基于改进Yolov5的地铁隧道附属设施与衬砌表观病害检测方法  

Metro tunnel accessorial facilities and lining diseases detection method based on improved Yolov5

作　　者：朱家松[1] 郑澳 雷占占 练敏青 杨军伍 李林超 ZHU Jiasong;ZHENG Ao;LEI Zhanzhan;LIAN Minqing;YANG Junwu;LI Linchao(College of Civil and Transportation Engineering,Shenzhen University,Shenzhen 518000,China;School of Architecture&Urban Planning,Shenzhen University,Shenzhen 518000,China;Shenzhen Metro Group Co.,Ltd.,Shenzhen 518026,China)

机构地区：[1]深圳大学土木与交通工程学院,广东深圳518000 [2]深圳大学建筑与城市规划学院,广东深圳518000 [3]深圳市地铁集团有限公司,广东深圳518026

出　　处：《铁道科学与工程学报》2023年第3期1008-1019,共12页Journal of Railway Science and Engineering

基　　金：国家重点研发计划资助项目(2018YFB2101000);深圳市地铁集团有限公司科研咨询服务(STJS-DT413-KY002/2021);深圳市科技创新委项目-稳定支持面上项目(20200812102651001);广东省区域联合基金青年项目(2020A1515110438);广东省自然科学基金面上项目(2022A1515010939)。

摘　　要：随着地铁隧道服役时间增长,隧道衬砌在多因素影响下病害频发,对隧道结构及临近附属设施造成不良影响,严重危及到行车安全。因此,亟需精确高效的地铁隧道病害及设施检测技术。然而,地铁隧道存在内部环境复杂,附属设施与衬砌病害纹理及灰度相似、目标尺度不一等检测难点,传统的人工巡检及数字图像处理方法均存在局限。针对上述问题,提出一种基于改进Yolov5的地铁隧道附属设施及衬砌表观病害检测模型。针对设施及病害的位置特征引入坐标注意力(Coordinate attention)引导模型对目标区域赋予更高权重,抑制背景噪声;采用Bi FPN(Bi-directional Feature Pyramid Network)特征融合网络提升小目标病害检测效果;并利用Ghost Bottleneck替代部分卷积减少模型参数,提高检测效率。为验证改进后模型检测性能,进行现场实验,构建样本数量为843的地铁隧道衬砌图像数据集。并采用随机裁剪、镜像翻转等数据增强方法,将样本量扩充至4 072。数据集上的实验结果表明,改进模型的平均精度均值(m AP)可达89.2%,较原模型提高了3.7%,有效提升了隧道环境中小目标病害的检测效果。且模型参数减少了12%,更有利边缘端部署。相比于其他隧道检测模型,改进后的模型在综合性能上更具优势,可为地铁隧道衬砌病害实时检测和附属设施数字化提供技术支持。With the increase of service time of subway tunnel,tunnel lining diseases occur frequently under the influence of many factors,which have a bad influence on the tunnel structure and accessorial facilities,and seriously endanger the traffic safety.Therefore,an accurate and efficient detection technology for subway tunnel diseases and facilities is urgently needed.However,the traditional manual inspection and digital image processing methods have limitations due to the complex internal environment,similar texture and gray value of accessorial facilities and lining diseases,and different target scales.To solve the above problems,this paper proposed an improved Yolov5 metro tunnel accessorial facilities and lining diseases detection method.According to the location characteristics of facilities and diseases.Coordinate attention mechanism was introduced to guide the model to give higher weight to the target area,and suppress background noise Bi-directional feature pyramid network was adopted to improve the efficiency of small scale diseases.Uses GhostBottleneck to replace part of convolution to compression model volume and improve detection efficiency.In order to verify the performance of the improved model,a field experiment was carried out,and a metro tunnel lining image dataset with 843samples was constructed.The sample size was increased to 4 072 by a series of data enhancement methods,such as random clipping and mirror flipping.The experimental results show that the mean average precision of the proposed model is 89.2%,3.7% higher than that of the original model,which improved the detection effect of small target in tunnel environment.The model parameters are reduced by 12%,which makes the model deploying to edge devices more conveniently.Compared with other tunnel detection models,the proposed model has more advantages in comprehensive performance,and can provide technical support for real-time detection of subway tunnel lining diseases and digitization of accessorial facilities.

关 键 词：地铁隧道 衬砌检测 深度学习 目标检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]