|本期目录/Table of Contents|

[1]沈 悦.铁路异物入侵深度识别网络快速训练算法[J].电子设计工程,2019,27(22):48-54.
 SHEN Yue.Fast training algorithm for railway obstacle intrusion based on deep identification network[J].SAMSON,2019,27(22):48-54.
点击复制

铁路异物入侵深度识别网络快速训练算法(PDF)
分享到:

《电子设计工程》[ISSN:1674-6236/CN:61-1477/TN]

卷:
27
期数:
2019年22期
页码:
48-54
栏目:
计算机技术应用
出版日期:
2019-11-20

文章信息/Info

Title:
Fast training algorithm for railway obstacle intrusion based on deep identification network
文章编号:
1674-6236(2019)22-0048-07
作者:
沈 悦12
(1. 北京交通大学 机械与电子控制工程学院,北京 100044;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),北京 100044)
Author(s):
SHEN Yue12
(1. School of Mechanical, Electronic and Control Engineering, Beijing Jiaotong University, Beijing 100044, China;2. Key Laboratory of Vehicle Advanced Manufacturing, Measuring and Control Technology (Beijing Jiaotong University),Ministry of Education, Beijing 100044, China)
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
Keywords:
railway obstacle classification convolution neural network (CNN) VGG16 deep network model transfer compression
分类号:
TP391.4; U215.8
DOI:
-
文献标志码:
A
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1 020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%。
Abstract:
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%.

参考文献/References:

[1] 郭保青,朱力强,史红梅. 基于快速DBSCAN聚类的铁路异物侵限检测算法[J]. 仪器仪表学报, 2012,33(2): 241-247.[2] 王尧,余祖俊,王中卫,等. 基于FPGA的铁路异物检测算法的硬件实现[J]. 铁道学报, 2016(3): 84-91.[3] 史红梅,柴华,王尧,等. 基于目标识别与跟踪的嵌入式铁路异物侵限检测算法研究[J]. 铁道学报, 2015(7): 58-65.[4] 张成. 基于深度学习的铁路异物侵限检测系统研究[D]. 北京: 北方工业大学, 2017.[5] 刘洪公. 基于深度学习理论的铁路异物检测与识别技术研究[D]. 河北: 石家庄铁道大学, 2017.[6] 李晓蕊. 基于迁移学习的铁路异物检测算法研究[D]. 北京: 北京交通大学, 2018.[7] 刘晨,曲长文,周强, 等. 基于卷积神经网络迁移学习的SAR图像目标分类[J]. 现代雷达, 2018, 40(3):38-42.[8] 王洋,余祖俊,朱力强,等. 基于CNN的高速铁路侵限异物特征快速提取算法[J]. 仪器仪表学报, 2017,38(5):1267-1275.[9] 曹文龙,芮建武,李敏. 神经网络模型压缩方法综述[J]. 计算机应用研究, 2019,36(3):649-656.[10]Han S, Mao H, Dally W J. Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding[J]. Fiber, 2015,56(4): 3-7.[11]Hinton G,Vinyals O,Dean J. Distilling the Knowledge in a Neural Network[J]. Computer Science, 2015,14(7):38-39.[12]Srinivas, Suraj Babu, R. Venkatesh. Data-free Parameter Pruning for Deep Neural Networks[C] //Proceedings of the British Machine Vision Conference 2015, BMVC 2015. Swansea, UK, 2015:1-12.[13]Denil M, Shakibi B, Dinh L, et a1. Predicting parameters in deep learning[C]// NIPS’13: Proceedings of the 2013 International Conference on Neural Information Processing System. North Miami Beach, FL, USA: Curran Associates, 2013:2148-2156.[14]Novikov A,Podoprikhin D,Osokin A,et al. Tensorizing neural networks[C]// Advances in Neural lnformation Processing Systems. Montreal, Quebec, Canada, 2015:442-450.[15]Courbariaux M,Bengio Y,David J P. Binary Connect: training deep neural networks with binary weights during propagations[C]// International Conference on Neural Information Processing Systems. MIT Press, 2015: 3123-3131.[16]Li H, Kadav A, Durdanovic I, et al. Pruning Filters for Efficient ConvNets[C]// International Conference on Learning Representations. Toulon, 2017: 230-242.[17]杨娟,曹浩宇,汪荣贵,等. 基于语义DCNN特征融合的细粒度车型识别模型[J]. 计算机辅助设计与图形学学报, 2019(1):141-157.[18]曾燕,陈岳林,蔡晓东. 结合全局与局部池化的深度哈希人脸识别算法[J]. 西安电子科技大学学报, 2018,45(5):169-175.[19]吴军,邱阳,卢忠亮.全局融合卷积神经网络的边缘分类的人脸性别识别[J]. 现代电子技术,2019(18):177-181,186[20]苗军,李凯,许少武.基于卷积神经网络多层特征融合的目标跟踪[J]. 现代电子技术,2018(24):114-116.

相似文献/References:

[1]高 飞,雷 涛.基于深度学习的超分辨率重建算法改进[J].电子设计工程,2019,27(06):64.
 GAO Fei,LEI Tao.Improvements of super?resolution algorithms based on deep learning[J].SAMSON,2019,27(22):64.
[2]赵东波,李 辉.基于椭圆皮肤模型和深度学习的静态手势识别[J].电子设计工程,2019,27(04):162.[doi:10.14022/j.cnki.dzsjgc.2019.04.035]
 ZHAO Dong?bo,LI Hui.A static gesture recognition method based on Elliptical skin model and deep learning[J].SAMSON,2019,27(22):162.[doi:10.14022/j.cnki.dzsjgc.2019.04.035]
[3]柳 畅,徐小杰.基于卷积神经网络的RGB-D图片分类[J].电子设计工程,2018,26(24):159.
 LIU Chang,XU Xiaojie.RGB-D object classification using convolutional neural networks[J].SAMSON,2018,26(22):159.
[4]张久鑫.基于卷积神经网络的室内行人识别与定位[J].电子设计工程,2018,26(24):140.
 ZHANG Jiuxin.Indoor pedestrian recognition and positioning based on CNN[J].SAMSON,2018,26(22):140.
[5]李一松.基于卷积神经网络的多光谱图像多标签场景分类[J].电子设计工程,2018,26(23):25.
 LI Yisong.CNN based multilabel scene classification method for multispectral image[J].SAMSON,2018,26(22):25.
[6]赵丹新,孙胜利.基于ResNet的遥感图像飞机目标检测新方法[J].电子设计工程,2018,26(22):164.
 ZHAO Danxin,SUN Shengli.A new method for target detection of remote sensing image based on ResNet.computer engineering and applications[J].SAMSON,2018,26(22):164.
[7]王宗伟.有限训练样本下基于深度领域自适应的机器人物体抓取[J].电子设计工程,2018,26(20):33.
 WANG Zongwei.Deep domain adaptation for grasping novel objects with limited training samples[J].SAMSON,2018,26(22):33.
[8]文 宇.基于CNN-LSTM网络分析金融二级市场数据[J].电子设计工程,2018,26(17):75.
 WEN Yu.Use CNN-LSTM network to analyze secondary market data[J].SAMSON,2018,26(22):75.
[9]凌佳佳,袁晓兵.联合噪声分类和掩码估计的语音增强方法[J].电子设计工程,2018,26(17):30.
 LING Jiajia,YUAN Xiaobing.Joint noise classification and mask estimation for speech enhancement[J].SAMSON,2018,26(22):30.
[10]张营营.生成对抗网络模型综述[J].电子设计工程,2018,26(05):34.
 ZHANG Yingying.A survey on generative adversarial networks[J].SAMSON,2018,26(22):34.

备注/Memo

备注/Memo:
收稿日期:2019-04-04 稿件编号:201904023基金项目: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)作者简介:沈 悦(1994—),女,回族,安徽滁州人,硕士研究生。研究方向:轨道交通检测图像处理和机器视觉检测。
更新日期/Last Update: 2019-11-21