The Research On GPR Pipeline Image Recognition Based On Semi-supervised 3D-CNN

The damage of old underground pipelines often causes huge economic losses,and even threatens the safety of people’s lives in severe cases.The lack of underground pipeline data will also hinder the construction process of modern cities.With the development of vehicle-mounted ground-penetrating radar(GPR)technology,the detection efficiency of GPR has been greatly improved.However,the heavy manual interpretation work limits the application of GPR in large-scale urban scenarios.Algorithms for accurate identification of GPR images can greatly reduce detection costs.Therefore,this paper mainly uses convolutional neural network to identify and classify GPR 3D pipeline images.Firstly,there is no publicly available GPR dataset.Based on the original data collected by the vehicle-mounted 3D GPR in the actual project,the preprocessing of the original data and the preparation of the dataset are completed.In view of the direct wave interference in the GPR image and the rapid attenuation of the amplitude of the electromagnetic wave with the increase of the propagation depth,resulting in the weak reflection echo of the deep target.In this paper,the mean filtering method is used to remove the direct wave interference,and the method of adaptively generating the gain function based on the envelope extreme value is used to compensate the echo.Finally,the characteristics of various underground targets in GPR 3D images are analyzed and explained,and GPR dataset is made by manual screening and labeling.Then,the supervised learning method of GPR 3D image based on convolutional neural network(CNN)is studied.Firstly,the problem that 2D convolution is difficult to extract the features between adjacent channels is expounded,and the necessity of using 3D convolution to extract target features in GPR 3D images is explained,and the advantages of using(3D-CNN)to identify and classify underground objects are verified by experiments.Then a network(A-3D-CNN)combined with attention mechanism is proposed to make the network pay more attention to the key areas in the GPR image and effectively improve the recognition accuracy of the network.Finally,the semi-supervised learning method of GPR 3D image based on CNN is studied.Aiming at the problem of limited labeled samples in GPR data,a method of assigning pseudo-labels to unlabeled samples through self-training method(PL algorithm)is firstly proposed to effectively utilize unlabeled samples.However,this method cannot confirm the correctness of pseudo-labels,and it is difficult to select reliable samples by predicting the confidence threshold to ensure better performance of the network.Next,aiming at the problem of reliable sample selection,this paper proposes a method for reliable sample selection based on dual classifiers(SVM-PL algorithm).The samples with more reliable labels are used to participate in network training,which effectively improves the classification performance of the network.