+高级检索
基于深度学习的盾构隧道衬砌病害识别方法
DOI:
作者:
作者单位:

作者简介:

通讯作者:

基金项目:


A Method of Disease Recognition for Shield Tunnel Lining Based on Deep Learning
Author:
Affiliation:

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
    摘要:

    隧道衬砌病害的检测是隧道维护和保障运营安全的重要环节.以基于CCD线阵相机移动式地铁衬砌病害检测系统的采集图片为研究对象,利用计算机科学最前沿的深度学习方法,提出了一种完全区别于传统手段的隧道病害识别方法,通过提取并建立隧道病害样本库,搭建深度学习框架,利用深度卷积神经网络(Convolutional Neural Network,CNN)训练样本,建立隧道衬砌特征图像分类系统.针对既有的CNN模型GoogLeNet,采用优化的卷积核,并改进了其inception模块与网络结构,获得了准确率超过95%的网络模型.通过实例对目前流行的深度学习框架(Caffe与Torch)以及图像对比度增强处理方法(如直方图均衡化处理(Histogram Equalization,HE))进行了测试.测试结果表明,深度学习方法用于隧道衬砌图像处理,具有准确率高,速度快,可扩展性好等特点,特别是对背景复杂条件下的图像处理更具鲁棒性.

    Abstract:

    Diseases detection and maintenance of tunnel lining is an important link to ensure the safety of tunnel in operation. Based on the images captured by CCD linear array camera in Movable Tunnel Inspection System, a new method was proposed. It is inspired by cutting-edge computer science-deep learning and different from the traditional ones entirely. The main idea is as follows: a) extracting lining diseases and establishing feature map database; b) building deep learning framework; c) training samples with convolutional neural network; and d) establishing a classification system of gray scale feature maps of tunnel lining. Aiming at CNN model GoogLeNet, inception module and overall architecture were improved by using improved convolutional kernels. The best test-set accuracy is over 95%. At the same time, the influence of different deep learning frameworks (Caffe and Torch) and image contrast enhancement method (such as histogram equalization, HE) were tested with examples. The results show that the deep learning method is applicable to the tunnel lining diseases detection. The advantages are high accuracy, high speed, good extensibility and very robust in complex cases.

    参考文献
    相似文献
    引证文献
文章指标
  • PDF下载次数:
  • HTML阅读次数:
  • 摘要点击次数:
  • 引用次数:
引用本文

薛亚东,李宜城.基于深度学习的盾构隧道衬砌病害识别方法[J].湖南大学学报:自然科学版,2018,45(3):100~109

复制
历史
  • 收稿日期:
  • 最后修改日期:
  • 录用日期:
  • 在线发布日期: 2018-03-27
  • 出版日期:
作者稿件一经被我刊录用,如无特别声明,即视作同意授予我刊论文整体的全部复制传播的权利,包括但不限于复制权、发行权、信息网络传播权、广播权、表演权、翻译权、汇编权、改编权等著作使用权转让给我刊,我刊有权根据工作需要,允许合作的数据库、新媒体平台及其他数字平台进行数字传播和国际传播等。特此声明。
关闭