基于多任务约束深度学习的水工隧洞缺陷检测方法研究  

Research on hydraulic tunnel defect detection method based on deep learning model with multi-task constraint

作　　者：邓旭方 徐轶 阿依胡兰·阿山 陈正虎[1] 蔡伟[1] DENG Xufang;XU Yi;AYIHULAN Ashan;CHEN Zhenghu;CAI Wei(China Yangtze Power Co.,Ltd.,Wuhan 430014,China;Changjiang Survey,Planning,Design and Research Co.,Ltd.,Wuhan 430010,China;National Dam Safety Engineering Research Center,Wuhan 430010,China;School of Remote Sensing and Information Engineering,Wuhan University,Wuhan 430072,China)

机构地区：[1]中国长江电力股份有限公司,湖北武汉430014 [2]长江勘测规划设计研究有限责任公司,湖北武汉430010 [3]国家大坝安全工程技术研究中心,湖北武汉430010 [4]武汉大学遥感信息工程学院,湖北武汉430072

出　　处：《水利水电快报》2024年第7期64-69,共6页Express Water Resources & Hydropower Information

基　　金：中国长江电力股份有限公司科研项目(Z212102007)。

摘　　要：由于水工隧洞环境复杂、缺陷类型多,数字图像处理技术在隧洞巡检中的应用存在缺陷识别精度低、鲁棒性差等问题。为此,提出了基于多任务约束深度学习的水工隧洞缺陷自动检测方法。该方法在Faster R-CNN深度学习原理的基础上,结合原始目标检测的定位约束、类别约束,增加特征到影像恢复的约束,可以较完整地保留图像以及缺陷特征信息,从而解决部分水工隧洞缺陷样本不全而引起的特征提取不鲁棒的问题。测试结果表明:多任务约束的Faster R-CNN深度学习方法能够较好地识别裂缝、渗水、掉块等多种病害缺陷,有效提高了识别精度,为水工隧洞工程缺陷检测提供了可靠方法。Digital image processing technology faces problems such as low recognition accuracy and poor robustness,due to the complex environment of hydraulic tunnels and many types of defects.Therefore,this article proposed an automatic tunnel defect detection method based on multi-task constraint deep learning.Based on the Faster R-CNN deep learning principle,combined with the localization and category constraints of original object detection,image self-coding recovery constraints were introduced to solve the problem of feature extraction been not robust due to the scarcity of some tunnel defect samples,and to fully preserve image and defect feature information.Tests had shown that Faster R-CNN with multi-task constraints can effectively identify defects such as cracks,water seepage,and falling blocks,providing a reliable detection method for practical tunnel engineering defects.

关 键 词：隧洞缺陷检测 多任务约束深度学习 Faster R-CNN 隧洞缺陷特征表达 

分 类 号：TV698.1[水利工程—水利水电工程]