基于全景图像的隧道渗漏水病害识别方法  

Identification Method of Tunnel Water Leakage Based on Panoramic Image

作　　者：许敏娟 陈莹莹 刘浩 XU Minjuan;CHEN Yingying;LIU Hao(Kunming Metro Operation Co.,Ltd.,Kunming 650021,China;Shanghai Tongyan Civil Engineering Technology Co.,Ltd.,Shanghai 200092,China;Shanghai Engineering Research Center of Detecting Equipment for Underground Infrastructure,Shanghai 200092,China;Jinan Rail Transit Group Co.,Ltd.,Jinan 250000,China;School of Traffic and Transportation,Beijing Jiaotong University,Beijing 100044,China)

机构地区：[1]昆明地铁运营有限公司,云南昆明650021 [2]上海同岩土木工程科技股份有限公司,上海200092 [3]上海地下基础设施安全检测与养护装备工程技术研究中心,上海200092 [4]济南轨道交通集团有限公司,山东济南250000 [5]北京交通大学交通运输学院,北京100044

出　　处：《铁道学报》2023年第8期184-192,共9页Journal of the China Railway Society

基　　金：山东省重点研发计划(2019JZZY010428);国家重点研发计划(2018YFB2101004)。

摘　　要：针对隧道衬砌结构渗漏水形态复杂,光照与背景干扰强,样本不平衡导致的误识别与漏识别问题,为提高渗漏水自动识别准确率和效率,提出基于全景图像的渗漏水病害快速检测方法。首先,依据自主研发的隧道检测车采集结构表观全景图像,构建专属的渗漏水定位检测及分割样本库;然后,针对渗漏水这类形态非刚性目标,改进了DeepLab V3+分割网络,引入可变形卷积以提高感受野尺度变化自适应能力,同时融合逐像素交叉熵损失函数与Focal Loss损失函数;最后,提出方向区域搜索算法,解决因窗口滑动预测方式而造成分割断裂的问题。研究结果表明:与UNet、DeepLab V3+相比,提出的改进算法分割精度平均交并比为91.02%,提高了3.3%;同时平均每张2560×2048像素图片识别耗时0.30 s,降低了23%。方法已集成于同济曙光病害自动识别软件,并已用于隧道检测工程作业中,取得良好的应用效果。Aiming at false and missing identification caused by complex water leakage form in tunnel lining structure,strong illumination and background interference and unbalanced samples,a fast detection method of water leakage based on panoramic image was proposed to improve the accuracy and efficiency of automatic recognition of leakage water.First of all,based on the apparent panoramic images of the tunnel structure collected by the independently-developed tunnel inspection vehicle,an exclusive sample database was constructed for water leakage detection and segmentation.Then,for non-rigid targets with complex shapes such as water leakage,the DeepLab V3+segmentation network was improved,to introduce deformable convolution to improve the adaptive ability of the receptive field scale change,and the pixel-by-pixel cross-entropy loss function and the Focal Loss loss function were combined.Finally,a direction region search algorithm was proposed to solve the problem of segmentation fracture caused by window sliding prediction.The results show that,compared with UNet and DeepLab V3+,the segmentation accuracy of the proposed improved algorithm is 91.02%on MIoU,an increase of 3.3%.At the same time,the average recognition time of each 2560×2048 pixel picture is 0.30s,down 23%.The method,integrated into Tongji Shuguang automatic identification software,and used in tunnel detection engineering,achieved good application results.

关 键 词：隧道检测 渗漏水 深度学习 目标检测 精细分割 

分 类 号：U457.2[建筑科学—桥梁与隧道工程]