Research On Intelligent Processing Method Of Ground Penetrating Radar Target Feature Information

Ground penetrating radar(GPR)is based on the difference of medium electrical parameters which includes dielectric constant and conductivity to detect the interface position and morphological characteristics of the target.It has some prominent merits,such as non-destructiveness,speediness and high accuracy,and has a wide application prospect in shallow surface target detection.Under the influence of electromagnetic scattering,attenuation and interference in complex geological environments,the content of GPR images is complicated,which brings great challenges to GPR target detection and media parameter inversion.In this paper,considering the characteristics of targets in GPR B-SCAN images with hyperbolic structure,using the powerful feature representation capability of deep neural networks(DNN),and using the characteristics of the echo data itself under the constraints of GPR physical mechanism,such as amplitude,phase and geometric structure features,we propose a GPR target feature information detection method that combines the physical mechanism of GPR and DNN.Further feature analysis and applications based on effective extraction of target feature hyperbolas,such as estimation of target geometry and inversion of dielectric parameters,these studies can meet the automatic detection task of GPR targets in complex scenarios,which are of great scientific and social significance.It has important scientific significance and social value.The main contents of this paper are summarized as follows:(1)Aiming at the problem of poor generalization ability of deep learning network model under the condition of small samples,a GPR image data augmentation method based on cycle-consistent generative adversarial network(Cycle GAN)is proposed to generate B-SCAN images with high fidelity,which is used to train the Faster R-CNN model to improve the accuracy of identifying the target feature hyperbola.In addition,due to the complexity of the geological environment,the anchored target area in the B-SCAN image usually contains interference signals.In order to suppress the interference and accurately extract the target feature hyperbola,a hyperbolic segmentation method based on convolution operation is proposed.Taking full advantage of the characteristics of hyperbolic structure symmetry and its direction difference,the corresponding convolution kernel template is designed,and the effective segmentation of feature area in B-SCAN images is realized through convolution operation.Finally,the least squares(LS)method is used to fit the quadratic curve to obtain the accurate feature hyperbolic image of the target.The experimental results verify the effectiveness of the deep learning network for extracting GPR target feature information.(2)Aiming at the problem that the data-driven GPR B-SCAN image feature hyperbolic detection method is prone to false detection and missed detection,a feature hyperbola detection method combining GPR physical model-driven and data-driven is proposed.Under the constraints of the GPR physical mechanism,according to the intensity characteristics and hyperbolic structure characteristics of the target in the GPR echo data,a convolutional neural network with a cascade structure is designed.Firstly,detect and remove the direct wave interference signal in the echo data,and then use convolutional neural network(CNN)to extract the semantic feature map of the B-SCAN image,and use the classifier to identify the characteristic hyperbola of the target.In addition,in order to deal with the problem that various interference signals affect the structural integrity of target feature hyperbola,a feature data completion method based on direction guidance is proposed to improve the accuracy of target feature hyperbola recognition.The experimental results show that the cascaded CNN network designed under the constraints of the GPR physical mechanism can effectively detect the target feature information in complex scenes,which provides an effective method for practical application.The network model has more practical application significance.(3)Aiming at the problems of the traditional migration imaging methods,such as the difficulty in determining the migration parameters and the strong background noise in the migration imaging results,a method is proposed to estimate the geometric size of the target directly according to the hyperbolic curve of the target feature,and a more accurate estimation result is obtained.Since the feature curve of the lower boundary of the target in the GPR image data is relatively weak,the method to detect it in the local space combined with the amplitude and phase change information of the GPR scattering electric field is discussed,and a convolution kernel template for detecting the horizontal linear structure feature is designed.Convolution was performed with B-SCAN images in local space to accurately extract the lower boundary feature curves of the target.On this basis,the extracted target upper and lower boundary characteristic curves are used to estimate the target geometric size information,such as burial depth,width and height,and the parameter estimation results of the target geometric size can be automatically output according to the input GPR image data.(4)Aiming at the problem that it is difficult to accurately invert the scattering field of GPR based on the physical model-driven method(the scattering field equation is non-linear and ill-posed),a target permittivity estimation method combining model solving and example learning is proposed.The nonlinear mapping process corresponding to the GPR scattering field equation is expanded into a multi-layer regression network for optimization and solution.The ability of neural networks to obtain prior information by learning a large number of sample examples can be used to solve the problem of model selection and accuracy of model-based methods,while the problem of determining the topology structure of deep learning networks can be solved by using model methods.At the same time,by analyzing of the main factors affecting the target scattering intensity,electromagnetic wave attenuation and phase change in the subsurface half space,the relevant parameter information for inverting the permittivity of the target is determined,such as target amplitude,buried depth and the permittivity of the background medium,and feed them into a regression network for inverting the target permittivity,the estimation result of the permittivity of the target is output in an end-to-end way.