基于深度学习的隧道衬砌结构物探地雷达图像自动识别

传统探地雷达图像识别方法中存在识别准确率不高,复杂目标体识别难度大,识别流程较为繁琐,不能实现端到端识别等问题,不能准确的识别实测数据.本文将深度学习中Faster R-CNN、YOLOv3这两种具有代表性的目标检测算法运用到探地雷达的图像识别当中.选择隧道的衬砌结构作为探测识别研究目标,制作了包含钢拱架、钢筋网、施工缝三类结构物标注的实测数据集.从准确率、召回率、平均准确率,准确率-召回率曲线等评价指标,分析了这两种算法在实测数据集上的表现.并对照典型的识别结果,结合这两种算法的原理说明了其运用到探地雷达图像自动识别上的特点.在测试集上,Faster R-CNN在钢拱架、钢筋网、施工缝三类结构物的识别中分别取得了95.5%、90.5%、96.8%平均准确率,YOLOv3则分别取得了90.0%、16.6%、95.3%平均准确率.实验结果表明两种深度学习目标检测方法在隧道衬砌探地雷达图像的识别上取得了良好的效果,其中Faster R-CNN整体效果更好,但会将多次波误识别为有效信号,YOLOv3较少误识别多次波,但是对钢筋网识别效果不好,两种方法搭配使用会形成优势互补.

Automatic recognition of ground penetrating radar image of tunnel liningstructure based on deep learning

Traditional ground penetrating radar image recognition methods have some problems,such as low recognition accuracy,difficult identification of complex target,cumbersome identification process and inability to realize end-to-end recognition.So the accurate identification of the measured data cannot be achieved well.In this paper,two representative target detection algorithms,Faster R-CNN and YOLOv3 were applied to the image recognition of ground penetrating radar.The lining structure of the tunnel was selected as the research target of detection and identification,and the measured data set of the three types of structures including steel arch,steel mesh and construction joint was made.The performance of these two algorithms on measured data sets was analyzed by evaluation indexes such as accuracy,recall,average accuracy,Precision-Recall curve,etc.Through the typical recognition results,combined with the principles of these two algorithms,their application characteristics in the automatic recognition of ground penetrating radar images were illustrated.In the test set,Faster R-CNN achieved an average accuracy of 95.5%,90.5%and 96.8%in the identification of steel arch,steel mesh and construction joints,respectively,and YOLOv3 achieved an average accuracy of 90.0%,16.6%and 95.3%,respectively.The experimental results show that the two deep learning target detection methods have achieved good results in the recognition of tunnel lining ground penetrating radar images.Faster R-CNN has better overall results than YOLOv3,but it would misidentify multiple waves as valid signals.YOLOv3 rarely misidentifies multiple waves,but it has poor recognition effect on steel mesh.The two algorithms will form complementary advantages on image recognition for lining structures of the tunnel.