度都得到了提升,平均精度达92.71%,耗时仅为25 ms,为电解电容的出厂检测提供很好的技术支持。In the detection of electrolytic capacitor in industrial environments,the Single Shot MultiBox Detector(SSD)algorithm is proposed to solve problems such as poor detection effect and high missed detection rate due to uneven illumination and other noise interference.In order to improve the accuracy of electrolytic capacitor detection,a Mobile Networks Version 3-Single Shot MultiBox Detector(MobileNetV3-SSD)algorithm based on the original SSD algorithm and the deep learning 3rd version is proposed.The improved Visual Geometry Group Networks 16(VGG16)model is used to extract the feature information of the input image,and four effective feature maps in the network are used to generate numerous priori frames of the target,which are screened by the Non Maximum Suppression(NMS)to determine the position of the frame,and to complete the surface recognition of capacitors.The algorithm also introduces the transfer learning to improve the stability and generalization ability of the algorithm model.Experiment results show that the recognition accuracy and running speeds of the proposed algorithm have been improved too,such as the average accuracy of 92.71% and a time-consuming of only 25 ms,the study can provide good technical support for the factory inspection of electrolytic capacitors.

关 键 词：电解电容 深度学习 迁移学习 识别 先验框 

分 类 号：TP317.4[自动化与计算机技术—计算机软件与理论]