摘要
为了提高探地雷达图像中的病害自动识别的效率和准确度,通过现场实测和正演模拟的方法获得并扩充训练数据集,用于训练YOLO v5模型,以实现对探地雷达图像中地下空洞与管线的快速准确分类。结果表明:①YOLO v5模型能较好的定位和区分地下空洞、金属管线和混凝土管线三类地下目标。②使用正演模拟对数据集增广能提升模型的精准度和召回率,但数据增广比例不宜过高。③由于混凝土相对介电常数更接近土壤,因此混凝土管道的识别准确率较低。研究可为地下病害检测和识别工作提供一定参考。
To improve the efficiency and accuracy of automatic defect recognition in ground-penetrating radar(GPR)images,a training dataset was obtained and expanded through field measurements and forward modeling methods.This dataset was used to train the YOLO v5 model,aiming to achieve rapid and accurate classification of underground cavities and pipelines in GPR images.The results indicate that:①The YOLO v5 model can effectively locate and distinguish such three types of underground targets as cavities,metal pipelines,and concrete pipelines.②Using forward simulation modeling to augment the dataset improves the model's precision and recall rates,although excessive data augmentation may negatively impact results.③Due to the dielectric constant of concrete being closer to that of soil,the recognition accuracy for concrete pipelines is lower.This research provides valuable insights for underground defect detection and recognition.
作者
江路路
尹轶
孟姿含
许佳毅
JIANG Lulu;YIN Yi;MENG Zihan;XU Jiayi(Guoneng Huangda Railway Co.,Ltd.,Dongying 257000,China;School of Civil Engineering,Beijing Jiaotong University,Beijing 100044,China)
出处
《国防交通工程与技术》
2024年第5期7-11,共5页
Traffic Engineering and Technology for National Defence
关键词
探地雷达
图像识别
YOLO
v5模型
地下目标探测
数据增广
病害检测
ground-penetrating radar
image recognition
YOLO v5 model
underground target detection
data augmentation
defect detection
