基于改进卷积神经网络的零件识别  

Part Recognition Based on Improved Convolution Neural Network

作　　者：杨乐 甘忠[1] 李义龙[1] 薛超 凌子昊 石望兴 YANG Le;GAN Zhong;LI Yi-long;XUE Chao;LING Zi-hao;SHI Wang-xing(School of Mechanical Engineering,Northwestern Polytechnical University,Xi’an 710072,China)

机构地区：[1]西北工业大学机电学院,陕西西安710072

出　　处：《仪表技术与传感器》2022年第5期82-87,共6页Instrument Technique and Sensor

摘　　要：针对自动化分拣和智能装配的螺纹连接件零件识别问题,提出利用卷积神经网络进行螺纹连接件的识别。首先,构建真实螺纹连接类零件图像数据集,其次,在对经典深度卷积神经网络Inception V3与Xception进行分析的基础上,提出利用全局平均池化代替全连接层改进两种卷积神经网络。最后在迁移学习的基础上,探究了网络冻结层数和数据批量大小对零件识别效果的影响。结果表明:改进的Inception V3和Xception在数据集上的最佳网络冻结层数为70层和20层,最佳数据批量大小为16,零件识别准确率可以达到90%以上,分类速度也得到了提升。Aiming at the identification of threaded connector parts in automatic sorting and intelligent assembly,a convolution neural network was proposed to identify threaded connectors.Firstly,the image data set of real threaded connection parts was constructed.Secondly,based on the analysis of the classical deep convolution neural networks Inception V3 and Xception,it was proposed to use gap instead of the full connection layer to improve the two convolution neural networks.Finally,on the basis of transfer learning,the effects of network freezing layers and batch size on part recognition were explored.The results show that the optimal number of network freezing layers of the improved Inception V3 and Xception on the data set is 70 and 20,the optimal batch size is 16,the part recognition accuracy can reach more than 90%,and the classification speed is improved.

关 键 词：螺纹连接件 零件识别 深度卷积神经网络 迁移学习 数据集 

分 类 号：TH131.3[机械工程—机械制造及自动化]