摘要
探地雷达是隧道衬砌检测常用方法之一,但是后期大部分数据解释工作需要人工完成。为快速检测隧道施工过程中的初期支护质量,研究了一种基于卷积神经网络的自动识别算法,对初期支护的钢拱架排列和脱空进行检测。该算法使用实测探地雷达图像作为训练数据,利用Mask R-CNN深度学习框架和ResNet101卷积神经网络进行探地雷达图像特征提取,对钢拱架和脱空两类对象进行分类和位置回归。该网络首先在COCO数据集上进行预训练,得到一组预训练权重,然后利用200张处理后的探地雷达图像对权重进行微调,最后用50张探地雷达图像对该网络进行测试。结果显示,该方法对钢拱架的识别精度达到98%,对脱空的识别精度达到92%,能够实现对隧道初期支护检测探地雷达图像的自动识别。
Ground penetrating radar(GPR)is one of most used tool for tunnel lining detection,but most of the data interpretation work needs to be completed by manpower.In order to rapidly evaluate the quality of the primary support,an automatic recognition algorithm based on convolutional neural network(CNN)is developed to detect the arrangement of the steel ribs and the presence of voids in the primary support structure.The algorithm adopts actual GPR image as training data and uses Mask R-CNN framework together with the ResNet 101 model for feather extraction,object classification and position regression.The designed network is first pre-trained on the COCO dataset to obtain a set of pre-trained weights,then fine tune is conducted to the weights based on 200 processed GPR images.Finally,this network is tested with 50 GPR images and the recognition accuracy reaches 98%for steel ribs and 92%for voids,showing that the designed method can recognize the targets in GPR images of the tunnel primary support automatically.
作者
张东昊
覃晖
ZHANG Donghao;QIN Hui(School of Civil Engineering,Dalian University of Technology,Dalian 116024)
出处
《现代隧道技术》
EI
CSCD
北大核心
2020年第S01期174-178,共5页
Modern Tunnelling Technology
基金
国家自然科学基金项目(41904095)
中央高校基本科研业务费(DUT19 JC23,DUT19 RC(4)020).
关键词
隧道
衬砌检测
探地雷达
目标识别
深度学习
卷积神经网络
Tunnel
Lining detection
Ground penetrating radar(GPR)
Target Recognition
Deep learning
Convolutional neural network(CNN)
