改进YOLOv5的公路隧道衬砌裂缝检测  

Highway Tunnel Lining Crack Detection Based on Improved YOLOv5

作　　者：鲁冠宏 吕成顺 田隽 南骁聪 马银强 刘健 解全一 LU Guan-hong;L Cheng-shun;TIAN Juan;NAN Xiao-cong;MA Yin-qiang;LIU Jian;XIE Quan-yi(School of Qilu Transportation,Shandong University,Jinan 250002,China;Shandong Hi-speed Company Limited,Jinan 250098,China;Shandong Hi-speed Engineering Test Co,Ltd,Jinan 250002,China;Shandong Research Institute of Industrial Technology,Jinan 250101,China)

机构地区：[1]山东大学齐鲁交通学院,济南250002 [2]山东高速股份有限公司,济南250098 [3]山东高速工程检测有限公司,济南250002 [4]山东省工业技术研究院,济南250101

出　　处：《科学技术与工程》2025年第7期2997-3006,共10页Science Technology and Engineering

基　　金：山东省自然科学基金(ZR2021QE279,ZR2022DKX001);山东省泰山学者工程资助项目(tstp20221153)。

摘　　要：高效、准确的衬砌裂缝检测可以为评估隧道结构安全提供依据。针对传统裂缝检测方法复杂和泛化能力弱的缺点,提出了一种基于深度学习的隧道衬砌裂缝检测网络YOLOv5-CT(YOLOv5 CBAM Transformer)。考虑到裂缝细长的形态,网络引入了Transformer模块来改善裂缝检测效果。Transformer模块较强的长距离依赖捕捉能力使得所提出的检测模型能够充分学习到裂缝区域的上下文信息。此外,该网络在特征融合部分还集成了卷积注意力机制CBAM(convolutional block attention module)。在自采集数据集上的实验结果表明:YOLOv5-CT的AP50和AP(average precision)分别可以达到85.2%和51.3%,相比于基线模型YOLOv5提高了8.9%和12.1%,在精度上优于YOLOX、YOLOv3-MobileNet等其他单阶段目标检测网络。在640×640像素条件下推理速度达到161.3 f/s(frames per second),可以实现隧道衬砌裂缝实时检测。Efficient and accurate crack detection can provide a basis for assessing the structural safety of tunnels.Aiming at the shortcomings of traditional crack detection methods,which are complex and weak in generalization ability,an improved algorithm YOLOv5-CT(YOLOv5 CBAM Transformer)for tunnel lining crack detection was proposed.Considering the slender morphology of the cracks,the network introduced the Transformer module to improve the crack detection effect.The strong long-range dependency capture ability of the Transformer module enabled the proposed detection model to fully learn the contextual information of the crack region.In addition,the network integrated the convolutional attention mechanism CBAM(convolutional block attention module)in neck.The experiment shows that the YOLOv5-CT can achieve AP50 and AP of 85.2%and 51.3%,respectively,which is an improvement of 8.9%and 12.1%compared to the baseline model YOLOv5.It is better than other one-stage object detection networks in terms of accuracy,and the inference speed reaches 161.3 fps under 640×640 pixel conditions,which meets real-time detection of tunnel lining cracks.

关 键 词：公路隧道 裂缝检测 TRANSFORMER 注意力机制 YOLOv5 

分 类 号：U457.2[建筑科学—桥梁与隧道工程] TP391.41[交通运输工程—道路与铁道工程]