基于深度语义分割的隧道渗漏水智能识别  

Automatic identification of tunnel leakage based on deep semantic segmentation

作　　者：周中[1,2] 张俊杰 龚琛杰 丁昊晖 ZHOU Zhong;ZHANG Junjie;GONG Chenjie;DING Haohui(School of Civil Engineering,Central South University,Changsha,Hunan 410075,China;Hunan Tieyuan Civil Engineering Testing Co.,Ltd.,Changsha,Hunan 410075,China)

机构地区：[1]中南大学土木工程学院,湖南长沙410075 [2]湖南铁院土木工程检测有限公司,湖南长沙410075

出　　处：《岩石力学与工程学报》2022年第10期2082-2093,共12页Chinese Journal of Rock Mechanics and Engineering

基　　金：国家自然科学基金资助项目(50908234);湖南省自然科学基金资助项目(2020JJ4743);湖南铁院土木工程检测有限公司开放课题(202106)。

摘　　要：为解决现有隧道渗漏水检测方法中存在的检测精度较低、抗干扰能力较差、检测速度较慢的问题,提出了一种基于深度语义分割的隧道渗漏水图像识别算法。该算法以DeepLabv3+语义分割算法为基础,首先采用轻量化分类网络EfficientNetv2作为主干网络,在减少网络参数的同时,提升了识别精度;其次融合卷积注意力机制模块,通过增大图像中有效通道的权重,提升网络对于渗漏水特征信息的提取能力;进而从图像识别精度、模型内存和处理速度3个方面与DeepLabv3+,PSPnet,Unet等传统语义分割算法进行对比实验。研究表明:提出算法的平均像素准确率和平均交并比分别为93.99%,89.87%,模型大小仅为33.4 MB,图像处理速度可达39.87 f/s。相较于3种对比算法,构建算法在隧道渗漏水病害检测的精度和效率上都有显著提升,且具有更为优越的边缘分割效果以及抗干扰能力,适用于复杂环境下的隧道渗漏水检测任务,能够更好地满足工程检测需求。Aiming to solve the challenges of low detection accuracy,poor anti-interference ability and slow detection speed in the traditional tunnel leakage detection methods,a depth semantic segmentation algorithm for tunnel leakage is proposed on the basis of the DeepLabv3+semantic segmentation algorithm.Firstly,the lightweight classification network EfficientNetv2 is used as the backbone network,which enhances the recognition accuracy while reduces network parameters.Secondly,Convolutional Block Attention Module(CBAM)is integrated to increase the weight of the effective channels in the image,thereby improving the ability of extraction of leakage feature information.Traditional semantic segmentation algorithms,including DeepLabv3+,PSPnet and Unet,are used for comparative experiments from three aspects:image recognition accuracy,model size,and detection speed.The results show that the mean pixel accuracy(mPA),mean intersection over union(mIoU),model size and image processing speed(FPS)of the proposed algorithm are 93.99%,89.87%,33.4 MB and 39.87 f/s,respectively.Compared with the three comparison algorithms,the detection accuracy and efficiency of the proposed algorithm have been significantly improved.Furthermore,the proposed algorithm has better edge segmentation effect andanti-interference ability,which is suitable for tunnel leakage detection tasks in complex environments,so as to better meet the needs of engineering detection.

关 键 词：隧道工程 渗漏水 语义分割 深度学习 神经网络 

分 类 号：U45[建筑科学—桥梁与隧道工程]