铁路异物入侵深度识别网络快速训练算法  

Fast training algorithm for railway obstacle intrusion based on deep identification network

作　　者：沈悦 SHEN Yue(School of Mechanical,Electronic and Control Engineering,Beijing Jiaotong University,Beijing 100044,China;Key Laboratory of Vehicle Advanced Manufacturing,Measuring and Control Technology(Beijing Jiaotong University),Ministry of Education,Beijing 100044,China)

机构地区：[1]北京交通大学机械与电子控制工程学院 [2]载运工具先进制造与测控技术教育部重点实验室(北京交通大学)

出　　处：《电子设计工程》2019年第22期48-54,共7页Electronic Design Engineering

基　　金：国家重点研发计划(基于空天车地信息协同的轨道交通运营与安全综合保障技术)(2016YFB1200100)

摘　　要：异物侵入铁路限界对铁路系统可靠性造成了极大的威胁。为达到高分类准确率及低模型内存占用率兼备的目的,针对既有技术方法中分类效果、泛化性能较差以及耗时久、模型占用空间大等问题,本文提供了一种快速训练算法,采用网络迁移压缩同时进行的方式,提出基于特征图L1或L2范数的递归式裁剪准则剔除冗余卷积核以压缩网络。对于单个相机新场景的目标分类任务,只需使用在混合场景数据上得到的最优分类网络模型通过压缩和微调训练便可以实现不同场景铁路异物分类的快速训练。实验表明,在基于铁路场景数据库的测试中,该算法可以将原始VGG16模型的参数消耗内存压缩1020倍,在不同的单个相机场景测试样本库上压缩后网络的分类误差最低为0.34%。Railway intrusion poses a great threat to the reliability of the railway system. It is significant to study the intelligent surveillance system that can adapt to various scenes. For achieving high classification accuracy and low network memory,this paper provides a fast training algorithm aimed at the problems of poor performance by the existing methods. The transfer and compression are simultaneously performed,and a recursive pruning algorithm based on the L1 or L2 norm of feature maps to eliminate redundant convolution kernel is proposed. For the classification task of the new single scene,it is only necessary to use the optimal network model obtained on the mixed scenarios data through compression and fine-tuning training,which can achieve rapid training in different scenes. Experiments showed that the algorithm compressed consumption of the original VGG16 model by 1 020 times,and the minimum classification error of the compressed network on the testing samples of different single camera scene is 0.34%.

关 键 词：铁路异物分类 卷积神经网络 VGG16深度网络模型 迁移压缩 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] U215.8[自动化与计算机技术—计算机科学与技术]