DOI: 10.3390/app13137540 ISSN: 2076-3417

3D-GPR-RM: A Method for Underground Pipeline Recognition Using 3-Dimensional GPR Images

Xu Bai, Yu Yang, Zhitao Wen, Shouming Wei, Jiayan Zhang, Jinlong Liu, Hongrui Li, Haoxiang Tian, Guanting Liu
  • Fluid Flow and Transfer Processes
  • Computer Science Applications
  • Process Chemistry and Technology
  • General Engineering
  • Instrumentation
  • General Materials Science

Ground penetrating radar (GPR), as a non-destructive and rapid detection instrument, has been widely used for underground pipeline detection. However, as the interpretation of 3-dimensional GPR images is still manually performed, the process is inefficient. Aiming at solving the challenges of automatic recognition for underground pipelines, we propose a recognition method based on a deep learning algorithm, which uses 3-dimensional GPR images and the improved 3D depth-wise separable convolution block. In order to expand the number of samples in the dataset, we propose a data augmentation method based on three-dimensional matrix rotation and use a wavelet-based denoising method to filter out the direct wave interference. To prove the effectiveness and efficiency of our method, we compared the classification performance of the improved 3D depth-wise separable convolutional block with the traditional 3D convolutional block and the ordinary 3D depth-wise separable convolutional block under the same conditions. According to the experiment’s results, the number of parameters of the method we proposed is 66.9% less than that of the traditional 3D convolution method, while the classification performance is similar. Furthermore, compared with ordinary 3D depth-wise separable convolution, our method can significantly improve the classification and recognition ability of the neural network, while the number of calculations and the number of parameters remain almost the same. This study demonstrates the effectiveness of 3D-CNN in the field of GPR image interpretation. An improved 3D depth-wise separable convolutional block is also proposed. It greatly reduces the amount of calculation and parameters while ensuring classification performance. It is better than the existing algorithms in performance. At the same time, to obtain the position and direction of the pipeline, in this study, a conic fitting method using the Canny operator is proposed to locate the vertices of B-Scan images and record their horizontal and vertical coordinates. This method can estimate the direction of the pipeline and it lays the foundation for future work such as measuring the pipeline depth.

More from our Archive