Dual-frequency ground penetrating radar underground target saliency feature fusion method

Abstract

The invention discloses a dual-frequency ground penetrating radar underground target saliency feature fusion method, which comprises the following steps of: before target feature detection, removing a direct wave strong interference signal in GPR echo data; designing a semantic feature extraction model, fusing a channel attention SE module and a space attention PSANet module into a feature extraction network, and adding phase consistency loss and phase change region gradient loss into a loss function during network model training; transpose convolution is carried out on the extracted low-frequency-band semantic feature map in the double-frequency GPR data, upsampling is carried out to the scale of the high-frequency-band semantic feature map, and high-frequency-band and low-frequency-band feature fusion is carried out. According to the fusion method, the high-frequency-band GPR echo signal contains target information with higher resolution, the low-frequency-band GPR can detect a target located at a deeper position underground, the contradiction problem between the GPR detection depth and the detection precision is effectively solved, and the method has the advantages of being high in precision and robustness.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of signal analysis and processing, and relates to feature fusion, in particular to a double-frequency ground penetrating radar underground target saliency feature fusion method. BACKGROUND

[0002] The information amount of data collected by a ground penetrating radar (GPR) system is related to the bandwidth of an antenna used, and a GPR can obtain higher resolution but has shallow detection depth when a high-frequency antenna is used, and has deep detection depth but low resolution when a low-frequency antenna is used. Different frequency GPR systems are used for data collection on the same survey line, which can effectively cope with the contradiction between detection depth and resolution. When the data collected by the double-frequency GPR system is jointly interpreted, the complementarity and redundancy between the data of different frequency bands should be fully utilized to mine more potential effective information, so as to obtain accurate and reliable detection results.

[0003] When the traditional method is used for fusion of GPR echo images of different frequencies, the features of images of different frequencies should be well matched, and the geological environment of the shallow surface layer is relatively complex, so in actual application, the GPR data of different frequencies usually need to be preprocessed, such as noise reduction, zero setting, registration, etc., and the precision and quality of GPR image fusion are difficult to guarantee. In addition, the traditional method is usually based on images, and the feature information in the original echo data is not fully utilized. SUMMARY

[0004] The technical problem to be solved by the application: the geophysical information detected by the GPR contains feature information of the target, and it is necessary to ensure that the GPR can detect a certain depth and also obtain higher resolution. The application provides a double-frequency ground penetrating radar underground target saliency feature fusion method, and proposes to use double-frequency GPR technology to make the echo data contain more complete target feature information. The construction of a double-frequency GPR data feature detection network model effectively extracts and fuses the target information components in the echo data, enhances the target features, and obtaining complete information representing the target is the problem to be solved by the application.

[0005] The technical scheme adopted by the application is as follows: A double-frequency ground penetrating radar underground target saliency feature fusion method, comprising the following steps: S1, data preprocessing: removing strong amplitude coupling direct wave interference signal preprocessing on the echo data collected by the double-frequency GPR; S2, semantic feature extraction: designing a semantic feature extraction network model for extracting target feature information in the GPR data, and using a parallel network architecture of a deep convolutional neural network combined with an attention mechanism to extract semantic feature information of the double-frequency GPR data; S3, make a data set and use it for training of a semantic feature extraction network model; add phase consistency loss and phase change area gradient loss in a loss function; S4, input high and low resolution GPR echo data collected by the dual-frequency GPR into the trained feature information extraction network respectively, and output GPR feature maps of different scales; S5, perform deconvolution operation on the low-resolution feature map and perform feature fusion with the high-resolution feature map, and output the fused target feature information.

[0006] When the traditional method is used to fuse the underground target features, the GPR image is usually processed, and the present application directly extracts features from the GPR raw echo data and fuses them, which is more efficient and accurate.

[0007] Firstly, according to the physical mechanism of GPR, a convolutional neural network model for extracting deep semantic features is designed according to the characteristics of the target with large amplitude intensity in the echo data, and high and low resolution echo data collected by the dual-frequency GPR are input into the semantic feature extraction network respectively, and the target semantic feature map is output; then, the geometric structure features and phase features of the target are combined for classification and recognition; finally, the low-resolution feature map is operated by deconvolution operation and the high-resolution feature map is fused, and the feature information of the extracted target is output.

[0008] The technical features and significant effects of the present application are as follows: The present application proposes a multi-resolution target feature information fusion method starting from GPR data, based on the analysis of the characteristics of multi-frequency GPR echo signals, combining the powerful feature expression ability of deep neural network, designing a low complexity network model under a unified network architecture, extracting different frequency GPR data features and fusing information. On the one hand, the attention mechanism is integrated into the semantic feature extraction network, and the model can more accurately focus on the target structure features under the constraint of GPR physical mechanism; on the other hand, the phase features of the echo signal are combined for target classification, and the robustness of the feature information extraction method is improved. The method proposed in the present application can effectively overcome the problem of weight selection of different resolution echo signals in traditional multi-frequency GPR data fusion, fuse the semantic features in different resolution GPR data, and effectively solve the contradiction between the detection depth and the detection accuracy of GPR. BRIEF DESCRIPTION OF DRAWINGS

[0009] Figure 1 is a flowchart of the underground target saliency feature fusion method of the dual-frequency ground penetrating radar of the present application; Figure 2 is a schematic diagram of the semantic feature extraction network structure. DETAILED DESCRIPTION

[0010] The application will be further described below in connection with the embodiments and drawings. Obviously, the described embodiments are only a part of the embodiments of the application, and not all the embodiments. Based on the embodiments of the application, all other embodiments obtained by those skilled in the art without creative work are within the protection scope of the application.

[0011] Embodiments A dual-frequency ground penetrating radar underground target saliency feature fusion method comprises the following steps: The collected GPR echo data is first positioned and the direct wave strong interference signal is removed by using the PCA algorithm, then the high and low frequency band GPR data feature information is extracted by using the trained semantic feature detection network, the low frequency band GPR feature map is transposed and convolution calculated to output the feature map with the same scale as the high frequency band, and finally the high and low frequency feature maps are information fused to obtain the target high resolution feature information.

[0012] As shown in Figure 1 A dual-frequency ground penetrating radar underground target saliency feature fusion method comprises the following steps: Step 1), the principal component analysis method is used to remove the direct wave strong interference signal of the echo data collected by the dual-frequency GPR, and the processes of the high frequency band and the low frequency band data of the dual-frequency GPR are the same, and the specific steps include: 1-1), assuming that the original data matrix collected is Data, the size is , wherein represents the number of time sampling points, represents the number of sampling channels (trace); 1-2), the square value of each channel signal, i.e. the energy, is calculated: , wherein N is the number of spatial measurement points, x(t, i) is the echo amplitude at time t and position i; 1-3), the time index corresponding to the energy peak value of each channel signal is determined , and according to the characteristics of the direct wave with strong intensity and short time, the time window corresponding to the direct wave is determined ; 1-4), the PCA algorithm is used to decompose Data into a background and an underground target effective signal low-rank matrix and a direct wave interference signal sparse matrix s; 1-5), in the sparse matrix s, the data in the time window of the direct wave is set to zero, and the remaining area is retained, and the specific implementation process includes: 1-5-1), a mask matrix M is constructed, wherein when , otherwise ; 1-5-2), perform mask operation on s: , ⊙ represents element-wise multiplication; 1-6) Merge the low-rank matrix and the processed sparse matrix, , output the reconstructed GPR data D after removing the direct wave; Step 2) Based on the principle that the greater the difference between the electrical characteristic parameters of the GPR underground half-space target and the background medium, the more obvious the intensity characteristics of the target in the echo data, the semantic feature extraction network is designed by combining the powerful feature expression ability of DCNN (Deep CNN) and the attention mechanism to obtain more complete target features. The parallel network architecture of the deep convolutional neural network combined with the attention mechanism is used to extract the semantic feature information of the dual-frequency GPR data respectively. The DCNN network structure for extracting the semantic feature information of the GPR data is as follows: Figure 2 As shown, the specific steps include: 2-1) In each convolutional layer of DCNN, convolution kernels of different scales, 3×3 and 5×5, are used to extract feature information, covering local details to global semantic information. At the same time, to prevent gradient vanishing and retain shallow features, a residual block of the ResNet structure is added after the multi-scale convolution; 2-2) The feature maps output by multi-scale convolution are spliced ​​in the channel dimension, and then the spatial dimension is compressed through global maximum pooling to generate a low-dimensional feature vector; 2-3), first use the channel attention SE module to select key features, then apply the spatial attention PSANet module to locate the target area and highlight the key features of the target in the GPR data; Step 3) Use GPRMAX electromagnetic simulation software to generate GPR data of different frequency bands, and combine the measured data to train the semantic feature detection network model. The specific steps include: 3-1) Divide the collected dual-frequency GPR data into training and test sets according to the high-frequency band and low-frequency band, and select the test set for model training; 3-2) Set relevant training parameters, including optimizer, learning rate, batch size, initialization weights, number of training rounds, regularization method, and loss function; In this embodiment, 5000 sets of GPR data are selected for model training. The optimizer uses the Adam algorithm, the learning rate is set to 0.001, the batch size is set to 160, the initialization weight adopts the He initialization method, the training round is set to 100, the Dropout regularization method is set to 0.2, and the loss function consists of three parts: mean absolute error function, phase consistency loss, and phase change region gradient loss. 3-3), according to the phase change characteristics of the target in the GPR echo signal, the phase consistency loss is added to the loss function and phase change region gradient loss , construct the total loss function as: , Where L1 represents the mean absolute error function, α represents the phase consistency loss weight, and β represents the gradient loss weight of the phase change region; , is the true phase, To predict the phase, e represents the natural exponential function; , Represents the vertical gradient of pixel (i, j), Represents the horizontal gradient of pixel (i, j); ε represents the smoothing parameter to avoid the denominator being zero; Step 4) Input the test set dual-frequency GPR data into the feature extraction model trained in step 3) and output the feature information detection results. The high-frequency and low-frequency data feature extraction processes are performed in parallel: The high-frequency part of the dual-frequency GPR data is input into the high-frequency band GPR data semantic feature extraction network, and the high-frequency band semantic feature map is output; The low-frequency part of the dual-frequency GPR data is input into the low-frequency band GPR data semantic feature extraction network, and the low-frequency band semantic feature map is output; Step 5) Align and fuse the feature maps of high and low frequency bands with different resolutions in the dual-frequency GPR data to output feature information containing a more complete structure of the target. The specific steps include: 5-1) Perform transposed convolution calculation on the extracted low-frequency GPR feature map, upsample it to the same size as the high-frequency GPR feature map, and complete the semantic feature map alignment operation; 5-2), the aligned semantic feature maps are concat-fused. The fused feature maps contain more complete target feature information and output information representing the saliency features of the target.

[0013] Through the above implementation cases, the advantages of this method are reflected in: On the one hand, in step 2), the GPR data feature extraction network is designed with GPR domain knowledge as a constraint, making the designed DCNN feature extraction model more interpretable; On the other hand, in step 3), in order to improve the accuracy of identifying target features, a phase constraint is added to the loss function, and the target classification is implemented by combining the intensity, structure and phase information of the target, which increases the robustness of the method. In addition, dual-frequency GPR is used for detection. The high-frequency GPR echo signal contains higher-resolution target information, and the low-frequency GPR can detect targets located deeper underground, effectively addressing the contradiction between GPR detection depth and detection accuracy.

[0014] In summary, the present invention effectively improves the performance of the underground target intelligent detection system.

[0015] Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the present invention, and that the scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for fusing salient features of underground targets using dual-frequency ground penetrating radar, characterized in that: The following steps are involved: S1, data preprocessing: preprocess the echo data collected by dual-frequency GPR to remove the strong amplitude coupled direct wave interference signal; S2, semantic feature extraction: Design a semantic feature extraction network model for extracting target feature information from GPR data. A parallel network architecture combining a deep convolutional neural network with an attention mechanism is used to extract semantic feature information from dual-frequency GPR data. S3, create a dataset and use it to train the semantic feature extraction network model; add phase consistency loss and phase change region gradient loss to the loss function; S4, inputting the high-resolution and low-resolution GPR echo data collected by the dual-frequency GPR into the trained feature information extraction network respectively, and outputting GPR feature maps of different scales; S5, performing a deconvolution operation on the low-resolution feature map and performing feature fusion with the high-resolution feature map, and outputting the fused target feature information.

2. The method for fusing underground target saliency features of a dual-frequency ground penetrating radar according to claim 1, characterized in that: The principal component analysis method is used to remove the strong interference signal of the direct wave from the echo data collected by the dual-frequency GPR. The process of processing the high-frequency band and low-frequency band data of the dual-frequency GPR is the same. The specific steps include: 1-1), assuming that the original data matrix collected is Data, and the size is ,in Indicates the number of time sampling points, Indicates the number of sampling channels (trace); 1-2), calculate the square value of each signal, that is, the energy: ， Where N is the number of spatial measurement points, and x(t,i) is the echo amplitude at time t and position i; 1-3), determine the time index corresponding to the energy peak of each signal , and according to the characteristics of the direct wave with high intensity and short time, determine the time window corresponding to the direct wave ; 1-4), PCA algorithm is used to decompose Data into low-rank matrices of background and underground target effective signals and the direct wave interference signal sparse matrix s; 1-5), in the sparse matrix s, the direct wave time window The data in the area is set to zero, and the rest of the area is retained; 1-6) Merge the low-rank matrix and the processed sparse matrix, , output the reconstructed GPR data D after removing the direct wave.

3. The method for fusing underground target saliency features of a dual-frequency ground penetrating radar according to claim 2, characterized in that: Steps 1-5) include: 1-5-1), construct the mask matrix M, where when ,otherwise ; 1-5-2), perform mask operation on s: , ⊙ represents element-wise multiplication.

4. The method for fusing underground target saliency features of a dual-frequency ground penetrating radar according to claim 1, characterized in that: S2 uses a parallel network architecture that combines a deep convolutional neural network with an attention mechanism to extract semantic feature information from dual-frequency GPR data. The specific steps include: 2-1) In each convolutional layer of DCNN, convolution kernels of different scales, 3×3 and 5×5, are used to extract feature information, covering local details to global semantic information. At the same time, to prevent gradient vanishing and retain shallow features, a residual block of the ResNet structure is added after the multi-scale convolution; 2-2) The feature maps output by multi-scale convolution are spliced ​​in the channel dimension, and then the spatial dimension is compressed through global maximum pooling to generate a low-dimensional feature vector; 2-3), first use the channel attention SE module to select key features, then apply the spatial attention PSANet module to locate the target area and highlight the key features of the target in the GPR data.

5. The method for fusing underground target saliency features of a dual-frequency ground penetrating radar according to claim 1, characterized in that: S3 generates GPR data in different frequency bands and combines it with the measured data to train the semantic feature detection network model. The specific steps include: 3-1) Divide the collected dual-frequency GPR data into training and test sets according to the high-frequency band and low-frequency band, and select the test set for model training; 3-2) Set relevant training parameters, including optimizer, learning rate, batch size, initialization weights, number of training rounds, regularization method, and loss function; 3-3), according to the phase change characteristics of the target in the GPR echo signal, the phase consistency loss is added to the loss function and phase change region gradient loss , construct the total loss function as: ， Where L1 represents the mean absolute error function, α represents the phase consistency loss weight, and β represents the gradient loss weight of the phase change region; ， is the true phase, To predict the phase, e represents the natural exponential function; ， Represents the vertical gradient of pixel (i, j), Represents the horizontal gradient of pixel (i, j); ε represents the smoothing parameter to avoid the denominator being zero.

6. The method for fusing underground target saliency features of a dual-frequency ground penetrating radar according to claim 1, characterized in that: S4 inputs the test set dual-frequency GPR data into the feature extraction model trained by S3 and outputs the feature information detection results. The high-frequency and low-frequency data feature extraction processes are performed in parallel: The high-frequency part of the dual-frequency GPR data is input into the high-frequency band GPR data semantic feature extraction network, and the high-frequency band semantic feature map is output; The low-frequency part of the dual-frequency GPR data is input into the low-frequency band GPR data semantic feature extraction network, and the low-frequency band semantic feature map is output.

7. The method for fusing underground target saliency features of a dual-frequency ground penetrating radar according to claim 1, characterized in that: The specific steps of S5 include: 5-1) Perform transposed convolution calculation on the extracted low-frequency GPR feature map, upsample it to the same size as the high-frequency GPR feature map, and complete the semantic feature map alignment operation; 5-2), the aligned semantic feature maps are concat-fused. The fused feature maps contain the target feature information and output information representing the saliency features of the target.

Images