基于轻量化深度迁移神经网络的电子元器件识别  

Electronic Components Recognition Based on Lightweight Deep Transfer Network

作　　者：夏玉果[1] 董天天 丁晟 XIA Yuguo;DONG Tiantian;DING Shen(School of Microelectronics,Jiangsu Vocational College of Information Technology,Wuxi Jiangsu 214153,China)

机构地区：[1]江苏信息职业技术学院微电子学院,江苏无锡214153

出　　处：《电子器件》2023年第6期1673-1679,共7页Chinese Journal of Electron Devices

基　　金：江苏省“333工程”科研资助项目计划(BRA2020318);江苏省高等学校自然科学研究项目(19KJB510027);江苏信息职业技术学院重点科研课题(JSITKY202201)。

摘　　要：由于电子元器件种类多、体积小,给识别的速度和准确度带来挑战,并且传统的深度神经网络由于网络模型参数多,测试时间较长,无法满足实时运行的需要,针对以上问题,提出一种轻量化深度迁移神经网络的电子元器件识别方法。以MobileNetV3为主干网络,首先利用ImageNet数据集上的预训练模型参数权重对MobileNetV3网络进行初始化,实现模型参数的迁移,然后在自建电子元器件数据集上进行重新训练,并进行微调,最后,利用训练好的模型对电子元器件的类别进行判定,并将其部署在树莓派。实验表明,所提出的方法在测试集上的识别率达92.76%,而且模型的参数量小,实时性好,适合在嵌入式系统中运行。Due to the variety and small size of electronic components,the speed and accuracy of recognition are challenged,and the tra-ditional deep neural network can not meet the needs of real-time operation because of the large network model parameter number and long prediction time.In view of the above problems,a lightweight deep transfer neural network method for electronic component recogni-tion is proposed.Firstly,MobileNetV3 is used as the backbone network and initialized by using the pre-train model weight on the Ima-geNet dataset,then the model parameters are transferred,retrained and fine tune on the electronic component dataset.Finally,the trained model is used to determine the category of electronic components and deploy them in Raspberry Pi.Experiments show that the recognition rate of the proposed method on the test set is 92.76%,and the model has small parameter number and good real-time per-formance,which is suitable for running on the embedded system.

关 键 词：元器件识别 轻量化网络 迁移学习 树莓派 

分 类 号：TN912.3[电子电信—通信与信息系统]