Research On Recognition And Detection Of Underground Pipeline Using GPR Based On Deep Learning

Urban underground pipelines play an important role in maintaining people’s daily life and industrial production,and provide important support for the development and modernization of cities.However,due to the aging of underground pipelines,unscientific and nonstandard management of early urban pipelines,improper selection of pipeline detection methods and many other factors,the accurate identification and detection of underground pipelines are facing great challenges.Therefore,how to efficiently identify and locate underground pipelines is of great significance in the process of urban construction.This thesis mainly studies the ground penetrating radar（GPR）simulation data of underground pipelines,and applies deep learning algorithms to the recognition and detection tasks of pipeline GPR two-dimensional images.The main research contents are as follows:1.The forward modeling models of targets such as drainage pipes,gas pipes,water pipes,and rock blocks in the ground penetrating radar detection mode are established,and the hyperbolic characteristics of target echo signals in the B-scan images of four targets are analyzed.For the problem of strong direct wave energy in the image,the direct wave suppression effect was evaluated from the average cancellation method,twodimensional filtering,SVD decomposition method,etc.The experiments show that the SVD decomposition method has an efficient direct wave removal effect,which can be selected as the direct wave removal algorithm.Aiming at the interference problems such as noise in the actual working environment,the denoising performance of the wavelet transform is studied.The results show that the wavelet transform combined with the rbio6.8 wavelet basis function and the soft and hard threshold function has a good performance in B-scan image denoising.2.The basic modules of various convolutional neural networks such as atrous convolution,channel attention,spatial attention,and residual structure are studied,and their effects are verified on the pipeline dataset constructed in this thesis.The results show that three structures,such as atrous convolution,spatial attention and residual structure,can improve the recognition ability of the network.Combined with the three modules,a residual network based on the fusion of attention and multi-scale features is designed.The results show that the highest recognition accuracy of the recognition network proposed in this thesis can reach 86.8%.3.The YOLOv4 network is improved from the two aspects of accuracy improvement and network structure lightweight.The YOLOv4 network based on Improved Aspp and Attention feature enhancement and the YOLOv4 lightweight network based on grouped convolution and depthwise separable convolution are proposed respectively.The experimental results show that the former improves the m AP by about 0.61% and 1.13%on the YOLOv3 and YOLOv4 structures,respectively;while the latter（lightweight network）can reduce the number of parameters to 3.77 M,which is 20 times lower than that of YOLOv4,and its m AP reaches 69.7 %,the detection accuracy is better than several mainstream lightweight improved YOLOv4 networks.