Augmentation method and system for ground penetrating radar b-scan data

Abstract

The application discloses an augmented method and system for ground penetrating radar B-scan data, which comprises the following steps: simulating and constructing a ground penetrating radar B-scan primary data set and performing data preprocessing to construct a ground penetrating radar B-scan data set; based on a U-net network, a residual connection scheme, a skip connection scheme, a convolution scheme and a self-attention scheme, a ground penetrating radar B-scan data augmentation primary model is constructed and a ground penetrating radar B-scan data augmentation model is obtained through training; and the obtained ground penetrating radar B-scan data augmentation model is used for actual data augmentation of the ground penetrating radar B-scan data. The application can not only realize the augmentation of the ground penetrating radar B-scan data, but also has higher reliability, better accuracy and better data augmentation effect.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of geological exploration, and particularly relates to an augmented method and system for ground penetrating radar B-scan data. BACKGROUND

[0002] Ground penetrating radar (GPR) is a non-destructive testing technique that uses high-frequency electromagnetic waves to detect underground structures and targets. It is widely used in geological exploration, engineering detection, archaeological research, military mine detection, and other fields. GPR data is usually presented in the form of B-scan images, where the grayscale or color reflects the electromagnetic reflection characteristics of the underground medium.

[0003] At present, with the rapid development of machine learning and deep learning technologies, more and more researchers have begun to use machine learning and deep learning technologies for GPR target detection and recognition. In such tasks, high-quality and large-scale labeled data sets are the basis for high-performance GPR target detection and recognition models. However, the acquisition of GPR data often faces many challenges: (1) high data collection cost: requires professional equipment and personnel, and it is difficult to detect certain specific scenes (such as dangerous areas, sensitive areas). (2) Lack of sample diversity: in actual detection, the type, size, depth of the target and the surrounding environment are complex and variable, making it difficult to collect samples covering all situations. (3) Labeling difficulty: accurate labeling of targets in B-scan images requires professional domain knowledge and is time-consuming and labor-intensive. These factors result in GPR labeled data sets that are usually small in size and limited in diversity, severely affecting the training effect of deep learning models related to GPR.

[0004] To solve the problem of data scarcity, researchers have begun to use data augmentation techniques. Traditional data augmentation methods (such as rotation, scaling, translation, and noise addition) generate new samples by simply transforming existing data, but these methods fail to fundamentally increase the diversity of data, and the generated samples have little difference in feature distribution from the original samples, with limited augmentation effect. In recent years, researchers have begun to use image generation techniques based on generative adversarial networks (GAN) for GPR data augmentation. However, when GAN is directly applied to GPR B-scan data augmentation, there are the following problems: GPR B-scan images have unique echo characteristics and physical significance, and the fusion of targets and backgrounds needs to comply with certain physical laws, making it difficult for existing GAN models to generate realistic and physically consistent GPR data. SUMMARY

[0005] One of the purposes of the present application is to provide a reliable and accurate method for augmenting GPR B-scan data.

[0006] The second object of the present application is to provide a system for implementing the augmentation method of the ground penetrating radar B-scan data.

[0007] The present application provides an augmentation method of ground penetrating radar B-scan data, comprising the following steps:

[0008] S1. Set up a simulated detection scene, set the material composition parameters and position and size parameters of the underground medium and the target, and simulate and construct a ground penetrating radar B-scan primary data set;

[0009] S2. Data preprocessing is performed on the ground penetrating radar B-scan primary data set obtained in step S1 to construct a ground penetrating radar B-scan data set;

[0010] S3. Construct a ground penetrating radar B-scan data augmentation primary model; the model comprises an image separation network, a generator network and a discriminator network connected in sequence;

[0011] The image separation network is constructed based on a U-net network, a residual connection scheme, a skip connection scheme and a convolution scheme; the image separation network is used to separate the target features and background features of the input image;

[0012] The generator network is constructed based on a U-net network, a skip connection scheme, a convolution scheme and a self-attention scheme; the generator network is used to generate a single-channel fusion image according to the target features and background features obtained by separation;

[0013] The discriminator network is constructed based on a convolution scheme; the discriminator network is used to judge whether the input image is real ground penetrating radar B-scan data or a single-channel fusion image generated by the generator, and outputs the corresponding discrimination probability;

[0014] S4. The ground penetrating radar B-scan data augmentation primary model constructed in step S3 is trained using the ground penetrating radar B-scan data set obtained in step S2 to obtain a ground penetrating radar B-scan data augmentation model;

[0015] S5. The ground penetrating radar B-scan data augmentation model obtained in step S4 is used to perform data augmentation of actual ground penetrating radar B-scan data.

[0016] The step S1 specifically comprises the following steps:

[0017] Set up a simulated detection scene, set the parameters of the underground medium, set the material composition, buried position and size parameters of the target, and set up a simulated detection scene;

[0018] generating a ground penetrating radar B-Scan image including a background and a target and a ground penetrating radar B-Scan image including only the background, subtracting the ground penetrating radar B-Scan image including only the background from the ground penetrating radar B-Scan image including the background and the target to obtain a ground penetrating radar B-Scan image containing only the target;

[0019] changing parameters of the underground medium, setting material composition, buried position and size parameters of the target, obtaining a plurality of corresponding ground penetrating radar B-Scan images including a background and a target, ground penetrating radar B-Scan images including only the background, and ground penetrating radar B-Scan images containing only the target;

[0020] all the obtained ground penetrating radar B-Scan images are used to form a ground penetrating radar B-scan primary data set.

[0021] The data preprocessing in step S2 specifically includes the following steps:

[0022] The images in the ground penetrating radar B-scan primary data set are center cropped in the image center to set a pixel range, so as to ensure the uniformity of the image size;

[0023] The cropped images are converted into single-channel grayscale images;

[0024] The single-channel grayscale images are converted into tensor format, and the pixel values are scaled to a set range;

[0025] Finally, the pixel values are normalized.

[0026] The processing process of the image separation network specifically includes the following steps:

[0027] The constructed image separation network includes an image separation encoder and an image separation decoder;

[0028] The image separation encoder includes an image separation first convolutional layer, an image separation first batch normalization layer, an image separation first activation function layer, an image separation first residual connection layer, an image separation second convolutional layer, an image separation second batch normalization layer, an image separation second activation function layer, an image separation second residual connection layer, an image separation third convolutional layer, an image separation third batch normalization layer, an image separation third activation function layer, an image separation third residual connection layer, an image separation fourth convolutional layer, an image separation fourth batch normalization layer, an image separation fourth activation function layer, and an image separation fourth residual connection layer, which are sequentially connected; the output of the image separation first residual connection layer is represented as an image separation first encoding feature; the output of the image separation second residual connection layer is represented as an image separation second encoding feature; the output of the image separation third residual connection layer is represented as an image separation third encoding feature; and the output of the image separation fourth residual connection layer is represented as an image separation fourth encoding feature;

[0029] The image separation decoder comprises an image separation first up-sampling layer, an image separation second up-sampling layer, an image separation third up-sampling layer, an image separation fourth up-sampling layer, an image separation output up-sampling layer and an image separation activation layer which are sequentially connected in series; the image separation fourth encoding feature is taken as the input of the image separation first up-sampling layer, and the image separation third encoding feature is taken as the skip connection input of the image separation first up-sampling layer; the output of the image separation first up-sampling layer is taken as the input of the image separation second up-sampling layer, and the image separation second encoding feature is taken as the skip connection input of the image separation second up-sampling layer; the output of the image separation second up-sampling layer is taken as the input of the image separation third up-sampling layer, and the image separation first encoding feature is taken as the skip connection input of the image separation third up-sampling layer; the output of the image separation third up-sampling layer is taken as the input of the image separation fourth up-sampling layer, and the image separation first encoding feature is also taken as the skip connection input of the image separation fourth up-sampling layer;

[0030] The output of the image separation activation layer comprises the separated prediction target image feature and the prediction background image feature.

[0031] The image separation output up-sampling layer is a convolution layer.

[0032] The image separation first up-sampling layer to the image separation fourth up-sampling layer have the same structure; the processing process of the image separation up-sampling layer comprises the following steps:

[0033] The input of the image separation up-sampling layer is processed through up-sampling, and then processed through a convolution layer to obtain an input processing feature;

[0034] If the size of the input processing feature does not match the size of the skip connection input, the size of the input processing feature is adjusted through size matching processing to obtain an input processing matching feature;

[0035] The input processing matching feature and the skip connection input are spliced in the channel dimension, and then sequentially processed through a convolution layer, a normalization layer, an activation function layer and an up-sampling residual connection layer to obtain the output of the image separation up-sampling layer.

[0036] The image separation first residual connection layer to the image separation fourth residual connection layer and the up-sampling residual connection layer all have the same structure; the processing process of the residual connection layer comprises the following steps:

[0037] The input of the residual connection layer is sequentially processed through a convolution layer, a normalization layer and an activation function layer, and then sequentially processed through a convolution layer and a normalization layer to obtain a residual connection processing feature;

[0038] The residual connection processing feature and the input of the residual connection layer are summed by element addition to obtain the output of the residual connection layer.

[0039] The processing procedure of the generator network specifically comprises the following steps:

[0040] The generator network comprises a background encoding branch, a target encoding branch, a feature remodeling branch and a decoding branch;

[0041] The background encoding branch comprises, in sequence, a background encoding first convolutional layer, a background encoding first normalization layer, a background encoding first activation function layer, a background encoding second convolutional layer, a background encoding second normalization layer, a background encoding second activation function layer, a background encoding third convolutional layer, a background encoding third normalization layer and a background encoding third activation function layer;

[0042] The target encoding branch comprises, in sequence, a target encoding first convolutional layer, a target encoding first normalization layer, a target encoding first activation function layer, a target encoding second convolutional layer, a target encoding second normalization layer, a target encoding second activation function layer, a target encoding third convolutional layer, a target encoding third normalization layer and a target encoding third activation function layer;

[0043] The feature remodeling branch comprises, in sequence, a feature concatenation layer and a self-attention layer; the feature concatenation layer concatenates the output of the background encoding third activation function layer and the output of the target encoding third activation function layer along the channel dimension to obtain a fusion feature map; after the fusion feature map is spatially flattened, the self-attention layer is used to process the fusion feature map to obtain a remodeled feature;

[0044] The decoding branch comprises a decoding first convolutional layer, a decoding first normalization layer, a decoding first activation function layer, a decoding first concatenation layer, a decoding first convolutional fusion layer, a decoding second convolutional layer, a decoding second normalization layer, a decoding second activation function layer, a decoding second concatenation layer, a decoding second convolutional fusion layer, a decoding third convolutional layer, a decoding third normalization layer, a decoding third activation function layer, a decoding output convolutional layer and an output decoding output activation function layer;

[0045] The decoding first convolutional layer, the decoding first normalization layer and the decoding first activation function layer are sequentially connected; the input of the decoding first convolutional layer is the output of the self-attention layer;

[0046] The output of the background encoding third activation function layer and the output of the target encoding third activation function layer are both subjected to size matching through upsampling processing, and then the output of the generation encoding first splicing layer is spliced with the output of the generation decoding first activation function layer along the channel dimension, and the splicing result is taken as the output of the generation encoding first splicing layer; the output of the encoding first splicing layer is taken as the input of the generation decoding first convolution fusion layer, and the output of the generation decoding first convolution fusion layer is taken as the input of the generation decoding second convolution layer;

[0047] The generation decoding second convolution layer, the generation decoding second normalization layer and the generation decoding second activation function layer are sequentially connected;

[0048] The output of the background encoding second activation function layer and the output of the target encoding second activation function layer are both subjected to size matching through upsampling processing, and then the output of the generation encoding second splicing layer is spliced with the output of the generation decoding second activation function layer along the channel dimension, and the splicing result is taken as the output of the generation encoding second splicing layer; the output of the encoding second splicing layer is taken as the input of the generation decoding second convolution fusion layer, and the output of the generation decoding second convolution fusion layer is taken as the input of the generation decoding third convolution layer;

[0049] The generation decoding third convolution layer, the generation decoding third normalization layer, the generation decoding third activation function layer, the generation decoding output convolution layer and the output decoding output activation function layer are sequentially connected.

[0050] The processing process of the self-attention layer includes the following steps:

[0051] The input of the self-attention layer is divided into three paths: the first path is processed by the self-attention first convolution layer to obtain an initial query tensor, and the initial query tensor is transposed to obtain a query tensor; the second path is processed by the self-attention second convolution layer to obtain a key tensor; the third path is processed by the self-attention third convolution layer to obtain a value tensor;

[0052] The query tensor and the key tensor are processed by matrix multiplication, normalized and processed by a Softmax function to obtain an attention distribution map;

[0053] The attention distribution map and the value tensor are processed by matrix multiplication to obtain a weighted initial feature, and an inverse flattening operation is performed to obtain a weighted feature;

[0054] The weighted feature obtained is multiplied by the weighted parameter to be learned Then, the input of the self-attention layer is superimposed to obtain the output of the self-attention layer.

[0055] The processing process of the discriminator network includes the following steps:

[0056] The discriminator network comprises, in sequence, a discriminant first convolutional layer, a discriminant first normalization layer, a discriminant first activation function layer, a discriminant second convolutional layer, a discriminant second normalization layer, a discriminant second activation function layer, a discriminant third convolutional layer, a discriminant third normalization layer, a discriminant third activation function layer, a discriminant fourth convolutional layer, a discriminant fourth normalization layer, a discriminant fourth activation function layer, a discriminant output convolutional layer, and a discriminant output normalization layer.

[0057] The output of the generator network is processed by the discriminator network to obtain an output of the discriminator network.

[0058] The application also provides a system for implementing the data augmentation method of the ground penetrating radar B-scan data, comprising a data construction module, a data processing module, a model construction module, a model training module, and a data augmentation module; the data construction module, the data processing module, the model construction module, the model training module, and the data augmentation module are sequentially connected; the data construction module is used for setting a simulation detection scene, setting material composition parameters and position and size parameters of underground media and targets, and simulating and constructing a ground penetrating radar B-scan primary data set, and uploading data information to the data processing module; the data processing module is used for performing data preprocessing on the obtained ground penetrating radar B-scan primary data set according to the received data information, to construct a ground penetrating radar B-scan data set, and uploading data information to the model construction module; the model construction module is used for constructing a ground penetrating radar B-scan data augmentation primary model according to the received data information, and uploading data information to the model training module; the model comprises, in sequence, an image separation network, a generator network, and a discriminator network; the image separation network is constructed based on a U-net network, a residual connection scheme, a skip connection scheme, and a convolution scheme; the image separation network is used for separating target features and background features of an input image; the generator network is constructed based on a U-net network, a skip connection scheme, a convolution scheme, and a self-attention scheme; the generator network is used for generating a single-channel fusion image according to the separated target features and background features; the discriminator network is constructed based on a convolution scheme; the discriminator network is used for judging whether an input image is real ground penetrating radar B-scan data or a single-channel fusion image generated by the generator, and outputting a corresponding discrimination probability; the model training module is used for training the constructed ground penetrating radar B-scan data augmentation primary model by using the obtained ground penetrating radar B-scan data set, to obtain a ground penetrating radar B-scan data augmentation model, and uploading data information to the data augmentation module; and the data augmentation module is used for performing actual data augmentation of the ground penetrating radar B-scan data by using the obtained ground penetrating radar B-scan data augmentation model according to the received data information.

[0059] The application provides the augmented method and system of the ground penetrating radar B-scan data, the training data set is constructed through simulation, and the ground penetrating radar B-scan data augmentation model is constructed based on the U-net network, the residual connection scheme, the skip connection scheme, the convolution scheme and the self-attention scheme and is trained, so that the application can not only realize the augmentation of the ground penetrating radar B-scan data, but also has higher reliability, better accuracy and better data augmentation effect. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 The method flowchart of the method of the application.

[0061] Figure 2 The structure schematic diagram of the image separation network of the method of the application.

[0062] Figure 3 The structure schematic diagram of the image separation up-sampling layer of the method of the application.

[0063] Figure 4 The structure schematic diagram of the residual connection layer of the method of the application.

[0064] Figure 5 The structure schematic diagram of the generator network of the method of the application.

[0065] Figure 6 The structure schematic diagram of the self-attention layer of the method of the application.

[0066] Figure 7 The structure schematic diagram of the discriminator network of the method of the application.

[0067] Figure 8 The experimental result schematic diagram of the image separation network of the embodiment of the method of the application; wherein, Figure 8 (a) is the schematic diagram of the original B-scan image containing background and target, Figure 8 (b) is the schematic diagram of the target B-scan image generated by the separation network, Figure 8 (c) is the schematic diagram of the background B-scan image generated by the separation network, Figure 8 (d) is the schematic diagram of the real target B-scan image, Figure 8 (e) is the schematic diagram of the real background B-scan image.

[0068] Figure 9 The experimental result schematic diagram of the generator network of the embodiment of the method of the application; wherein, Figure 9 (a) is the schematic diagram of the background B-scan image generated by the image separation network, Figure 9 (b) is the schematic diagram of the target B-scan image generated by the image separation network, Figure 9(c) is a B-scan image diagram generated by the generator network, Figure 9 (d) is a real B-scan image diagram.

[0069] Figure 10 The input image diagram of the method embodiment of the application.

[0070] Figure 11 The background separation result diagram of the method embodiment of the application.

[0071] Figure 12 The target separation result diagram of the method embodiment of the application.

[0072] Figure 13 The Cartesian product combination diagram of the background and the target of the method embodiment of the application; wherein, Figure 13 (a) is a fusion effect diagram of the first separated background model_1 generated by the generator network in turn with five different separated targets, Figure 13 (b) is a fusion effect diagram of the second separated background model_2 generated by the generator network in turn with five different separated targets, Figure 13 (c) is a fusion effect diagram of the third separated background model_3 generated by the generator network in turn with five different separated targets, Figure 13 (d) is a fusion effect diagram of the fourth separated background model_4 generated by the generator network in turn with five different separated targets, Figure 13 (e) is a fusion effect diagram of the fifth separated background model_5 generated by the generator network in turn with five different separated targets.

[0073] Figure 14 The final generated image and the real image comparison diagram of the method embodiment of the application; wherein, Figure 14 (a) is a real B-scan image diagram, Figure 14 (b) is a generated B-scan image diagram.

[0074] Figure 15 The functional module diagram of the system of the application. DETAILED DESCRIPTION

[0075] As Figure 1 The method flow diagram of the method of the application is shown: the augmented method of ground penetrating radar B-scan data disclosed by the application comprises the following steps:

[0076] S1. Set up a simulated detection scene, set the material composition parameters and position size parameters of the underground medium and the target, and simulate and construct a ground penetrating radar B-scan primary data set; specifically comprising the following steps:

[0077] Setting an emulated detection scene, setting parameters of underground medium, setting material composition, buried position and size parameters of target;

[0078] Generating ground penetrating radar B-Scan images including background and target and ground penetrating radar B-Scan images including only background, subtracting ground penetrating radar B-Scan images including only background from ground penetrating radar B-Scan images including background and target to obtain ground penetrating radar B-Scan images containing only target;

[0079] Changing parameters of underground medium, setting material composition, buried position and size parameters of target, obtaining several corresponding ground penetrating radar B-Scan images including background and target, ground penetrating radar B-Scan images including only background and ground penetrating radar B-Scan images containing only target;

[0080] Constructing ground penetrating radar B-scan primary data set by using all obtained ground penetrating radar B-Scan images.

[0081] In specific implementation, when the simulation calculation is performed by the preset simulation software GprMax, first, a cubic grid with an edge length of 2 mm is established in a calculation domain with a size of 1.5 m*0.45 m*0.002 m, and a pseudo-random seed is taken as a driving force to randomly select one of the five soil models of sandy soil, clay, loam, mixed soil and low-loss high-dielectric-constant special soil as a matrix; the fractal_box command is applied to the selected soil to set the fractal dimension 1.5, the disturbance ratio 1, the weight 1 and the random volume fraction, and generate a fractal disturbed soil layer with grading statistical characteristics; on the basis of the soil background, according to the volume fractions of 74.29% aggregate, 20.71% asphalt mortar and 5% air void, the Gaussian random field is superimposed on the continuous dielectric constant distribution after the fractal disturbance to disperse the three-phase discrete medium, and the HDF5 format background model and the matching material list are output; then, the background model is executed twice with the same forward: the first time is target-free simulation, only the background medium is reserved; the second time is to randomly place a cylindrical metal target with a radius of 0.04 m and a height of 0.002 m in the same background, and keep the transmitting waveform, antenna trajectory, boundary condition and soil electromagnetic parameters completely consistent; the point-by-point subtraction of the Ez fields obtained by the two simulations obtains the difference field reflecting only the target scattering, and is saved in HDF5 format; further, the difference field and the corresponding "targeted" field and "target-free" field are mapped into a single-channel grayscale image with a size of 256*256 pixels and a resolution of 300 DPI, and after removing the coordinate axes and blank edges, the image is directly used as the network input; by changing the pseudo-random seed, samples covering various soil electromagnetic properties, different target configurations and spatial scenarios can be batch-generated, and the data set can be expanded on demand. 420 GPRB-scan images containing target echoes and background echoes, 420 GPRB-scan images without target echoes and background echoes and 420 GPRB-scan images with only target echoes and without background echoes are obtained by using the GPRMAX software simulation, which are used for network training. The size of all GPRB-scan images is 256*256*1.

[0082] S2. Data preprocessing is performed on the ground penetrating radar B-scan primary data set obtained in step S1 to construct a ground penetrating radar B-scan data set; specifically including the following steps:

[0083] The images in the ground penetrating radar B-scan primary data set are center cropped in a set pixel range (preferably 256*256 pixels) at the center of the image to ensure the uniformity of the image size;

[0084] The cropped images are converted into single-channel grayscale images;

[0085] The single-channel grayscale images are converted into tensor format, and the pixel values are scaled to a set range (preferably );

[0086] Finally, the pixel values are normalized (preferably normalized with a mean of 0.5 and a standard deviation of 0.5, so that the output tensor values are distributed in the interval [0, 1]);

[0087] S3. Construct a ground penetrating radar B-scan data augmentation primary model; the model comprises an image separation network, a generator network and a discriminator network connected in series;

[0088] The image separation network is constructed based on the U-net network, the residual connection scheme, the skip connection scheme and the convolution scheme; the image separation network is used to separate the target features and the background features of the input image;

[0089] The generator network is constructed based on the U-net network, the skip connection scheme, the convolution scheme and the self-attention scheme; the generator network is used to generate a single-channel fusion image according to the separated target features and background features;

[0090] The discriminator network is constructed based on the convolution scheme; the discriminator network is used to judge whether the input image is real ground penetrating radar B-scan data or a single-channel fusion image generated by the generator, and output the corresponding discrimination probability;

[0091] In specific implementation, the processing process of the image separation network specifically includes the following steps (as shown in Figure 2 FIG. 1):

[0092] The constructed image separation network comprises an image separation encoder and an image separation decoder;

[0093] The image separation encoder comprises, in sequence, an image separation first convolutional layer (with a step size of 2 and a padding of 1 and using a reflective padding mode), an image separation first batch normalization layer, an image separation first activation function layer (preferably a LeakyReLU activation function with a negative slope of 0.2), an image separation first residual connection layer, an image separation second convolutional layer (with a step size of 2 and a padding of 1), an image separation second batch normalization layer, an image separation second activation function layer (preferably a LeakyReLU activation function with a negative slope of 0.2), an image separation second residual connection layer, an image separation third convolutional layer (with a step size of 2 and a padding of 1), an image separation third batch normalization layer, an image separation third activation function layer (preferably a LeakyReLU activation function with a negative slope of 0.2), an image separation third residual connection layer, an image separation fourth convolutional layer (with a step size of 2 and a padding of 1), an image separation fourth batch normalization layer, an image separation fourth activation function layer (preferably a LeakyReLU activation function with a negative slope of 0.2), and an image separation fourth residual connection layer; the output of the image separation first residual connection layer represents an image separation first encoded feature; the output of the image separation second residual connection layer represents an image separation second encoded feature; the output of the image separation third residual connection layer represents an image separation third encoded feature; the output of the image separation fourth residual connection layer represents an image separation fourth encoded feature; and each of the first to fourth convolutional layers uses a 4x4 convolutional kernel;

[0094] The image separation decoder comprises, in sequence, an image separation first up-sampling layer, an image separation second up-sampling layer, an image separation third up-sampling layer, an image separation fourth up-sampling layer, an image separation output up-sampling layer, and an image separation activation layer; the image separation fourth encoded feature is input to the image separation first up-sampling layer, and the image separation third encoded feature is input to the image separation first up-sampling layer as a skip connection input; the output of the image separation first up-sampling layer is input to the image separation second up-sampling layer, and the image separation second encoded feature is input to the image separation second up-sampling layer as a skip connection input; the output of the image separation second up-sampling layer is input to the image separation third up-sampling layer, and the image separation first encoded feature is input to the image separation third up-sampling layer as a skip connection input; the output of the image separation third up-sampling layer is input to the image separation fourth up-sampling layer, and the image separation first encoded feature is also input to the image separation fourth up-sampling layer as a skip connection input to enhance high-resolution features, thereby forming a double skip connection structure;

[0095] In the processing of the image separation fourth up-sampling layer, the actual up-sampling operation is actually constrained due to the size limitation of the skip connection feature, and finally a global up-sampling mechanism is used at the end of the decoder to ensure that the output feature map reaches the preset image size requirement;

[0096] The output of the image separation activation layer includes separated prediction target image features and prediction background image features.

[0097] The image separation output up-sampling layer is a convolution layer, the convolution kernel is 3*3, the step is 1, the padding is 1, and the reflection padding mode is adopted.

[0098] In specific implementation, the image separation first up-sampling layer to the image separation fourth up-sampling layer have the same structure; the processing process (structure as shown in Figure 3 FIG. 1) of the image separation up-sampling layer includes the following steps:

[0099] The input of the image separation up-sampling layer is processed by up-sampling, and then processed by a convolution layer to obtain input processing features;

[0100] If the size of the input processing features does not match the size of the skip connection input, the size of the input processing features is adjusted by size matching processing to obtain input processing matching features;

[0101] The input processing matching features and the skip connection input are spliced in the channel dimension, and then sequentially processed by a convolution layer, a normalization layer, an activation function layer and an up-sampling residual connection layer to obtain the output of the image separation up-sampling layer;

[0102] The image separation first residual connection layer to the image separation fourth residual connection layer and the up-sampling residual connection layer have the same structure; the processing process (as shown in Figure 4 FIG. 2) of the residual connection layer includes the following steps:

[0103] The input of the residual connection layer is sequentially processed by a convolution layer, a normalization layer and an activation function layer, and then sequentially processed by a convolution layer and a normalization layer to obtain residual connection processing features;

[0104] The residual connection processing features and the input of the residual connection layer are summed by element addition to obtain the output of the residual connection layer.

[0105] In specific implementation, the processing process (as shown in Figure 5 FIG. 3) of the generator network specifically includes the following steps:

[0106] The generator network includes a generated background encoding branch, a generated target encoding branch, a feature remodeling branch and a generated decoding branch;

[0107] The background generation coding branch comprises, in sequence, a background generation coding first convolutional layer (with a convolution kernel of 4*4, a step length of 2 and padding of 1), a background generation coding first normalization layer, a background generation coding first activation function layer (using a LeakyReLU activation function), a background generation coding second convolutional layer (with a convolution kernel of 4*4, a step length of 2 and padding of 1), a background generation coding second normalization layer, a background generation coding second activation function layer (using a LeakyReLU activation function), a background generation coding third convolutional layer (with a convolution kernel of 4*4, a step length of 2 and padding of 1), a background generation coding third normalization layer and a background generation coding third activation function layer (using a LeakyReLU activation function);

[0108] The target generation coding branch comprises, in sequence, a target generation coding first convolutional layer (with a convolution kernel of 4*4, a step length of 2 and padding of 1), a target generation coding first normalization layer, a target generation coding first activation function layer (using a LeakyReLU activation function), a target generation coding second convolutional layer (with a convolution kernel of 4*4, a step length of 2 and padding of 1), a target generation coding second normalization layer, a target generation coding second activation function layer (using a LeakyReLU activation function), a target generation coding third convolutional layer (with a convolution kernel of 4*4, a step length of 2 and padding of 1), a target generation coding third normalization layer and a target generation coding third activation function layer (using a LeakyReLU activation function);

[0109] The feature remodeling branch comprises, in sequence, a feature splicing layer and a self-attention layer; the feature splicing layer splices the output of the background generation coding third activation function layer and the output of the target generation coding third activation function layer along a channel dimension to obtain a fusion feature map; after the fusion feature map is spatially flattened, the self-attention layer is used to process the fusion feature map to obtain a remodeled feature;

[0110] The decoding generation branch comprises a decoding generation first convolutional layer (with a convolution kernel of 4*4, a step length of 2 and padding of 1), a decoding generation first normalization layer, a decoding generation first activation function layer (using a ReLU activation function), a decoding generation first splicing layer, a decoding generation first convolutional fusion layer (using a 1*1 convolution), a decoding generation second convolutional layer (with a convolution kernel of 4*4, a step length of 2 and padding of 1), a decoding generation second normalization layer, a decoding generation second activation function layer (using a ReLU activation function), a decoding generation second splicing layer, a decoding generation second convolutional fusion layer (using a 1*1 convolution), a decoding generation third convolutional layer (with a convolution kernel of 4*4, a step length of 2 and padding of 1), a decoding generation third normalization layer, a decoding generation third activation function layer (using a ReLU activation function), a decoding generation output convolutional layer (with a convolution kernel of 3*3, a step length of 1 and padding of 1) and an output decoding output activation function layer (using a Tanh activation function);

[0111] The decoding first convolutional layer, the decoding first normalization layer, and the decoding first activation function layer are sequentially connected in series; the input of the decoding first convolutional layer is the output of the self-attention layer;

[0112] The output of the background encoding third activation function layer and the output of the target encoding third activation function layer are both subjected to size matching through upsampling processing, then are spliced along the channel dimension with the output of the decoding first activation function layer through the decoding first splicing layer, and the splicing result is taken as the output of the decoding first splicing layer; the output of the decoding first splicing layer is taken as the input of the decoding first convolutional fusion layer, and the output of the decoding first convolutional fusion layer is taken as the input of the decoding second convolutional layer;

[0113] The decoding second convolutional layer, the decoding second normalization layer, and the decoding second activation function layer are sequentially connected in series;

[0114] The output of the background encoding second activation function layer and the output of the target encoding second activation function layer are both subjected to size matching through upsampling processing, then are spliced along the channel dimension with the output of the decoding second activation function layer through the decoding second splicing layer, and the splicing result is taken as the output of the decoding second splicing layer; the output of the decoding second splicing layer is taken as the input of the decoding second convolutional fusion layer, and the output of the decoding second convolutional fusion layer is taken as the input of the decoding third convolutional layer;

[0115] The decoding third convolutional layer, the decoding third normalization layer, the decoding third activation function layer, the decoding output convolutional layer, and the decoding output activation function layer are sequentially connected in series.

[0116] In a specific implementation, the processing procedure of the self-attention layer (as shown in Figure 6 The processing procedure of the self-attention layer includes the following steps.

[0117] The input of the self-attention layer is divided into three paths: the first path is subjected to processing through the self-attention first convolutional layer (using a 1*1 convolution) to obtain an initial query tensor, and the initial query tensor is transposed to obtain a query tensor; the second path is subjected to processing through the self-attention second convolutional layer (using a 1*1 convolution) to obtain a key tensor; and the third path is subjected to processing through the self-attention third convolutional layer (using a 1*1 convolution) to obtain a value tensor;

[0118] The query tensor and the key tensor are subjected to processing through matrix multiplication, normalization, and a Softmax function to obtain an attention distribution map;

[0119] The attention distribution map and the value tensor are subjected to processing through matrix multiplication to obtain a weighted initial feature, and an inverse flattening operation is performed to obtain a weighted feature;

[0120] The obtained weighted features are multiplied by the weighted parameters to be learned The output of the self-attention layer is obtained by superimposing the input of the self-attention layer.

[0121] In a specific implementation, the processing procedure of the discriminator network (as shown in the figure) specifically includes the following steps: Figure 7

[0122] The discriminator network includes, in sequence, a first discriminant convolutional layer (with a convolution kernel of 4*4, a step length of 2, and a padding of 1), a first discriminant normalization layer (using spectral normalization), a first discriminant activation function layer, a second discriminant convolutional layer (with a convolution kernel of 4*4, a step length of 2, and a padding of 1), a second discriminant normalization layer (using spectral normalization), a second discriminant activation function layer, a third discriminant convolutional layer (with a convolution kernel of 4*4, a step length of 2, and a padding of 1), a third discriminant normalization layer (using spectral normalization), a third discriminant activation function layer, a fourth discriminant convolutional layer (with a convolution kernel of 4*4, a step length of 2, and a padding of 1), a fourth discriminant normalization layer (using spectral normalization), a fourth discriminant activation function layer, a discriminant output convolutional layer (with a convolution kernel of 4*4, a step length of 2, and a padding of 1), and a discriminant output normalization layer (using spectral normalization).

[0123] The output of the generator network is processed by the discriminator network to obtain the output of the discriminator network.

[0124] S4. Using the ground penetrating radar B-scan data set obtained in step S2, the ground penetrating radar B-scan data augmentation primary model constructed in step S3 is trained to obtain a ground penetrating radar B-scan data augmentation model; specifically including the following steps:

[0125] During the training process, the image separation network is trained alone, and the generator network and the discriminator network are trained together.

[0126] During the training process of the image separation network, the weighted sum of the mean absolute error loss, the joint similarity loss, and the perception loss is selected as the total loss function, the pre-trained deep neural network (preferably the first 9 layers of convolutional features of VGG16) is used to extract high-level features, and the L1 distance is used to measure the difference between the predicted result and the true value in the feature space; the optimizer uses the Adam algorithm, and the initial learning rate is set to During the training, an early stopping mechanism is set, and when the validation loss is lower than the historical best value, the current network parameters are automatically saved;

[0127] ​In the training process of the generator network and the discriminator network, a weighted sum value of the adversarial loss, the L1 loss and the perceptual loss is selected as a total loss function; an Adam optimizer is used for parameter updating. The learning rate of the generator is set to 0.0001; the cosine annealing strategy is used for learning rate scheduling, the training period is set to 2000 rounds, the minimum learning rate is set to 0.000001; the learning rate gradually decreases from the initial value according to the cosine function law, and decreases to the minimum value at 2000 rounds; the strategy realizes fine adjustment of model parameters by reducing the learning rate in the later training period, reduces the oscillation in the convergence process, and improves the final convergence accuracy; the training adopts an asymmetric strategy of discriminator priority, that is, in each iteration, the generator parameters are fixed to train the discriminator for 3 times, and then the discriminator parameters are fixed to train the generator for 1 time. This design is based on the characteristic that the discriminator converges faster than the generator, and through multiple rounds of training, the discriminability of the discriminator to the sample authenticity is enhanced, the generator is prevented from falling into a suboptimal solution due to early deception of the discriminator, and the effective training of the generative adversarial network is ensured; the early stopping strategy sets double early stopping conditions to avoid model overfitting and retain optimal parameters. The comprehensive score is calculated by weighting the peak signal-to-noise ratio (PSNR), the structural similarity (SSIM) and the Fréchet Inception distance (FID) with weights of 1, 10 and -1 respectively, and the higher the score, the better the comprehensive quality of the generated image. If the verification score does not improve for 600 consecutive rounds, or the FID value does not decrease for 300 consecutive rounds, the training is terminated and the current optimal generator parameters are saved. This mechanism monitors the performance of the validation set, stops training in time, saves computing resources and prevents overfitting.

[0128] S5. Adopting the ground penetrating radar B-scan data augmentation model obtained in step S4 to perform data augmentation of actual ground penetrating radar B-scan data; in the application, only the trained image separation network and the generator network in the ground penetrating radar B-scan data augmentation model obtained in step S4 are used, and the image generated by the generator network is the data augmentation result of the ground penetrating radar B-scan data finally obtained; the discriminator network is only used in the training process of the model and does not participate in the data augmentation process.

[0129] The effects of the method are further described below in combination with an embodiment:

[0130] The image separation network is constructed based on the U-net network, the residual connection scheme, the skip connection scheme and the convolution scheme; the image separation network is used to separate the target features and the background features of the input image; the experimental results are shown in Figure 8 ;

[0131] The saved best model weight (based on the lowest loss of the validation set) is loaded into the image separation network architecture, and the test set is forward inferred to generate the predicted output background image and target image. Then, the evaluation indicators are calculated for the background branch and the target branch respectively: for the background branch, the L1 error is 0.0025, the PSNR is 52.24 dB, the SSIM is 0.9985, and the FID is 2.04, indicating that the network has extremely high pixel-level accuracy, structural fidelity and feature distribution consistency in background echo reconstruction; for the target branch, the L1 error is 0.0387, the PSNR is 32.16 dB, the SSIM is 0.9574, and the FID is 33.03, showing that the network can effectively separate the target echo and retain key details, while maintaining good consistency in perceptual quality with the real target image. The average FID value is 17.54, further verifying the closeness of the overall distribution of the generated image to the real image. These indicators collectively demonstrate the excellent performance of the image separation network in separating GPR B-Scan images under complex geological conditions.

[0132] The generator network is constructed based on the U-net network, the skip connection scheme, the convolution scheme and the self-attention scheme; the generator network is used to generate a single-channel fusion image according to the separated target feature and background feature; the experimental results are as shown in Figure 9

[0133] The best model weight is loaded into the network architecture, and the test set is forward inferred to obtain the predicted output. Then, the indicators are calculated for the fusion branch: the PSNR is 55.51 dB, and the SSIM is 0.9984, indicating that the network has extremely high pixel-level accuracy and structural fidelity in fusion reconstruction; at the same time, the Fréchet Inception Distance (FID) is 0.26, showing that the network effectively captures the distribution characteristics of the fusion image while maintaining reasonable overall perceptual similarity.

[0134] Then, the obtained ground penetrating radar B-scan data augmentation model is used to perform actual ground penetrating radar B-scan data augmentation: the input image is as shown in Figure 10 , the background separation result is as shown in Figure 11 , the target separation result is as shown in Figure 12 , the Cartesian product combination of the background and the target is as shown in Figure 13 , and the comparison chart of the finally generated image and the real image is as shown in Figure 14 .

[0135] ​Through intuitive comparison between the generated image and the real image, it can be clearly observed that both have high consistency in vision and are difficult to distinguish, which preliminarily verifies the effectiveness of the method from a qualitative perspective. Further quantitative evaluation results provide strong data support for this: the peak signal-to-noise ratio (PSNR) of the generated data is as high as 52.06, and the structural similarity (SSIM) index reaches 0.996, which shows that the generated image is almost the same as the real image in terms of pixel-level fidelity and structural information integrity; at the same time, the extremely low Fréchet inception distance (FID) score (0.896) proves that both are extremely close in high-level feature distribution, and have excellent visual authenticity. Finally, this embodiment proves from multiple dimensions that the scheme can generate a large amount of high-quality and high-fidelity B-scan data.

[0136] As Figure 15The system of the application is shown in the functional module diagram: the system for implementing the augmented method of the ground penetrating radar B-scan data disclosed in the application comprises a data construction module, a data processing module, a model construction module, a model training module and a data augmentation module; the data construction module, the data processing module, the model construction module, the model training module and the data augmentation module are sequentially connected; the data construction module is used for setting a simulation detection scene, setting the material composition parameters and the position and size parameters of the underground medium and the target, simulating and constructing a ground penetrating radar B-scan primary data set, and uploading the data information to the data processing module; the data processing module is used for performing data preprocessing on the obtained ground penetrating radar B-scan primary data set according to the received data information, constructing a ground penetrating radar B-scan data set, and uploading the data information to the model construction module; the model construction module is used for constructing a ground penetrating radar B-scan data augmentation primary model according to the received data information, and uploading the data information to the model training module; the model comprises an image separation network, a generator network and a discriminator network which are sequentially connected; the image separation network is constructed based on a U-net network, a residual connection scheme, a skip connection scheme and a convolution scheme; the image separation network is used for separating the target features and the background features of the input image; the generator network is constructed based on a U-net network, a skip connection scheme, a convolution scheme and a self-attention scheme; the generator network is used for generating a single-channel fusion image according to the separated target features and background features; the discriminator network is constructed based on a convolution scheme; the discriminator network is used for judging whether the input image is real ground penetrating radar B-scan data or a single-channel fusion image generated by the generator, and outputting the corresponding discrimination probability; the model training module is used for training the constructed ground penetrating radar B-scan data augmentation primary model by using the obtained ground penetrating radar B-scan data set, obtaining a ground penetrating radar B-scan data augmentation model, and uploading the data information to the data augmentation module; the data augmentation module is used for performing actual data augmentation of the ground penetrating radar B-scan data by using the obtained ground penetrating radar B-scan data augmentation model according to the received data information.

Claims

1. An augmentation method for ground penetrating radar B-scan data, the method comprising: Comprise the following steps: S1. Set up a simulated detection scene, set the material composition parameters and position and size parameters of the underground medium and target, and simulate and construct a ground penetrating radar B-scan primary data set; S2. Data preprocessing is performed on the ground penetrating radar B-scan primary data set obtained in step S1 to construct a ground penetrating radar B-scan data set; S3. Construct a ground penetrating radar B-scan data augmentation primary model; the model comprises an image separation network, a generator network and a discriminator network connected in series; The image separation network is constructed based on a U-net network, a residual connection scheme, a skip connection scheme and a convolution scheme; the image separation network is used to separate the target features and background features of the input image; The generator network is constructed based on a U-net network, a skip connection scheme, a convolution scheme and a self-attention scheme; The generator network is used to generate a single-channel fusion image according to the target features and background features obtained by separation; The discriminator network is constructed based on a convolution scheme; the discriminator network is used to judge whether the input image is real ground penetrating radar B-scan data or a single-channel fusion image generated by the generator, and outputs the corresponding discrimination probability; S4. The ground penetrating radar B-scan data augmentation model is obtained by training the ground penetrating radar B-scan data augmentation primary model constructed in step S3 using the ground penetrating radar B-scan data set obtained in step S2; S5. The ground penetrating radar B-scan data augmentation model obtained in step S4 is used to perform actual ground penetrating radar B-scan data augmentation; In specific implementation, the processing process of the generator network specifically comprises the following steps: The generator network comprises a background encoding branch, a target encoding branch, a feature remodeling branch and a generation decoding branch; The background encoding branch comprises a background encoding first convolutional layer, a background encoding first normalization layer, a background encoding first activation function layer, a background encoding second convolutional layer, a background encoding second normalization layer, a background encoding second activation function layer, a background encoding third convolutional layer, a background encoding third normalization layer and a background encoding third activation function layer connected in series; The target encoding branch comprises a target encoding first convolutional layer, a target encoding first normalization layer, a target encoding first activation function layer, a target encoding second convolutional layer, a target encoding second normalization layer, a target encoding second activation function layer, a target encoding third convolutional layer, a target encoding third normalization layer and a target encoding third activation function layer connected in series; The feature remodeling branch comprises a feature concatenation layer and a self-attention layer connected in series; The feature concatenation layer concatenates the outputs of the background encoding third activation function layer and the target encoding third activation function layer along the channel dimension to obtain a fusion feature map; after spatial flattening of the fusion feature map, the remodeling feature is obtained by processing through the self-attention layer; The generating decoding branch comprises a generating decoding first convolutional layer, a generating decoding first normalization layer, a generating decoding first activation function layer, a generating encoding first splicing layer, a generating decoding first convolutional fusion layer, a generating decoding second convolutional layer, a generating decoding second normalization layer, a generating decoding second activation function layer, a generating encoding second splicing layer, a generating decoding second convolutional fusion layer, a generating decoding third convolutional layer, a generating decoding third normalization layer, a generating decoding third activation function layer, a generating decoding output convolutional layer and an output decoding output activation function layer; The generating decoding first convolutional layer, the generating decoding first normalization layer and the generating decoding first activation function layer are sequentially connected; the input of the decoding first convolutional layer is the output of the self-attention layer; The output of the generating background encoding third activation function layer and the output of the generating target encoding third activation function layer are subjected to size matching through upsampling processing, and then are spliced along the channel dimension through the output of the generating decoding first activation function layer and the output of the generating encoding first splicing layer, and the splicing result is taken as the output of the generating encoding first splicing layer; the output of the encoding first splicing layer is taken as the input of the generating decoding first convolutional fusion layer, and the output of the generating decoding first convolutional fusion layer is taken as the input of the generating decoding second convolutional layer; The generating decoding second convolutional layer, the generating decoding second normalization layer and the generating decoding second activation function layer are sequentially connected; The output of the generating background encoding second activation function layer and the output of the generating target encoding second activation function layer are subjected to size matching through upsampling processing, and then are spliced along the channel dimension through the output of the generating decoding second activation function layer and the output of the generating encoding second splicing layer, and the splicing result is taken as the output of the generating encoding second splicing layer; the output of the encoding second splicing layer is taken as the input of the generating decoding second convolutional fusion layer, and the output of the generating decoding second convolutional fusion layer is taken as the input of the generating decoding third convolutional layer; The generating decoding third convolutional layer, the generating decoding third normalization layer, the generating decoding third activation function layer, the generating decoding output convolutional layer and the output decoding output activation function layer are sequentially connected.

2. The augmented method of ground penetrating radar B-scan data of claim 1, wherein The step S1 specifically comprises the following steps: An artificial detection scene is set, parameters of underground medium are set, and material composition, buried depth position and size parameters of the target are set; A ground penetrating radar B-Scan image including a background and a target and a ground penetrating radar B-Scan image including only the background are generated, and the ground penetrating radar B-Scan image including the background and the target is subtracted from the ground penetrating radar B-Scan image including only the background to obtain a ground penetrating radar B-Scan image containing only the target; Parameters of underground medium are changed, material composition, buried depth position and size parameters of the target are set, and a plurality of corresponding ground penetrating radar B-Scan images including a background and a target, ground penetrating radar B-Scan images including only the background and ground penetrating radar B-Scan images containing only the target are obtained; All the obtained ground penetrating radar B-Scan images constitute a ground penetrating radar B-scan primary data set.

3. The augmented method of ground penetrating radar B-scan data of claim 2, wherein The data preprocessing of the step S2 specifically comprises the following steps: The image in the ground penetrating radar B-scan primary data set is centrally cropped in the center of the image to set the center of the pixel range, so as to ensure the uniformity of the image size; The cropped image is converted into a single-channel grayscale image; The single-channel grayscale image is converted into a tensor format, and the pixel value is scaled to a set range; Finally, the pixel value is normalized.

4. The augmented method of ground penetrating radar B-scan data of claim 1, wherein The processing process of the image separation network specifically includes the following steps: The constructed image separation network includes an image separation encoder and an image separation decoder; The image separation encoder includes an image separation first convolutional layer, an image separation first batch normalization layer, an image separation first activation function layer, an image separation first residual connection layer, an image separation second convolutional layer, an image separation second batch normalization layer, an image separation second activation function layer, an image separation second residual connection layer, an image separation third convolutional layer, an image separation third batch normalization layer, an image separation third activation function layer, an image separation third residual connection layer, an image separation fourth convolutional layer, an image separation fourth batch normalization layer, an image separation fourth activation function layer, and an image separation fourth residual connection layer, which are sequentially connected in series; The output of the image separation first residual connection layer is represented as an image separation first encoding feature; The output of the image separation second residual connection layer is represented as an image separation second encoding feature; The output of the image separation third residual connection layer is represented as an image separation third encoding feature; The output of the image separation fourth residual connection layer is represented as an image separation fourth encoding feature; The image separation decoder includes an image separation first upsampling layer, an image separation second upsampling layer, an image separation third upsampling layer, an image separation fourth upsampling layer, an image separation output upsampling layer, and an image separation activation layer, which are sequentially connected in series; the image separation fourth encoding feature is input to the image separation first upsampling layer, and the image separation third encoding feature is input to the image separation first upsampling layer as a skip connection; the output of the image separation first upsampling layer is input to the image separation second upsampling layer, and the image separation second encoding feature is input to the image separation second upsampling layer as a skip connection; the output of the image separation second upsampling layer is input to the image separation third upsampling layer, and the image separation first encoding feature is input to the image separation third upsampling layer as a skip connection; the output of the image separation third upsampling layer is input to the image separation fourth upsampling layer, and the image separation first encoding feature is also input to the image separation fourth upsampling layer as a skip connection; The output of the image separation activation layer includes a predicted target image feature and a predicted background image feature separated; The image separation output upsampling layer is a convolutional layer.

5. The augmented method of ground penetrating radar B-scan data of claim 4, wherein The image separation first upsampling layer and the image separation fourth upsampling layer have the same structure; the processing process of the image separation upsampling layer includes the following steps: The input of the image separation upsampling layer is processed by upsampling, and then processed by a convolutional layer to obtain an input processing feature; If the size of the input processing feature does not match the size of the skip connection input, the size of the input processing feature is adjusted by size matching processing to obtain an input processing matching feature; The input processing matching feature and the jump connection input are spliced in a channel dimension, and then sequentially processed by a convolution layer, a normalization layer, an activation function layer, and an up-sampling residual connection layer to obtain an output of the image separation up-sampling layer.

6. The augmented method of ground penetrating radar B-scan data of claim 5, wherein The image separation first residual connection layer, the image separation fourth residual connection layer, and the up-sampling residual connection layer all have the same structure; the processing process of the residual connection layer includes the following steps: The input of the residual connection layer is sequentially processed by a convolution layer, a normalization layer, and an activation function layer, and then sequentially processed by a convolution layer and a normalization layer to obtain residual connection processing features; The residual connection processing features and the input of the residual connection layer are summed by element addition to obtain the output of the residual connection layer.

7. The augmented method of ground penetrating radar B-scan data of claim 6, wherein The processing process of the self-attention layer specifically includes the following steps: The input of the self-attention layer is divided into three paths: the first path is processed by a self-attention first convolution layer to obtain a query initial tensor, and the query initial tensor is transposed to obtain a query tensor; the second path is processed by a self-attention second convolution layer to obtain a key tensor; and the third path is processed by a self-attention third convolution layer to obtain a value tensor; The query tensor and the key tensor are processed by matrix multiplication, normalized, and processed by a Softmax function to obtain an attention distribution map; The attention distribution map and the value tensor are processed by matrix multiplication to obtain weighted initial features, and the weighted initial features are de-flattened to obtain weighted features; The obtained weighted features are multiplied by the weighted parameters to be learned After that, the output of the self-attention layer is obtained by superimposing the input of the self-attention layer.

8. The augmented method of ground penetrating radar B-scan data of claim 7, wherein The processing process of the discriminator network specifically includes the following steps: The discriminator network includes a discriminant first convolution layer, a discriminant first normalization layer, a discriminant first activation function layer, a discriminant second convolution layer, a discriminant second normalization layer, a discriminant second activation function layer, a discriminant third convolution layer, a discriminant third normalization layer, a discriminant third activation function layer, a discriminant fourth convolution layer, a discriminant fourth normalization layer, a discriminant fourth activation function layer, a discriminant output convolution layer, and a discriminant output normalization layer, which are sequentially connected. The output of the generator network is processed by the discriminator network to obtain the output of the discriminator network.

9. A system for implementing the augmentation method of the B-scan data of a ground penetrating radar according to one of claims 1 to 8, characterized in that The data construction module, the data processing module, the model construction module, the model training module, and the data augmentation module are sequentially connected; the data construction module is used to set a simulation detection scene, set material composition parameters and position and size parameters of underground media and targets, and simulate and construct a ground penetrating radar B-scan primary data set, and upload data information to the data processing module; The data processing module is used to perform data preprocessing on the obtained ground penetrating radar B-scan primary data set according to the received data information, to construct a ground penetrating radar B-scan data set, and upload data information to the model construction module; The model construction module is configured to construct a ground penetrating radar B-scan data augmentation primary model according to the received data information, and upload the data information to the model training module; the model comprises an image separation network, a generator network and a discriminator network connected in sequence; the image separation network is constructed based on a U-net network, a residual connection scheme, a skip connection scheme and a convolution scheme; the image separation network is configured to separate target features and background features of an input image; The generator network is constructed based on the U-net network, the skip connection scheme, the convolution scheme and a self-attention scheme; The generator network is configured to generate a single-channel fusion image according to the separated target features and background features; the discriminator network is constructed based on a convolution scheme; the discriminator network is configured to determine whether an input image is real ground penetrating radar B-scan data or a single-channel fusion image generated by the generator, and output a corresponding discrimination probability; The model training module is configured to train the constructed ground penetrating radar B-scan data augmentation primary model using the obtained ground penetrating radar B-scan data set, obtain a ground penetrating radar B-scan data augmentation model, and upload the data information to the data augmentation module; The data augmentation module is configured to perform actual data augmentation of the ground penetrating radar B-scan data using the obtained ground penetrating radar B-scan data augmentation model.

Images