基于YOLOX-G算法的隧道裂缝实时检测  

Real-time detection of tunnel cracks based on YOLOX-G algorithm

作　　者：周中[1] 闫龙宾 张俊杰 杨豪 ZHOU Zhong;YAN Longbin;ZHANG Junjie;YANG Hao(School of Civil Engineering,Central South University,Changsha 410075,China;School of Traffic and Transportation Engineering,Changsha University of Science&Technology,Changsha 410114,China;National Engineering Research Center of Highway Maintenance Technology,Changsha University of Science&Technology,Changsha 410114,China)

机构地区：[1]中南大学土木工程学院,湖南长沙410075 [2]长沙理工大学交通运输工程学院,湖南长沙410114 [3]长沙理工大学公路养护技术国家工程研究中心,湖南长沙410114

出　　处：《铁道科学与工程学报》2023年第7期2751-2762,共12页Journal of Railway Science 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);中南大学研究生科研创新项目(1053320213484)。

摘　　要：裂缝是隧道衬砌最常见的病害之一,影响隧道的结构耐久性和运营安全性。由于现役隧道日常检修任务艰巨,因此对隧道裂缝的高效智能化检测至关重要。针对隧道衬砌裂缝传统检测方法检测精度低、检测速度慢等问题,基于YOLOX算法提出一种新的YOLOX-G隧道衬砌裂缝图像检测算法。采用Ghostnet替换YOLOX的CSPDarknet主干网络,在加强特征提取网络中利用Ghost卷积代替原卷积块,用GIOU损失函数代替IOU损失函数。将YOLOX-G算法与YOLOX,YOLOv5,YOLOv3,SSD和Faster RCNN 5种算法在构建的隧道裂缝图像数据集上进行实验对比,结果显示:YOLOX-G算法的F1分数为85.29%,相较于其他5种算法分别提高了4.26%,6.49%,7.29%,17.23%和4.53%;AP值为90.14%,相较于其他5种算法分别提高了7.28%,10.93%,11.53%,17.65%和10.38%。此外,YOLOX-G算法模型数据大小为38.1 M,相对于YOLOX算法模型压缩了81.59%;检测单张图片的时间为15.12 ms,FPS为66.14帧/s,相较于其他5种算法分别提高了18.89帧/s,13.92帧/s,21.41帧/s,25.72帧/s和49.69帧/s。因此,提出的YOLOX-G算法满足移动设备对模型大小的要求及对帧率的需求,可以实现对隧道衬砌裂缝高速度、高精度、实时动态性检测。Cracks are one of the most common diseases of tunnel linings,affecting the structural durability and operational safety of tunnels.Due to the huge daily maintenance tasks of active tunnels,efficient and intelligent detection of tunnel cracks is crucial.Aiming at the problem of low detection accuracy and slow detection speed of traditional detection methods of tunnel lining cracks,a new YOLOX-G tunnel lining crack image detection algorithm was proposed based on the YOLOX algorithm.Firstly,the Ghostnet was used to replace the YOLOX’s CSPDarknet backbone network.Secondly,the Ghost convolution was used to replace the original convolution block in the enhanced feature extraction network.Finally,the IOU loss function was replaced by the GIOU loss function.Comparing the YOLOX-G algorithm with other five algorithms of YOLOX,YOLOv5,YOLOv3,SSD,and Faster RCNN on the constructed tunnel crack image dataset,the results are drawn as follows.The F1 score of the YOLOX-G algorithm is 85.29%,which is 4.26%,6.49%,7.29%,17.23%,4.53%higher than the other five algorithms,respectively.The AP value of the YOLOX-G algorithm is 90.14%,which is 7.28%,10.93%,11.53%,17.65%,10.38%higher than the other five algorithms,respectively.In addition,the data size of the YOLOX-G algorithm model is 38.1 M,which is 81.59%compressed compared with that of the YOLOX algorithm model.For the YOLOX-G algorithm,the time to detect a single image is 15.12 ms,and the FPS is 66.14 frames/s,which is higher than the other five algorithms of 18.89 frames/s,13.92 frames/s,21.41 frames/s,25.72 frames/s,49.69 frames/s,respectively.Therefore,the proposed YOLOX-G algorithm meets the requirements of mobile devices for model size and frame rate,and can realize high-speed,high-precision,and real-time dynamic detection of tunnel lining cracks.

关 键 词：隧道工程 裂缝检测 深度学习 神经网络 

分 类 号：U45[建筑科学—桥梁与隧道工程]