基于CNN的高速铁路侵限异物特征快速提取算法
DOI:
作者:
作者单位:

北京交通大学机械与电子控制工程学院北京100044

作者简介:

通讯作者:

中图分类号:

TP391.4TH701U215.8

基金项目:

国家重点研发计划高速铁路系统安全保障课题(2016YFB1200401)项目资助


Fast feature extraction algorithm for highspeed railway clearance intruding objects based on CNN
Author:
Affiliation:

School of Mechanical, Electronic and Control Engineering, Beijing Jiaotong University, Beijing 100044, China

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    高速铁路异物侵限检测系统用来检测是否有异物侵入高速铁路安全限界。为增加系统的可靠性,提出了一种基于卷积神经网络(CNN)的特征快速提取算法。针对特征计算速度缓慢的问题,提出简化的全连接网络结构;针对准确率因简化网络结构而下降的问题,提出将卷积层的卷积核进行预先训练;最后为防止因全连接而导致的对称性特征提取,提出加入稀疏性参数的快速特征提取算法。改进后的卷积神经网络,在保证准确率的基础上加快了计算速度,同时满足了实时性和高准确率的要求。实验表明处理单幅图像的速度为0.15 s,准确率为99.5%。

    Abstract:

    The highspeed railway clearance intrusion detection system is used to detect whether there is object intruding the safety clearance of the highspeed railway. To enhance the reliability of the system, a new CNN based fast feature extraction algorithm is proposed. Aiming at the problem of slow feature calculation speed, a simplified full connected network structure is proposed, and the structure of the neural network is simplified to two full connected convolutional layers. To avoid the accuracy decreasing caused by simplifying network structure, the convolutional kernels of the convolutional layers are pretrained. Finally, in order to prevent the symmetric feature extraction caused by full connection, fast feature extraction algorithm with sparse parameters added is proposed, and the network is trained with sparse coding algorithm. The improved CNN accelerates the calculation speed while ensures the accuracy. At the same time, the new algorithm satisfies the requirements of real time capability and high accuracy. Experiment result shows that the speed of processing single image is 0.15 s and the accuracy is 99.5%.

    参考文献
    相似文献
    引证文献
引用本文

王洋,余祖俊,朱力强,郭保青.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

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:
  • 最后修改日期:
  • 录用日期:
  • 在线发布日期: 2017-07-10
  • 出版日期: