Research on Classification Method of Tunnel Structure Disease Type Based on Deep Learning
王维①TWANG Wei;黄宝森②HUANG Bao-sen;陈颖③CHEN Ying;
王鲁杰③WANG Lu-jie;郭春生①GUO Chun-sheng;李家平③LI Jia-ping
①上海勘察设计研究院(集团)有限公司,上海200438;
②绍兴京越地铁有限公司,绍兴312099;
③上海地铁监护管理有限公司,上海 200070
①SGIDI Engineering Consulting ( Group ) Co.,Ltd. ,Shanghai 200438,China ;
②Shaoxing Jingyue Metro Co.,Ltd.,Shaoxing 312099,China ;
③Shanghai Metro Monitoring Management Co.,Ltd. ,Shanghai 200070,China
摘要:受建设条件、运营环境等复杂因素影响,隧道结构在运营期间将不可避免出现诸如渗水.缺损、裂缝等多种病害。以相机拍摄或三维激光扫描为基础的快速隧道病害检测方法是未来发展的趋势。近年来,国内外的许多学者对如何快速处理数量庞大的隧道影像数据、自动识别病害特征,进行了一定的研究,但这些研究大多采用相机照片建立深度学习数据集。本文结合人工标注的上海部分地铁隧道病害可见光相片与三维扫描灰度影像,分别建立了具有一定规模的隧道病害数据集,实验并对比了多种图像分类卷积神经网络在自建数据集上的表现,此外,还比较了可见光数据集与灰度数据集的独立训练效果和混合训练效果,并对模型错分情况进行了讨论。
Abstract:Affected by the complex factors such as construction conditions and operation environment,the tunnel structure willinevitably have various diseases such as leakage, missing block and cracks during the operation periodl.Rapid tunel disease detectionmethods hased on camera photography or 3D laser scanning are future trends.Many scholas have conducted some research on how toquickly process a large nmumber of tunnel images and automatically identify disease characteristics, but most of these studies use cameraphotos to establish deep learning data sets. In this paper, the tunnel disease datasets wih a certain sale are estabished hased on themanually labeled visible light photos of some subway tunnel diseases in Shanghai and the thre-dimensional scanned gray scale imagesExperiments are carried out to compare the performance of various image classification Convolutional neural network on the self buildatasets. In addition, the independent training effect and mixed training effect of visible light datasets and gray scale datasets are compared.and the model misclassification is discussed.
关键词:深度学习;数据集;隧道病害;隧道三维激光扫描影像
Key words: deep learning; dataset; tunnel defect; tunnel image of 3D laser scan
价值工程-文章出处：王维, 黄宝森, 陈颖, 王鲁杰, 郭春生, 李家平. 基于深度学习的隧道结构病害类型识别方法研究[J]. 价值工程, 2023, 42 (21): 138-141. |
