3D GPR Target Recognition Based On Convolutional Neural Network

With the rapid development of urbanization and the continuous development of urban underground space resources,urban road collapse accidents occur frequently.A large number of road quality investigations have found that road collapse is closely related to underground concealed diseases.Ground penetrating radar(GPR)is a non-destructive and rapid detection technique using electromagnetic wave reflection to determine the distribution of underground media.It has become the preferred method for road disease detection.However,existing GPR data interpretation relies on expert experience.The rate of misjudgment and missing judgment is large,and the speed is slow.Target recognition is an essential part of data interpretation.Therefore,GPR intelligent target recognition has become a research hotspot and difficulty in recent years.Existing GPR intelligent target recognition methods mainly include those based on traditional machine learning and deep learning.The former requires manual feature extraction,while the latter mainly uses GPR two-dimensional(2D)data for recognition,which has low accuracy in identifying road diseases.To solve these problems,this paper proposes a GPR three-dimensional(3D)data target recognition method based on convolutional neural networks(CNN),which takes advantage of high information content of 3D data and automatic feature extraction by deep learning to realize intelligent recognition of urban road diseases.The main research work and achievements of this paper are as follows:(1)This paper summarizes the existing methods of GPR target recognition,analyzes the existing problems,and expounds on the basic principles of GPR and convolutional neural networks.Moreover,shallow CNN and faster region-based CNN designed by ourselves are used to carry out preliminary verification of underground target recognition based on GPR B-scan data.(2)An underground target recognition method based on three-view fusion is proposed.In this method,the features in the front view,side view,and top view of GPR 3D data are fused to improve the recognition accuracy.The proposed method is validated using laboratory data and compared to the state-of-the-art methods to verify its superiority.(3)An underground object recognition method based on 3D Convolutional Networks(C3D)and multiple mirror encoding(MME)is proposed.In this method,the C3 D is applied to extract the spatio-temporal features between parallel B-scans since it has a strong spatio-temporal feature learning ability.The MME method is proposed to enhance the spatio-temporal features of the target,as well as to make all 3D data the same size to satisfy the requirements of the network input.Experimental results show that the performance of the proposed method is better than that of the state-of-the-art methods,and the performance can be further improved by combining the C3 D with the corresponding 2D network.In addition,the method is suitable for the multi-target classification of small data sets.The proposed method provides a new solution to GPR target recognition based on 3D data with different widths.