探地雷达数值模拟与道路裂缝图像检测的深度学习增强方法  

Deep learning-enhanced numerical simulation of ground penetrating radar and image detection of road cracks

作　　者：刘震 顾兴宇[1] 李骏 董侨[1] 蒋继望 LIU Zhen;GU XingYu;LI Jun;DONG Qiao;JIANG JiWang(School of Transportation,Southeast University,Nanjing 211189,China;Kaili Highway Administration Bureau of Guizhou Province,Guiyang 550008,China)

机构地区：[1]东南大学交通学院,南京211189 [2]贵州省凯里公路局,贵阳550008

出　　处：《地球物理学报》2024年第6期2455-2471,共17页Chinese Journal of Geophysics

基　　金：国家重点研发计划(2021YFC3000074)资助。

摘　　要：针对路面结构内部裂缝的无损检测与智能识别需求,提出一种基于改进的YOLOv8和AutoAugment增强的探地雷达检测与图像自动识别方法.结合Gprmax数值模拟、室内模型试验及现场探地雷达测试,分析得出内部裂缝在探地雷达图像上呈现出明显的"双曲线"特征,且裂缝宽度与双曲线的幅度成正比.利用MALA探地雷达GX-750检测获取路面结构内部裂缝病害图像,经多重滤波处理后采用640×640的像素框进行截取.针对裂缝特征在探地雷达图像中尺度较小的问题,通过对最新的YOLOv8模型增加一层像素为160×160的输出层得到改进的YOLOv8模型.同时,引入SKNet注意力机制进一步增加感受野,并采用power IoU损失函数以降低模型训练的损失.针对原始图像数据集,采用AutoAugment无监督自动增强方法,通过近端策略优化的强化学习算法寻找最佳增强策略及其概率和强度,实现了探地雷达数据集的有效扩充.在扩充的数据集上进行训练与测试,结果表明改进的YOLOv8模型取得了90.7%的平均检测精度和90.1%的F1分数,相比原始YOLOv8模型分别提升了6.3%和5.9%,也大幅度超越了主流的目标检测模型.在进行图像增强后,模型的平均检测精度和F1分数分别提升了4.1%和4.6%,对各类尺度的裂缝图像检测也体现出良好的鲁棒性.在养护路段应用探地雷达图像智能识别方法的检测结果与取芯验证结果相吻合,表明提出的改进模型在实际工程应用中是可靠的.Aiming at the requirements of nondestructive testing and intelligent recognition of cracks inside the pavement structures,a ground penetrating radar(GPR)detection and automatic identification method based on improved you only look once version 8(YOLOv8)and AutoAugment enhancement was proposed.Combined with Gprmax numerical simulation,laboratory model test,and field GPR measurement,it is concluded that the internal cracks showed a prominent"hyperbola"feature on GPR images,and the width of the cracks is proportional to the size of the hyperbola.MALA GPR with GX-750 was used to detect the crack images inside the pavement structure,and a 640×640 pixel frame was used to intercept the image after multiple filtering processing.Aiming at the problem of small crack features in GPR images,the improved YOLOv8 model was obtained by adding an output layer with pixels of 160×160 to the latest YOLOv8 model.Meanwhile,Selective Kernel Networks(SKNet)attention mechanism was introduced to increase the sensitivity field further and adopted the loss function of power IoU to decrease the training loss of the model.As for the original image data set,the AutoAugment method with unsupervised automatic augment was applied to find the best enhancement strategy and its probability and intensity through the reinforcement learning algorithm optimized by the proximal strategy to realize the effective expansion of the GPR image data set.After training and testing on the expanded data set,the results show that the improved YOLOv8 model has achieved a mean average precision(mAP)of 90.7%and an F1 score of 90.1%,which is 6.3%and 5.9%higher than that of the original YOLOv8 model,and also significantly exceeds the mainstream target detection model.After image enhancement,the mAP and F1 score of the model increased by 4.1%and 4.6%,respectively,showing good robustness to crack image detection of various scales.The application of the GPR image intelligent recognition method in the maintenance section is consistent with the coring verification results,

关 键 词：道路工程 路面结构内部裂缝 探地雷达检测 正演模拟 深度学习 图像增强 

分 类 号：P631[天文地球—地质矿产勘探] U418[天文地球—地质学]