改进的VGG-16卷积神经网络算法在丁腈橡胶片材识别中的应用  

Application of improved VGG-16 convolutional neural network algorithm on nitrile rubber sheet recognition

作　　者：何琛 李云红[1] 谢蓉蓉 HE Chen;LI Yunhong;XIE Rongrong(School of Electronics and Information,Xi’an Polytechnic University,Xi’an 710048,China)

机构地区：[1]西安工程大学电子信息学院,陕西西安710048

出　　处：《西安工程大学学报》2021年第2期41-47,共7页Journal of Xi’an Polytechnic University

基　　金：陕西省科技厅青年科学基金项目(2019JQ-255);西安市科技局高校人才服务企业项目(2019217114GXRC007CG008-GXYD7.2,2019217114GXRC007CG008-GXYD7.8);国家级大学生创新创业训练计划项目(S202010709003)。

摘　　要：针对VGG-16卷积神经网络识别丁腈橡胶片材时,出现了过拟合、参数量大、准确率较低的问题,提出在缩减原网络深度基础上改进的VGG-16卷积神经网络识别算法。通过嵌入多分辨率分组卷积、混合池化取代最大池化、增加自适配归一化(switchable normalization,SN)的方法,优化了网络结构。实验结果表明:该方法训练时未出现过拟合,参数量约下降至VGG-16的0.098%,相对仅缩减深度的VGG-16网络,识别准确率提高了7.17%,算法可应用于某些固体火箭发动机内绝热层材料的识别。In order to solve the problems of over fitting,large amount of parameters and low accuracy of VGG-16 convolutional neural network in nitrile rubber sheet recognition,an improved VGG-16 convolutional neural network recognition algorithm based on reducing the depth of the original network was proposed.The network structure was optimized by embedding multi-resolution packet convolution,replacing maximum pooling with mixed pooling,and adding switched normalization.The experimental results show that there is no over fitting in the training process,and the parameters are approximately reduced to 0.098%of VGG-16.Compared with VGG-16 network which only reduced the depth,the recognition accuracy was improved by 7.17%,which could be applied to the material recognition of some solid rocket motors with inner insulation.

关 键 词：丁腈橡胶片材 卷积神经网络 VGG-16 归一化 混合池化 多分辨率分组卷积 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]