Research On Interpretation Method Of Underground Cylindrical Object Based On GPR Image And Deep Learning

Ground Penetrating Radar(GPR)technology,as a nondestructive testing method,has been widely used in underground infrastructure construction.At present,the interpretation of GPR images mostly relies on manual interpretation,which is inefficient and timeconsuming.In view of the interpretation requirements of GPR underground cylindrical objects,this paper combines deep learning to conduct in-depth research on the interpretation of GPR images.Based on improved deep learning Framework and image processing methods,data enhancement,image denoising,object detection and parameter estimation methods for GPR images are studied.It aims to provide a set of intelligent interpretation methods for GPR images and provide strong support for GPR detection in underground space.The main content and research work of this thesis include the following aspects:(1)A GPR image enhancement method based on generative adversarial network is proposed.Building a Wasserstein Generative Adversarial Network(WGAN)for augmenting the GPR image dataset of subterranean cylindrical objects.WGAN is trained by simulated GPR images and a small number of field GPR images.After the training is completed,a large number of realistic GPR images can be generated by WGAN.It solves the problem of insufficient datasets and improves the performance of subsequent image denoising networks.(2)A multi-scale convolutional autoencoder(MCAE)based GPR image denoising method is proposed.By improving the network structure of the encoder and decoder in the convolutional autoencoder,the features of different scales in the image are effectively extracted.The experimental results on simulated,generated and field datasets show that the proposed scheme can effectively denoise GPR images at low signal-to-noise ratio.It prepares for subsequent GPR image object detection.(3)A GPR target region detection method based on ATRD algorithm is proposed.The proposed ATRD algorithm includes the steps of adaptive normalization,binarization,dilation operation,contour detection and region framing.The algorithm can quickly determine the target area with a small amount of calculation,which meets the requirements of real-time detection.The detected target area will be used as the input of the next step recognition network.(4)A method for identifying diameter parameters of GPR cylindrical objects based on the CNN-LSTM framework is proposed.The constructed CNN-LSTM framework integrates convolutional neural network(CNN)and long short-term memory(LSTM)network to extract spatial location features and temporal features of hyperbolic regions.It transforms the diameter estimation task into a hyperbolic region classification task.Experimental results on both simulated and field datasets demonstrate that the proposed scheme has good performance in diameter identification.On the simulated data set,the recognition accuracy reached 99.5%.On the field data set,the recognition accuracy reached 92.5%.