改进YOLOv5的探地雷达常见地下管线识别

Identification of common underground pipelines by ground penetrating radar based on improved YOLOv5

针对当前探地雷达(Ground Penetrating Radar,GPR)图像识别准确率低,且小目标识别困难等问题,本文将YOLOv5(You Only Look Once)和ConvNeXt网络相结合,提出一种适用于GPR图像的常见地下管线识别网络ConvNeXt-YOLOv5,显著提高了雷达图像中小目标的识别精度.同时,为了解决因地下管线原始信息不足而导致的模型训练精度低的问题,使用实测GPR图像、时域有限差分法(Finite-Difference Time-Domain,FDTD)与混类增强算法Mixup模拟图像构建数据集.然后,在此数据集上验证了本文所提方法对常见地下管线的识别效果,并用识别后的二维GPR B-scan图像还原出地下管线的三维图像.实验结果表明,对镀锌输水钢管、PVC管、电缆线和含水塑料瓶等常见地下管线进行识别时,ConvNeXt-YOLOv5表现出优良的识别性能.尤其是对PVC管的识别精度提升明显.与YOLOv4、YOLOv5、YOLOv7和Faster R-CNN模型相比,其平均精度均值分别提高了3.83%、2.82%、1.9%和3.72%.同时,三维探地雷达图像更直观的还原了地下管线信息.

Aiming at the current problems of low accuracy of ground-penetrating radar image recognition and difficulty in recognizing small targets,this paper combines YOLOv5 and ConvNeXt network and proposes a common underground pipeline recognition network ConvNeXt-YOLOv5 for GPR images,which significantly improves the recognition accuracy of small targets in radar images.Meanwhile,to solve the problem of low model training accuracy due to insufficient raw information on underground pipelines,a dataset is constructed using real GPR images,time-domain finite-difference method with the mixup enhancement algorithm Mixup simulated images.Then,the recognition effect of the proposed method on underground pipelines is verified on this dataset,and the recognized 2D GPR B-scan image is used to restore the 3D image of underground pipelines.The experimental results show that ConvNeXt-YOLOv5 exhibits excellent recognition performance when recognizing common underground pipelines,such as galvanized water transmission steel pipes,PVC pipes,cable wires,and plastic bottles containing water.In particular,the recognition accuracy of PVC pipes is significantly improved.Compared with YOLOv4,YOLOv5,YOLOv7,and Faster R-CNN model,its average accuracy mean is improved by 3.83%,2.82%,1.9%,and 3.72%,respectively.Meanwhile,the 3D ground-penetrating radar image more intuitively restores the underground pipeline information.