GPR data reconstruction method and device based on RA-U-Net network
Through the RA-U-Net network method, the problem of low quality of GPR data reconstruction is solved, efficient reconstruction of missing data in space sampling and direct wave suppression is achieved, and the recognition accuracy of underground target signals and ground penetrating radar are improved. Imaging resolution.
Patent Information
- Application Number
- CN202411050781.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-01
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2044-08-01
AI Technical Summary
The prior art when reconstructing GPR data, the data quality is low, resulting in the lack of scattered echo spatial sampling of underground targets, affecting imaging quality and subsequent analysis.
Using the data reconstruction method based on the RA-U-Net network, the target B-scan data is obtained in the first media scene, the background B-scan data is obtained in the second media scene after the pipeline target is removed, and the training input B is input. -scan data and tag B-scan data iteratively trains the RA-U-Net network to obtain the reconstructed B-scan data with direct waves removed.
The data recovery capability and direct wave suppression capability of GPR data in spatial sampling missing scenarios is improved, and the quality of reconstructed data and the recognition accuracy of underground target signals are improved.
Smart Images

Figure CN119089196A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a data processing method, and in particular to a GPR data reconstruction method and device based on a RA-U-Net network. Background Art
[0002] Ground Penetrating Radar (GPR) is a non-destructive testing technology that emits high-frequency electromagnetic waves for shallow underground detection. It obtains information about underground structures by collecting scattered signals from underground areas. The transmitting antenna of the ground penetrating radar emits electromagnetic waves into the ground at each scanning point on the scanning line. When the electromagnetic wave signal encounters an abnormal body with different electromagnetic parameters from the background at the interface of different media, it will be reflected and scattered. Part of the signal will return to the surface and be received by the ground penetrating radar receiving antenna. These received signals can generate one-dimensional (A-scan) signals, two-dimensional (B-scan) profiles or three-dimensional (C-scan) images of underground structures, which are used to detect targets or evaluate the spatial position, shape, size and dielectric properties of underground structures.
[0003] Specifically, an A-scan echo data can be obtained at each position on the survey line, and multiple A-scan echo data obtained at each position on the survey line constitute GPR B-scan data. Furthermore, multiple GPR B-scan data corresponding to multiple one-dimensional survey lines can constitute a GPR C-scan data.
[0004] When collecting data, ground penetrating radar is easily affected by various factors, resulting in incomplete data collection, such as environmental factors such as uneven surface and complex terrain, external factors such as construction conditions and technical failures, and internal factors such as mismatch between sampling interval and sampling speed. The above situations will lead to random or regular data loss, resulting in the scattering echo of underground targets also showing spatial sampling loss, which will lead to a decrease in GPR imaging quality and affect subsequent data interpretation and analysis.
[0005] Current methods for recovering missing data are mainly interpolation and deep learning methods. Researchers use simple linear interpolation methods or simply ignore missing traces, resulting in low quality of reconstructed GPR data, which has a negative impact on the interpretation and analysis of GPR data and the identification of target signals. Summary of the invention
[0006] The technical problem to be solved by the present invention is the low quality of GPR data reconstructed by the existing method. In order to achieve higher quality reconstruction of missing data, the present invention provides a GPR data reconstruction method and device based on the RA-U-Net network.
[0007] The content of the present invention includes:
[0008] In a first aspect, an embodiment of the present invention provides a GPR data reconstruction method based on a RA-U-Net network, comprising:
[0009] Acquiring target B-scan data in a first media scene, and acquiring background B-scan data in a second media scene, wherein the first media scene includes a pipeline target, and the second media scene is a media scene obtained by removing the pipeline target from the first media scene;
[0010] Replacing a portion of the A-scan data contained in the target B-scan data with a zero value to obtain training input B-scan data;
[0011] Subtracting the target B-scan data from the background B-scan data to obtain label B-scan data corresponding to the training input B-scan data;
[0012] Iteratively training the RA-U-Net network using the training input B-scan data and the corresponding label B-scan data to obtain a trained RA-U-Net network, wherein the RA-U-Net network is constructed based on the U-Net network;
[0013] The B-scan data to be reconstructed is input into the trained RA-U-Net network for data reconstruction processing to obtain the reconstructed B-scan data with the direct wave removed.
[0014] Optionally, acquiring target B-scan data in the first media scene, and acquiring background B-scan data in the second media scene, includes:
[0015] Constructing the first media scene including the pipeline target;
[0016] In the first medium scenario, scanning is performed along a one-dimensional survey line to obtain the target B-scan data, the size of the target B-scan data is M×N, N is the number of scanning points on the one-dimensional survey line, M is the number of time sampling points, and M and N are positive integers;
[0017] Removing the pipeline target from the first media scene to obtain the second media scene;
[0018] In the second medium scenario, scanning is performed along the one-dimensional measurement line to obtain the background B-scan data, and the size of the background B-scan data is M×N.
[0019] Optionally, before iteratively training the RA-U-Net network using the training input B-scan data and the corresponding label B-scan data to obtain a trained RA-U-Net network, the method further includes:
[0020] A RA-U-Net network is constructed based on a U-Net network, wherein the U-Net network is an encoder-decoder network architecture, including an encoder, a decoder, and a jump connection module, wherein the encoder is used to gradually downsample the input into a feature map, the decoder is used to gradually restore the image resolution, the jump connection module is used to connect the feature map in the encoder with the corresponding layer of the decoder, and the encoder includes a two-dimensional convolution module;
[0021] Constructing a nested residual module, so that the nested residual module replaces the two-dimensional convolution module in the encoder, and the nested residual module includes a shallow residual block;
[0022] An attention gating module is constructed and introduced into the jump connection module, wherein the attention gating module is used to perform attention weighting on the high-level features output by the nested residual module and the feature map transmitted after upsampling by the decoder.
[0023] Optionally, the shallow residual block includes an input path and a residual path, the input path includes a first convolutional layer and a batch normalization layer in sequence, the residual path includes a second convolutional layer to match the size and number of channels of the output of the input path and the output of the residual path, and the residual path is used to achieve an identity mapping from input to output;
[0024] The output of the shallow residual block satisfies:
[0025] L(x)=F(x)+H(x);
[0026] Among them, x represents the feature of the shallow residual block input, F(x) is the output of the input path, and H(x) is the output of the residual path.
[0027] Optionally, a loss value for iterative training of the RA-U-Net network is determined based on Huber loss, MS-SSIM loss and perceptual loss;
[0028] The Huber loss satisfies:
[0029]
[0030] Among them, y is the true value, f(x) is the predicted value, and δ is the hyperparameter of the Huber loss function, which is used to control the threshold;
[0031] The MS-SSIM loss satisfies:
[0032]
[0033] The original image is the first scale, and the low-pass filtering and double downsampling are applied once to the second scale. After M-1 iterations, it is the Mth scale. In the above formula, M is the total number of scales included in the calculation of multi-scale structural similarity loss, μ f(x) is the mean of the predicted values, μ y is the mean of the true values, σ f(x) is the standard deviation of the predicted value, σ y is the standard deviation of the true value, σ f(x)y is the covariance, c1 and c2 are preset constants, β m and γ m Used to define the relative importance of the two components in each scale;
[0034] The perceptual loss satisfies:
[0035]
[0036] Where N represents the number of feature maps, i=1,2,3,…,N represents the i-th feature map, ||·||2 represents the L2 norm, represents the square of the L2 norm, F(y) is the feature representation of y in the target layer in the pre-trained convolutional neural network, G(f(x)) is the feature representation of f(x) in the target layer, and the target layer is located in the convolutional layer or the fully connected layer in the pre-trained convolutional neural network.
[0037] In a second aspect, an embodiment of the present invention provides a GPR data reconstruction device based on a RA-U-Net network, comprising:
[0038] An acquisition module, used to acquire target B-scan data in a first media scene, and acquire background B-scan data in a second media scene, wherein the first media scene includes a pipeline target, and the second media scene is a media scene obtained by removing the pipeline target from the first media scene;
[0039] A first training data construction module is used to replace part of the A-scan data included in the target B-scan data with zero values to obtain training input B-scan data;
[0040] A second training data construction module, configured to subtract the target B-scan data from the background B-scan data to obtain label B-scan data corresponding to the training input B-scan data;
[0041] A network training module, used for iteratively training the RA-U-Net network using the training input B-scan data and the corresponding label B-scan data to obtain a trained RA-U-Net network, wherein the RA-U-Net network is constructed based on the U-Net network;
[0042] The data reconstruction module is used to input the B-scan data to be reconstructed into the trained RA-U-Net network for data reconstruction processing to obtain the reconstructed B-scan data with the direct wave removed.
[0043] In a third aspect, an embodiment of the present invention provides an electronic device, comprising: a memory, a processor, and a program stored in the memory and executable on the processor; the processor is used to read the program in the memory to implement the steps in the GPR data reconstruction method based on the RA-U-Net network as described in the first aspect.
[0044] In a fourth aspect, an embodiment of the present invention provides a readable storage medium for storing a program, which, when executed by a processor, implements the steps in the GPR data reconstruction method based on the RA-U-Net network as described in the first aspect.
[0045] The beneficial effect of the present invention is that, in an embodiment of the present application, target B-scan data is obtained in a first media scene including a pipeline target, and background B-scan data is obtained in a corresponding second media scene after the pipeline target is removed, and training input B-scan data is obtained by replacing part of the A-scan data included in the target B-scan data with a zero value; label B-scan data corresponding to the training input B-scan data is obtained by subtracting the target B-scan data from the background B-scan data; finally, the RA-U-Net network is iteratively trained using the training input B-scan data and the corresponding label B-scan data to obtain a trained RA-U-Net network, and the RA-U-Net network is constructed based on the U-Net network. The beneficial effect of the present invention is that GPR data with missing spatial sampling parts are input into the RA-U-Net network to obtain reconstructed GPR data with direct waves removed. Through the above method, not only the data reconstruction problem of GPR B-scan in the case of missing spatial sampling can be solved, but also the reconstructed data can be processed to remove direct waves, thereby improving the data recovery capability and direct wave suppression capability of GPR in the scenario of missing spatial sampling. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Attached Figure 1 One of the flow charts of the GPR data reconstruction method based on the RA-U-Net network provided in an embodiment of the present invention;
[0047] Attached Figure 2 A schematic diagram of the process of constructing a data set provided by an embodiment of the present invention;
[0048] Attached Figure 3 A schematic diagram of the structure of a nested residual module provided by an embodiment of the present invention;
[0049] Attached Figure 4 A schematic diagram of the structure of a shallow residual block provided by an embodiment of the present invention;
[0050] Attached Figure 5 A schematic diagram of the structure of an attention gating module provided in an embodiment of the present invention;
[0051] Attached Figure 6 A schematic diagram of the RA-U-Net network training and data reconstruction process provided by an embodiment of the present invention;
[0052] Attached Figure 7 A schematic diagram of the structure of the RA-U-Net network provided in an embodiment of the present invention;
[0053] Attached Figure 8a A schematic diagram of original GPR data of processing results of four GPR B-scan data provided by an embodiment of the present invention;
[0054] Attached Figure 8b A schematic diagram of the results of removing direct waves by using the mean value method for processing four GPR B-scan data provided by an embodiment of the present invention;
[0055] Attached Figure 8c A schematic diagram of missing data generated by spatial random sampling of processing results of four GPR B-scan data provided by an embodiment of the present invention;
[0056] Attached Figure 8d A schematic diagram of RA-U-Net network test output results of processing results of four GPR B-scan data provided by an embodiment of the present invention;
[0057] Attached Figure 8e A schematic diagram of real label data of the processing results of four GPR B-scan data provided by an embodiment of the present invention;
[0058] Attached Fig. 9 A schematic diagram of a GPR data reconstruction device based on a RA-U-Net network provided in an embodiment of the present invention;
[0059] Attached Fig.10 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0060] In the embodiments of the present application, the term "and / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. In the embodiments of the present application, the term "multiple" refers to two or more, and other quantifiers are similar. The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here, and the objects distinguished by "first" and "second" are generally of the same type, and the number of objects is not limited. For example, the first object can be one or more.
[0061] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0062] See also Figure 1 , Figure 1 A flow chart of a GPR data reconstruction method based on a RA-U-Net network provided in an embodiment of the present invention is provided. The method specifically comprises the following steps:
[0063] Step 101, acquiring target B-scan data in a first media scene and acquiring background B-scan data in a second media scene, wherein the first media scene includes a pipeline target and the second media scene is a media scene obtained by removing the pipeline target from the first media scene.
[0064] Step 102: Part of the A-scan data included in the target B-scan data is replaced with zero values to obtain training input B-scan data.
[0065] Step 103: subtract the target B-scan data from the background B-scan data to obtain label B-scan data corresponding to the training input B-scan data.
[0066] Step 104, iteratively train the RA-U-Net network using the training input B-scan data and the corresponding label B-scan data to obtain a trained RA-U-Net network, where the RA-U-Net network is constructed based on the U-Net network.
[0067] Step 105, input the B-scan data to be reconstructed into the trained RA-U-Net network for data reconstruction processing to obtain the reconstructed B-scan data with the direct wave removed.
[0068] See also Figure 2 , a uniform or random media scene with a horizontal or undulating interface is pre-constructed, in which a pipeline target is buried, and the obtained media scene is the first media scene (also called the target media scene). The media scene obtained after removing the pipeline target from the first media scene is the second media scene (also called the background media scene).
[0069] Optionally, in some embodiments, step 101 includes:
[0070] Constructing the first media scene including the pipeline target;
[0071] In the first medium scenario, scanning is performed along a one-dimensional survey line to obtain the target B-scan data, the size of the target B-scan data is M×N, N is the number of scanning points on the one-dimensional survey line, M is the number of time sampling points, and M and N are positive integers;
[0072] Removing the pipeline target from the first media scene to obtain the second media scene;
[0073] In the second medium scenario, scanning is performed along the one-dimensional measurement line to obtain the background B-scan data, and the size of the background B-scan data is M×N.
[0074] Specifically, simulation software (such as gprMax) is used to scan the first media scene along a one-dimensional survey line. The scanning survey line is perpendicular to the axial direction of the underground pipeline, and the survey line includes N scanning points. At each scanning point, the transmitting antenna transmits electromagnetic waves to the underground, and the receiving antenna receives the scattered echo, and obtains an A-scan echo data with M time sampling points. N scanning points obtain a total of N A-scan echo data, and each A-scan echo data has M time sampling points, forming target B-scan data of size M×N. Repeat the above steps in the second media scene to obtain background B-scan data of size M×N.
[0075] The target B-scan data is randomly extracted for A-scan data according to a preset percentage (denoted as P). During the extraction process, the spatial order and size of the target B-scan data are kept unchanged. After extraction, the A-scan data of these measurement points are replaced with zero values to obtain B-scan data with a spatial sampling missing rate of P, which can be used as input data for RA-U-Net network training later. Among them, P can be set and adjusted according to actual needs, and the specifics are not limited here.
[0076] The ground penetrating radar transmits electromagnetic wave signals to the surface, and the direct coupling wave between the transmitting and receiving antennas and the reflected wave from the surface form a direct wave. The target B-scan data and the background B-scan data have the same background information. The two are subtracted to obtain the B-scan data without the direct wave, which can be used as the real label data for RA-U-Net network training later.
[0077] When the electromagnetic wave signal emitted by the ground penetrating radar passes through the ground, part of it is reflected by the ground. The surface reflected wave and the direct coupling wave of the transceiver antenna form a strong direct wave signal, which will interfere with the detection of the target scattered signal. The suppression of the direct wave interference signal is of great significance to improving the detection accuracy of underground target signals and the imaging resolution of the ground penetrating radar. Due to the unevenness of the ground surface, the direct wave signal reflected by the electromagnetic wave from the ground surface has non-uniform characteristics, making it difficult to suppress the direct wave. Through the above method, the GPRB-scan data without the direct wave can be obtained to achieve the suppression of the direct wave.
[0078] In the specific implementation, Y groups of GPR sample data are collected according to the above method, and each group of GPR sample data includes a training input B-scan data and its corresponding label B-scan data. According to actual needs, all B-scan data in the Y groups of GPR sample data are resized to 256×256 to obtain a sample data set for iterative training of the RA-U-Net network.
[0079] It should be noted that multiple GPR B-scan data corresponding to multiple one-dimensional survey lines can form a GPR C-scan data. The processing of GPR C-scan data is achieved by processing the B-scan data one by one and then combining them into C-scan. The details are not repeated here.
[0080] Optionally, in some embodiments, before step 104, the method further includes:
[0081] A RA-U-Net network is constructed based on a U-Net network, wherein the U-Net network is an encoder-decoder network architecture, including an encoder, a decoder, and a jump connection module, wherein the encoder is used to gradually downsample the input into a feature map, the decoder is used to gradually restore the image resolution, the jump connection module is used to connect the feature map in the encoder with the corresponding layer of the decoder, and the encoder includes a two-dimensional convolution module;
[0082] Constructing a nested residual module, so that the nested residual module replaces the two-dimensional convolution module in the encoder, and the nested residual module includes a shallow residual block;
[0083] An attention gating module is constructed and introduced into the jump connection module, wherein the attention gating module is used to perform attention weighting on the high-level features output by the nested residual module and the feature map transmitted after upsampling by the decoder.
[0084] The RA-U-Net network is constructed based on the U-Net network. The U-Net network consists of a symmetrical encoder and decoder. The encoder gradually downsamples the input to a feature map, and the decoder is used to gradually restore the image resolution. The jump connection connects the feature map in the encoder to the corresponding level of the decoder. A nested residual module is designed to replace the traditional two-dimensional convolution module in the encoder. The nested residual module is mainly composed of shallow residual blocks. An attention mechanism is added to the network, and an attention gating module is constructed and introduced into the jump connection of the network. The high-level features output by the encoder residual module and the feature map transmitted after upsampling by the decoder are weighted for attention, and then feature fusion is performed.
[0085] The nested residual module is mainly composed of shallow residual blocks. Each convolutional layer of the nested residual module adds a skip connection, which is a nested combination of multiple shallow residual blocks. Figure 3 and Figure 4 , Figure 4 It is a schematic diagram of the structure of the shallow residual block (ResidualBlock, ResB), that is, Figure 3 The ResB module shown in Figure 4 There are two branches in it. The left branch is a skip connection, which contains a 1×1 convolution to match the size and number of channels of the right branch output; the right branch contains a two-dimensional convolution.
[0086] like Figure 3 As shown in the figure, the nested residual module processes the input features in two paths. Path 1 contains a shallow residual block; Path 2 includes a shallow residual block, a convolution layer, and a batch normalization layer (Batch Normalization, BN) in sequence. The convolution kernel size of the convolution layer is 3×3, the step size is 1, and the padding is 1. The results of the two paths are added and processed by the ReLU activation function before output. Through the above settings, the neural network can learn the residual information of each layer instead of completely relearning the original function, making it easier to train deep networks.
[0087] Optionally, in some embodiments, the shallow residual block includes an input path and a residual path, the input path includes a first convolutional layer and a batch normalization layer in sequence, the residual path includes a second convolutional layer to match the size and number of channels of the output of the input path and the output of the residual path, and the residual path is used to achieve an identity mapping from input to output.
[0088] The output of the shallow residual block satisfies:
[0089] L(x)=F(x)+H(x);
[0090] Among them, x represents the feature of the shallow residual block input, F(x) is the output of the input path, and H(x) is the output of the residual path.
[0091] Specifically, Figure 4 As shown in the figure, the shallow residual block consists of two main parts: the input path and the residual path. The input path is the traditional neural network structure, which includes the convolution layer (i.e., the first convolution layer) and the batch normalization layer in sequence. The convolution kernel size of the first convolution layer is 3×3, the step size is 1, and the padding is 1. The residual path is an identity mapping from input to output, i.e., a jump connection, including a convolution layer of size 1×1 (i.e., the second convolution layer) to match the size and number of channels of the input path output and the residual path output. The sum of the results of the input path and the residual path, and then the ReLU activation function, is the output of the shallow residual block.
[0092] In this embodiment, a nested residual module is designed to replace the traditional two-dimensional convolution module in the U-Net network encoder, which can make full use of the residual information of each layer, alleviate the gradient vanishing problem in deep network training, enhance the feature extraction capability and accelerate the efficiency of network convergence.
[0093] Construct an attention gate (AG) module, such as Figure 5 As shown, the attention gate input feature signal and the gate signal are the high-level features x output by the residual block of the encoder layer l. l The feature map g transmitted after upsampling by the decoder is weighted by attention and then feature fusion is performed. The attention coefficient α is calculated l , the process can be described as:
[0094]
[0095] Among them, the attention coefficient α i ∈[0,1], g i The feature vectors corresponding to each pixel value i of the two feature maps respectively; W x and W g Represents the feature x l and g, ψ is the transformation coefficient of the combination of two features; W g , W x , ψ both include convolution and batch normalization operations, the convolution kernel size is 1×1, the step size is 1, the padding is 0, and b g and b ψ is the bias term corresponding to the convolution; Θ att represents ψ, W x , Wg 、b g 、b ψ The above series of parameters; i represents the spatial dimension, c represents the channel dimension, σ1 is the ReLU activation function, and σ2 is the Sigmoid function.
[0096] The output of the attention gating module is the attention coefficient and the input high-level feature x l Element-wise multiplication of :
[0097]
[0098] By constructing an attention gating module and introducing it into the jump connection of the network, the interaction between the data-missing area and other related parts can be strengthened, more detailed information can be reconstructed during the signal recovery process, and the continuity and accuracy of signal recovery can be enhanced.
[0099] In this embodiment, the design of the nested residual module can alleviate the gradient vanishing problem in deep network training, enhance the feature extraction capability and accelerate the efficiency of network convergence; the design of the attention module can strengthen the interaction between the data missing area and other related parts, and enhance the continuity and accuracy of signal recovery.
[0100] Finally, the network parameters of the RA-U-Net network are initialized, and the RA-U-Net network is iteratively trained using the training input B-scan data and the corresponding label B-scan data to obtain a trained RA-U-Net network.
[0101] In step 104, the specific steps of training the RA-U-Net network using the above-obtained data set are as follows:
[0102] S1: Initialize the training parameters of the RA-U-Net network and use the dataset obtained in the above method for training.
[0103] S2: Batch training is adopted for the RA-U-Net network, and the batch size is set to p, that is, each batch trains p GPR sample data with a spatial sampling missing rate of P, and each training round performs k batch training. is a round-up function, Y, k, and p are all positive integers. At the beginning of each training round, the order of the training set samples is disrupted so that the batch GPR sample data is randomly selected to improve the effect of network training.
[0104] S3: Use a set of B-scan data with a spatial sampling missing rate of P that has not been trained in the k sets of training data in S2 as the input for RA-U-Net network training, and remove the B-scan data of the direct wave as the true label data to train the RA-U-Net network.
[0105] S4: Repeat steps S2 to S3 until one training round is completed.
[0106] S5: Repeat steps S2 to S4 until the loss error of the RA-U-Net network tends to be stable, and the training of the RA-U-Net network is completed. Specifically, in the network training, let L n is the loss error of the nth training round, L n-1 is the loss error of the n-1th training round, L n-2 is the loss error of the n-2th training round, L n-3 is the loss error of the n-3th training round, where n ≥ 4, when all i (i = 1, 2, 3) satisfy |L n -L n-i When |≤ε, ε is a constant with a very small relative loss magnitude, indicating that the model has reached stability after the nth training.
[0107] Optionally, in some embodiments, the loss value for iterative training of the RA-U-Net network is determined based on Huber loss, multi-scale structural similarity (MS-SSIM) loss, and perceptual loss;
[0108] The Huber loss satisfies:
[0109]
[0110] Among them, y is the true value, f(x) is the predicted value, and δ is the hyperparameter of the Huber loss function, which is used to control the threshold. The mean square error is used within the threshold, and the absolute error is used outside the threshold, as follows:
[0111]
[0112] The MS-SSIM loss satisfies:
[0113]
[0114] In the above formula, M is the total number of scales included in calculating the multi-scale structural similarity loss, μ f(x) is the mean of the predicted values, μ y is the mean of the true values, σ f(x0 is the standard deviation of the predicted value, σ y is the standard deviation of the true value, σ f(x)y is the covariance, c1 and c2 are preset constants, β m and γ mIt is used to define the relative importance of the two components in each scale. In the specific implementation, in order to avoid the instability of division by zero, two constants c1=0.01 are given 2 and c2 = 0.03 2 .
[0115] It should be understood that the MS-SSIM loss calculates the weighted sum of structural similarity losses at multiple scales, which can be achieved by blurring or sampling the image. The original image is the first scale, low-pass filtering and double downsampling are applied once for one iteration to the second scale, and after M-1 iterations to the Mth scale.
[0116] The perceptual loss satisfies:
[0117]
[0118] Where N represents the number of feature maps, i=1,2,3,…,N represents the i-th feature map, ||·||2 represents the L2 norm, represents the square of the L2 norm, F(y) is the feature representation of y in the target layer in the pre-trained convolutional neural network, G(f(x)) is the feature representation of f(x) in the target layer, and the target layer is located in the convolutional layer or the fully connected layer in the pre-trained convolutional neural network.
[0119] In this embodiment, the loss value of iterative training of the RA-U-Net network can be expressed as:
[0120] L total =f(L Huber ,L MS-SSIM ,L perceptual )
[0121] L Huber It is the Huber loss, which is better at handling outliers than the mean square error. The Huber loss function combines the mean square error and the absolute error, using the mean square error when the error is small and the absolute error when the error is large, thereby balancing the sensitivity to outliers.
[0122] L MS-SSIM Denotes the MS-SSIM loss, which computes the weighted sum of structural similarity losses at multiple scales, which can be achieved by blurring or upsampling the image.
[0123] L perceptualRepresents perceptual loss, which measures the difference between images based on perceptual similarity. It captures the texture, structure, and semantic information of the image by extracting feature representations in a neural network, rather than just the difference in pixel values. Typically, perceptual loss uses a pre-trained convolutional neural network (such as VGG) to extract features, and uses the difference between the generated matrix and the true matrix on these features as the loss. This design enables perceptual loss to better guide deep models to generate more realistic data.
[0124] In this embodiment, a comprehensive loss function is designed, including Huber loss, multi-scale structural similarity loss, and perceptual loss. The combination of the three is used as the loss value function for iterative training of the RA-U-Net network, which can guide the network to converge to a valid solution more efficiently and accurately.
[0125] In an embodiment of the present application, target B-scan data is obtained in a first media scene containing a pipeline target, and background B-scan data is obtained in a corresponding second media scene after removing the pipeline target. The training input B-scan data is obtained by replacing part of the A-scan data contained in the target B-scan data with a zero value; the label B-scan data corresponding to the training input B-scan data is obtained by subtracting the target B-scan data from the background B-scan data; finally, the RA-U-Net network is iteratively trained using the training input B-scan data and the corresponding label B-scan data to obtain a trained RA-U-Net network, which is constructed based on the U-Net network. The beneficial effect of the present invention is that the GPR data with missing spatial sampling parts is input into the RA-U-Net network to obtain the reconstructed GPR data with the direct wave removed. Through the above method, not only can the recovery of the GPR signal with missing spatial sampling parts be realized efficiently and accurately, but also the suppression of the direct wave can be realized, thereby improving the recognition accuracy of underground target signals and the imaging resolution of ground penetrating radar.
[0126] By inputting the GPR data with missing spatial sampling into the trained RA-U-Net network, the GPR reconstructed data with suppressed direct waves can be obtained. The processing of GPR C-scan data can be realized by performing the above processing on the multiple B-scan data that make up the GPR C-scan data one by one and then combining them.
[0127] See also Figure 6 , a specific embodiment is used as an example to illustrate below.
[0128] In this embodiment, 570 sets of GPR B-scan simulation data are obtained using the gprMax simulation software. The C-scan profile is a spatial combination of multiple B-scans, so the present invention is not limited to the processing of GPR B-scan data, but is also applicable to the processing of GPRC-scan data. The processing of GPR C-scan data is achieved by processing the multiple B-scan data that make up the C-scan data one by one and then combining them, and the details are not repeated here. The GPR B-scan simulation data set is divided into a training set, a validation set, and a test set according to 7:1:2, that is, 399 sets of data are used to train the RA-U-Net network, and 57 and 114 sets of data are used for network verification and testing.
[0129] First, a random medium scene with an undulating interface is constructed, in which a horizontal cylindrical pipeline target is buried (the first medium scene can be obtained at this time). The gprMax simulation software is used to scan the detection scene along a one-dimensional survey line, and the survey line direction is perpendicular to the axis direction of the underground pipeline. At each position of the survey line, an A-scan echo data is obtained. The multiple A-scan echo data obtained at each position on the survey line constitute a GPR B-scan data containing the pipeline target, that is, the target B-scan data. In the simulation scene, the pipeline target is removed and the background medium scene is kept unchanged. At this time, the second medium scene can be obtained. Repeat the above process in the second medium scene to obtain GPR B-scan data without pipeline targets, that is, background B-scan data. The target B-scan data is randomly extracted by A-scan at 70%. During the extraction process, the spatial order and size of each channel of the target B-scan data are kept unchanged. After extraction, the A-scan data of these measurement points are replaced with zero values to obtain GPR data with a spatial sampling missing rate of 70%, which is the training input B-scan data. The target B-scan data is subtracted from the background B-scan data (the former minus the latter) to obtain the GPR B-scan data with suppressed direct waves, which is the corresponding label B-scan data.
[0130] According to the above method, 570 sets of GPR B-scan data were collected, and all data sizes were resized to 256 × 256 to obtain the data set for network training. Among them, 399 sets of GPR data with a spatial sampling missing rate of 70% and GPR B-scan data with suppressed direct waves were used as network input and real label data sets to train the RA-U-Net network, and 57 sets of simulation data and 114 sets of simulation data were used for network verification and testing.
[0131] Construct the RA-U-Net network. Based on the U-Net encoder-decoder network architecture, design a nested residual module to replace the traditional two-dimensional convolutional module in the encoder. The structure of the nested residual module is as follows: Figure 3 and Figure 4 As shown in Figure 1. Construct an attention gating module and introduce it into the skip connection of the network. Perform attention weighting on the high-level features output by the encoder residual block and the feature map transmitted after upsampling by the decoder, and then perform feature fusion. The structure of the attention gating is shown in Figure 1. Figure 5 As shown. Design a combined loss function, including Huber loss, multi-scale structural similarity loss, and perceptual loss, and use their linear combination as the network training loss function. This embodiment uses a linear combination of the three, but is not limited to a linear combination, and can also be a nonlinear combination. The RA-U-Net network structure is shown in Figure 7 shown.
[0132] The RA-U-Net network was trained using 399 sets of simulation data. The training rounds were set to 30 rounds, the initial learning rate was set to 0.01, the batch training size was 4, and 4 samples were trained in one batch training. The network training loss function was a linear combination of Huber loss, MS-SSIM loss, and perceptual loss:
[0133]
[0134] Weights of the three loss functions The perceptual losses are 0.5, 0.3, and 0.2 respectively, and the features are extracted from the four layers of “conv1_1”, “conv2_1”, “conv3_1”, and “conv4_1” of the pre-trained VGG19 network.
[0135] During the training process of the RA-U-Net network, each batch training inputs 4 groups of GPR data with a spatial sampling missing rate of 70% and the corresponding real label data into the network training, outputs the reconstructed GPR data with suppressed direct waves, and calculates the loss function value based on the output data compared with the real label data, and accumulates it to the loss value of the previous batch training. In one training round, all samples are trained according to the batch size, and the parameters of the network are updated and optimized according to the loss function value, and then the network training of the next training round is carried out. After 30 training rounds, the loss function value tends to be stable, the network training is completed, and the model and parameter values of the network training are saved.
[0136] 114 sets of GPR data with a spatial sampling missing rate of 70% were input into the trained RA-U-Net network to test the training effect of the network. The network trained with the single cylindrical pipeline target GPR simulation data was tested on the double cylindrical pipeline target GPR simulation data. Figure 8a-8e The figure includes the processing results of four GPR B-scan data, wherein the figures in the first column and the second column are the simulation data results of the single cylindrical pipeline target, and the figures in the third column and the fourth column are the simulation data results of the double cylindrical pipeline target. Figure 8a The original GPR data is displayed. The direct wave signal in the image is strong and uneven, while the target signal is relatively weak, which will seriously affect the detection and identification of the target signal; Figure 8b The traditional mean value method is used to suppress the direct wave. However, due to the non-uniform direct wave signal caused by the undulating surface, the direct wave signal cannot be effectively suppressed. Figure 8c It is the GPR data with a spatial sampling missing rate of 70%; Figure 8d This is the output result of the RA-U-Net network test. It not only recovers the target signal, but also suppresses the direct wave interference signal. Figure 8e The network test results have achieved good visual restoration results based on the real labeled data. Figure 8a-8e From the test results of dual-target data in the third column of the figure, it can be seen that the network has a good recovery effect on the target hyperbolic signal, but other details, such as the mutual interference wave signals between targets, are not recovered.
[0137] Table 1 gives Figure 8a-8e The accuracy results of the four data sets and the average accuracy results of the 114 data sets in the test set are shown in the table. The mean square error (MSE), peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are calculated by comparing the model test results with the real label data. The results in the table show that even if the spatial sampling missing rate of the GPR data is large (70%), the single-target GPR data (samples 1 and 2) of the test set still achieve high accuracy in the three indicators; there are no dual-target GPR data samples in the training set, but the dual-target GPR data (samples 3 and 4) tested also achieved high accuracy values, indicating that the reconstruction effect of the network is very good; the three average indicator values of the 114 test data sets also achieved high accuracy.
[0138] Table 1 Test results Accuracy results
[0139]
[0140] The GPR missing data reconstruction and direct wave suppression method based on the RA-U-Net network implemented by the present invention can not only reconstruct the GPR data with missing spatial sampling parts with high precision, but also suppress the direct wave signals on the horizontal ground or undulating ground during network training, completing the dual-target tasks of GPR missing signal recovery and direct wave suppression, and facilitating the extraction of underground target echo features. The present invention effectively achieves the dual-target tasks of missing signal recovery and direct wave signal suppression, and well solves the problem that direct waves cannot be effectively suppressed when the surface is uneven.
[0141] like Fig. 9 As shown, the embodiment of the present invention further provides a GPR data reconstruction device 900 based on the RA-U-Net network, comprising:
[0142] An acquisition module 901 is used to acquire target B-scan data in a first media scene and acquire background B-scan data in a second media scene, wherein the first media scene includes a pipeline target and the second media scene is a media scene obtained by removing the pipeline target from the first media scene;
[0143] A first training data construction module 902 is used to replace a portion of the A-scan data included in the target B-scan data with a zero value to obtain training input B-scan data;
[0144] A second training data construction module 903, configured to obtain label B-scan data corresponding to the training input B-scan data by subtracting the target B-scan data from the background B-scan data;
[0145] A network training module 904 is used to iteratively train the RA-U-Net network using the training input B-scan data and the corresponding label B-scan data to obtain a trained RA-U-Net network, wherein the RA-U-Net network is constructed based on the U-Net network;
[0146] The data reconstruction module 905 is used to input the B-scan data to be reconstructed into the trained RA-U-Net network for data reconstruction processing to obtain the reconstructed B-scan data with the direct wave removed.
[0147] Optionally, the acquisition module 901 includes:
[0148] A construction unit, configured to construct the first media scene including the pipeline target;
[0149] A first scanning unit is used to scan along a one-dimensional survey line in the first medium scene to obtain the target B-scan data, wherein the size of the target B-scan data is M×N, where N is the number of scanning points on the one-dimensional survey line, M is the number of time sampling points, and M and N are positive integers;
[0150] A processing unit, configured to remove the pipeline target in the first media scene to obtain the second media scene;
[0151] The second scanning unit is used to scan along the one-dimensional measurement line in the second medium scene to obtain the background B-scan data, where the size of the background B-scan data is M×N.
[0152] Optionally, the RA-U-Net network-based GPR data reconstruction device 900 further includes:
[0153] A first construction module is used to construct a RA-U-Net network based on a U-Net network, wherein the U-Net network is an encoder-decoder network architecture, including an encoder, a decoder, and a jump connection module, wherein the encoder is used to gradually downsample the input into a feature map, the decoder is used to gradually restore the image resolution, the jump connection module is used to connect the feature map in the encoder with the corresponding layer of the decoder, and the encoder includes a two-dimensional convolution module;
[0154] A second construction module is used to construct a nested residual module, so that the nested residual module replaces the two-dimensional convolution module in the encoder, and the nested residual module includes a shallow residual block;
[0155] The third building module is used to construct an attention gating module and introduce the attention gating module into the jump connection module, wherein the attention gating module is used to perform attention weighting on the high-level features output by the nested residual module and the feature map transmitted after upsampling by the decoder.
[0156] Optionally, the shallow residual block includes an input path and a residual path, the input path includes a first convolutional layer and a batch normalization layer in sequence, the residual path includes a second convolutional layer to match the size and number of channels of the output of the input path and the output of the residual path, and the residual path is used to achieve an identity mapping from input to output;
[0157] The output of the shallow residual block satisfies:
[0158] L(x)=F(x)+H(x);
[0159] Among them, x represents the feature of the shallow residual block input, F(x) is the output of the input path, and H(x) is the output of the residual path.
[0160] Optionally, a loss value for iterative training of the RA-U-Net network is determined based on Huber loss, MS-SSIM loss and perceptual loss;
[0161] The Huber loss satisfies:
[0162]
[0163] Among them, y is the true value, f(x) is the predicted value, and δ is the hyperparameter of the Huber loss function, which is used to control the threshold;
[0164] The MS-SSIM loss satisfies:
[0165]
[0166] The original image is the first scale, and the low-pass filtering and double downsampling are applied once to the second scale. After M-1 iterations, it is the Mth scale. In the above formula, M is the total number of scales included in the calculation of multi-scale structural similarity loss, μ f(x) is the mean of the predicted values, μ y is the mean of the true values, σ f(x) is the standard deviation of the predicted value, σ y is the standard deviation of the true value, σ f(x)y is the covariance, c1 and c2 are preset constants, β m and γ m Used to define the relative importance of the two components in each scale;
[0167] The perceptual loss satisfies:
[0168]
[0169] Where N represents the number of feature maps, i=1,2,3,…,N represents the i-th feature map, ||·||2 represents the L2 norm, represents the square of the L2 norm, F(y) is the feature representation of y in the target layer in the pre-trained convolutional neural network, G(f(x)) is the feature representation of f(x) in the target layer, and the target layer is located in the convolutional layer or the fully connected layer in the pre-trained convolutional neural network.
[0170] The RA-U-Net network-based GPR data reconstruction device 900 provided in the embodiment of the present application can execute the above method embodiment, and its implementation principle and technical effect are similar, which will not be repeated in this embodiment.
[0171] It should be noted that the division of units in the embodiments of the present application is schematic and is only a logical function division. There may be other division methods in actual implementation. In addition, each functional unit in each embodiment of the present application may be integrated into a processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.
[0172] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a processor-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) or a processor (processor) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (RandomAccess Memory, RAM), disk or optical disk and other media that can store program code.
[0173] like Fig.10 As shown, an embodiment of the present application provides an electronic device 1000, including: a memory 1002, a processor 1001, and a program stored in the memory 1002 and executable on the processor 1001; the processor 1001 is used to read the program in the memory 1002 to implement the steps in the GPR data reconstruction method based on the RA-U-Net network as described above.
[0174] The embodiment of the present application also provides a readable storage medium, on which a program is stored. When the program is executed by a processor, the various processes of the above-mentioned GPR data reconstruction method embodiment based on the RA-U-Net network are implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here. Among them, the readable storage medium can be any available medium or data storage device that can be accessed by the processor, including but not limited to magnetic storage (such as floppy disk, hard disk, magnetic tape, magneto-optical disk (MO), etc.), optical storage (such as compact disk (CD), digital video disk (DVD), Blu-ray disk (BD), high-definition versatile disk (HVD), etc.), and semiconductor memory (such as read-only memory (ROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read only memory (EEPROM), non-volatile memory (NAND FLASH), solid-state drive (Solid State Disk or Solid State Drive, SSD)), etc.
[0175] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.
[0176] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, disk, CD), and includes a number of instructions for a terminal (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present application.
[0177] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.
Claims
1. A GPR data reconstruction method based on RA-U-Net network, characterized in that: include: Acquiring target B-scan data in a first media scene, and acquiring background B-scan data in a second media scene, wherein the first media scene includes a pipeline target, and the second media scene is a media scene obtained by removing the pipeline target from the first media scene; Replacing a portion of the A-scan data contained in the target B-scan data with a zero value to obtain training input B-scan data; Subtracting the target B-scan data from the background B-scan data to obtain label B-scan data corresponding to the training input B-scan data; Iteratively training the RA-U-Net network using the training input B-scan data and the corresponding label B-scan data to obtain a trained RA-U-Net network, wherein the RA-U-Net network is constructed based on the U-Net network; The B-scan data to be reconstructed is input into the trained RA-U-Net network for data reconstruction processing to obtain the reconstructed B-scan data with the direct wave removed.
2. The method according to claim 1, characterized in that: The step of acquiring target B-scan data in the first medium scene and acquiring background B-scan data in the second medium scene includes: Constructing the first media scene including the pipeline target; In the first medium scenario, scanning is performed along a one-dimensional survey line to obtain the target B-scan data, the size of the target B-scan data is M×N, N is the number of scanning points on the one-dimensional survey line, M is the number of time sampling points, and M and N are positive integers; Removing the pipeline target from the first media scene to obtain the second media scene; In the second medium scenario, scanning is performed along the one-dimensional measurement line to obtain the background B-scan data, and the size of the background B-scan data is M×N.
3. The method according to claim 1, characterized in that: Before iteratively training the RA-U-Net network using the training input B-scan data and the corresponding label B-scan data to obtain a trained RA-U-Net network, the method further includes: A RA-U-Net network is constructed based on a U-Net network, wherein the U-Net network is an encoder-decoder network architecture, including an encoder, a decoder, and a jump connection module, wherein the encoder is used to gradually downsample the input into a feature map, the decoder is used to gradually restore the image resolution, the jump connection module is used to connect the feature map in the encoder with the corresponding layer of the decoder, and the encoder includes a two-dimensional convolution module; Constructing a nested residual module, so that the nested residual module replaces the two-dimensional convolution module in the encoder, and the nested residual module includes a shallow residual block; An attention gating module is constructed and introduced into the jump connection module, wherein the attention gating module is used to perform attention weighting on the high-level features output by the nested residual module and the feature map transmitted after upsampling by the decoder.
4. The method according to claim 3, characterized in that: The shallow residual block includes an input path and a residual path, wherein the input path includes a first convolution layer and a batch normalization layer in sequence, and the residual path includes a second convolution layer to match the size and number of channels of the output of the input path and the output of the residual path, and the residual path Used to achieve the identity mapping from input to output; The output of the shallow residual block satisfies: L(x)=F(x)+H)(x); Among them, x represents the feature of the shallow residual block input, F(x) is the output of the input path, and H(x) is the output of the residual path.
5. The method according to any one of claims 1 to 4, characterized in that The loss value of the iterative training of the RA-U-Net network is determined based on Huber loss, MS-SSIM loss and perceptual loss; The Huber loss satisfies: Among them, y is the true value, f(x) is the predicted value, and δ is the hyperparameter of the Huber loss function, which is used to control the threshold; The MS-SSIM loss satisfies: The original image is the first scale, and the low-pass filtering and double downsampling are applied once to the second scale. After M-1 iterations, it is the Mth scale. In the above formula, M is the total number of scales included in the calculation of multi-scale structural similarity loss, μ f(x) is the mean of the predicted values, μ y is the mean of the true values, σ f(x) is the standard deviation of the predicted value, σ y is the standard deviation of the true value, σ f(x)y is the covariance, c1 and c2 are preset constants, β m and γ m Used to define the relative importance of the two components in each scale; The perceptual loss satisfies: Where N represents the number of feature maps, i=1,2,3,…,N represents the i-th feature map, ||·||2 represents the L2 norm, represents the square of the L2 norm, F(y) is the feature representation of y in the target layer in the pre-trained convolutional neural network, G(f(x)) is the feature representation of f(x) in the target layer, and the target layer is located in the convolutional layer or the fully connected layer in the pre-trained convolutional neural network.
6. A GPR data reconstruction device based on RA-U-Net network, characterized in that: include: An acquisition module, used to acquire target B-scan data in a first media scene, and acquire background B-scan data in a second media scene, wherein the first media scene includes a pipeline target, and the second media scene is a media scene obtained by removing the pipeline target from the first media scene; A first training data construction module is used to replace part of the A-scan data included in the target B-scan data with zero values to obtain training input B-scan data; A second training data construction module, configured to subtract the target B-scan data from the background B-scan data to obtain label B-scan data corresponding to the training input B-scan data; A network training module, used for iteratively training the RA-U-Net network using the training input B-scan data and the corresponding label B-scan data to obtain a trained RA-U-Net network, wherein the RA-U-Net network is constructed based on the U-Net network; The data reconstruction module is used to input the B-scan data to be reconstructed into the trained RA-U-Net network for data reconstruction processing to obtain the reconstructed B-scan data with the direct wave removed.
7. An electronic device comprising: A memory, a processor, and a program stored in the memory and executable on the processor; wherein the processor is used to read the program in the memory to implement the steps in the GPR data reconstruction method based on the RA-U-Net network as described in any one of claims 1 to 5.
8. A readable storage medium for storing a program, characterized in that: When the program is executed by a processor, the steps in the GPR data reconstruction method based on the RA-U-Net network as described in any one of claims 1 to 5 are implemented.
Citation Information
Patent Citations
Fetal heart and fetal lung deep learning combined segmentation method based on U-Net ultrasound
CN111242956A
Medical image segmentation method based on CHS-Net network
CN113096071A
Ground penetrating radar B-scan image feature enhancement method and system based on deep learning
CN114331890A
Unsupervised seismic data reconstruction method and device based on sparse constraint
CN114545494A
Ground penetrating radar back projection imaging method and system
CN114966560A
