Ground penetrating radar hyperbolic target detection method based on state space model
Through the hyperbolic target detection method of ground penetrating radar based on state space model, the arbitrary line length data in the B-scan data of the ground penetrating radar is processed, and the problem of reduced detection accuracy in traditional methods is solved, achieving efficient and accurate hyperbolic target detection.
Patent Information
- Application Number
- CN202510189466.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-20
- Publication Date
- 2025-06-27
AI Technical Summary
The existing deep learning-based ground penetrating radar hyperbolic recognition method cannot process ground penetrating radar B-scan data of arbitrary line lengths, resulting in data size reconstruction in data preprocessing may lead to distortion of the shape characteristics of the target to be detected, thereby reducing detection accuracy.
The hyperbolic target detection method of ground penetrating radar based on state space model is used to model the ground penetrating radar B-scan data into sequences, and the data is featured through the residual neural network and the deep neural network of the state space model. The time-dimensional decoder of the transposed convolution operation is upsampled and decoded, and the detection result image is output and post-processed to obtain the final hyperbolic target detection and positioning results.
This method can identify and detect hyperbolic targets in B-scan data of any line length, improve the accuracy and efficiency of hyperbolic target detection of ground penetrating radar, and solve the problem of reducing detection accuracy of traditional methods when processing arbitrary line length data.
Smart Images
Figure CN120214724A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of ground penetrating radar data processing, and particularly relates to a method for detecting hyperbola targets of ground penetrating radar based on a state space model. Background Art
[0002] The ground penetrating radar (GPR) technology is an electromagnetic geophysical method for effectively detecting underground targets based on the discontinuity of underground media. Due to its characteristics of non-destructive, high precision, high efficiency, etc., it has been widely used in the fields of civil engineering, archaeology, mineral exploration, military target detection, etc. During the detection process of ground penetrating radar, the instrument moves along the measurement line on the surface of the detected area at a specified speed, and time series data is synchronously recorded, and finally B-scan data of the entire measurement line is formed. In the B-scan data, the main target feature is the hyperbola morphological feature. Therefore, identifying the hyperbola in the ground penetrating radar B-scan data can realize the detection, positioning and interpretation of the targets in the detected area. However, the traditional empirical-based ground penetrating radar B-scan data interpretation technology is time-consuming and laborious. Therefore, the automatic interpretation method for ground penetrating radar data for hyperbola feature recognition is the mainstream direction.
[0003] In recent years, since the method based on deep learning can directly learn the feature representation of the detection target from the ground penetrating radar B-scan data, this type of method has become the mainstream method in the automatic interpretation method for ground penetrating radar hyperbola recognition. The deep learning methods that have been applied to the automatic interpretation of ground penetrating radar hyperbola recognition include Faster R-CNN, Mask R-CNN, SSD, and the YOLO series. However, the above deep learning methods require the input of an image with a fixed size in practical applications and cannot process the ground penetrating radar B-scan data with an arbitrary measurement line length. When directly using the ground penetrating radar B-scan data with a long measurement line for hyperbola target detection, the data size reconstruction in data preprocessing may cause the distortion of the shape features of the target to be detected, thereby reducing the detection accuracy. Summary of the Invention
[0004] In order to solve the problem that the existing ground penetrating radar hyperbola recognition method based on deep learning cannot process the ground penetrating radar B-scan data with an arbitrary measurement line length, the present invention provides a method for detecting hyperbola targets of ground penetrating radar based on a state space model. This method detects hyperbola targets in B-scan data with an arbitrary measurement line length by modeling the ground penetrating radar B-scan data as a sequence. This method can identify and detect hyperbola targets in B-scan data with an arbitrary measurement line length, and is an effective method for detecting hyperbola targets of ground penetrating radar.
[0005] The technical solution for realizing the present invention is as follows:
[0006] A ground penetrating radar hyperbola target detection method based on a state space model, and the specific process is as follows:
[0007] Step 1, use the time dimension encoder of the residual neural network (ResNet) to perform feature encoding on the time dimension signals of each scan line of the ground penetrating radar B-scan data, and obtain a feature data matrix containing high-dimensional features of the time dimension signals;
[0008] Step 2, use the deep neural network of the state space model (State Space Model, SSM), and use the state space model to extract bidirectional correlation features from the feature data matrix along the survey line (distance) dimension, and obtain a feature data matrix containing hyperbola target features in the survey line dimension;
[0009] Step 3, use the time dimension decoder of the transposed convolution (ConvTranspose) operation to perform upsampling decoding on the feature data matrix obtained in Step 2 and output a detection result image; the final hyperbola target detection and positioning result is obtained by post-processing the detection result image.
[0010] Optionally, the state space model of the present invention is:
[0011]
[0012] y(n) = Ch(n) + Dx(n)
[0013] where x(n), h(n), and y(n) are the discrete input sequence, system hidden state, and output sequence respectively, matrices A, B, C, and D are learnable linear layers, Δt is the discrete step size, represents the Hadamard product operation of matrices.
[0014] Optionally, the discrete step size of the present invention is:
[0015] Δt = Sig(Eh(n - 1) + Fx(n))
[0016] where E and F are learnable linear layers, and Sig() is the Sigmoid activation function.
[0017] Optionally, when the entire network model composed of the encoder, deep neural network, and decoder of the present invention is jointly trained, the training parameters of the state space model include θ S = {A, B, C, D, E, F}, and the state space model uses two loop mechanisms to calculate the input features as an output containing hyperbola features where the first loop processes the feature matrix from n = 1 to N scans Obtain the hidden state matrix H, and the second loop processes the feature matrix by scanning it in the opposite direction Obtain the hidden feature matrix H b ; Use the hidden state matrix H and the hidden feature matrix H b Calculate the output sequence.
[0018] Optionally, in the present invention, the first loop processes the feature matrix by scanning from n = 1 to N Obtain the hidden state matrix H, specifically:[[]]
[0019]
[0020] h(n) = tanh(Δh(n) + h(n - 1))
[0021] where h(n) is an element of the hidden state matrix H.
[0022] Optionally, in the present invention, the second loop processes the feature matrix by scanning it in the opposite direction Obtain the hidden feature matrix H b , specifically:[[]]
[0023]
[0024] h b (n) = tanh(Δh b (n) + h b (n - 1))
[0025] where, in the second loop, the expression of h b (n) is the same as that of h(n), and h b (n) is an element of the hidden feature matrix H b .
[0026] Optionally, in the present invention, the calculation of the output sequence using the hidden state matrix H and the hidden feature matrix H b is as follows:[[]]
[0027]
[0028] Optionally, the encoder in the present invention is composed of a two-dimensional convolutional layer that performs convolution along the time dimension of the B-scan data, two feature extraction modules composed of a residual neural network module and a max pooling layer, and a module composed of a flattening layer and a linear layer.
[0029] Optionally, the encoder in the present invention is composed of a linear layer, three decoding modules composed of a residual neural network module and a transposed convolutional layer, and a Sigmoid activation function layer.
[0030] Optionally, the residual neural network module in the decoder and encoder of the present invention is composed of two two-dimensional convolutional layers and a residual connection.
[0031] Beneficial effects:
[0032] The ground penetrating radar hyperbola target detection method based on the state space model proposed by the present invention, compared with the existing technology, can identify and detect hyperbola targets in B-scan data with any survey line length, and is an effective ground penetrating radar hyperbola target detection method. Brief description of the drawings
[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.
[0034] Figure 1 It is the processing flow chart of the method of the present invention;
[0035] Figure 2 It is the structural flow chart of the time dimension encoder based on the residual neural network (ResNet);
[0036] Figure 3 It is the block diagram of the continuous state space model and the discrete state space model;
[0037] Figure 4 It is the structural flow chart of the deep neural network based on the state space model;
[0038] Figure 5 It is the structural flow chart of the time dimension decoder based on the transposed convolution (ConvTranspose);
[0039] Figure 6 It is an example of simulation and measured data and their label images;
[0040] Figure 7 It is the training loss function curve;
[0041] Figure 8 It is the experimental result of the simulation data;
[0042] Figure 9 It is the experimental result of the measured data. Specific implementation manners
[0043] The following will make a detailed description of the implementation manner of the method of the present invention with reference to the drawings.
[0044] An embodiment of the present application provides a ground penetrating radar hyperbola target detection method based on the state space model, and the processing flow is asFigure 1 As shown in the figure, its specific components include:
[0045] Step 1: Use the time - dimension encoder of the Residual Neural Network (ResNet) to perform feature encoding on the time - dimension signals of each scan line of the ground - penetrating radar B - scan data, and obtain a feature data matrix containing the high - dimensional features of the time - dimension signals.
[0046] As Figure 2 shown, in this embodiment, the encoder consists of a two - dimensional convolutional layer that performs convolution along the time dimension of the B - scan data, two feature extraction modules composed of residual neural network modules and max - pooling layers, and a module composed of a flattening layer and a linear layer. Among them, each residual neural network module consists of two two - dimensional convolutional layers and a residual connection. The convolutional layer is used to extract the features of the time dimension of the B - scan data, and the residual connection is used to maintain the feature details and improve the gradient propagation. A max - pooling layer is used between each residual module to reduce the spatial dimension of the features. The output of the encoder is a semantic feature map with a smaller spatial size and more channels.
[0047] The parameters of each module of the encoder are shown in Table 1.
[0048] Table 1 Structural parameters of the time - dimension encoder
[0049]
[0050]
[0051] Step 2: Use the deep neural network of the State Space Model (SSM) to extract bidirectional correlation features from the feature data matrix along the survey line (distance) dimension by using the state space model, and obtain a feature data matrix containing hyperbolic target features in the survey line dimension;
[0052] As Figure 4 shown, the state space model can describe a system state representation, which maps the input sequence at the current moment to the hidden state representation and derives the predicted output sequence.
[0053] As Figure 3 (a) shown, the continuous state space model can be described by the following two equations,
[0054]
[0055] Among them, \(x(t)\) is the input sequence at time \(t\), \(h(t)\) is the system hidden state at time \(t\), and \(y(t)\) is the output sequence at time \(t\). The first equation is the state equation, where the state transition matrix \(A\) and the input matrix \(B\) respectively reflect how the current state and input affect the change of the state; the second equation is the output equation, where the output matrix \(C\) and the direct transfer matrix \(D\) respectively describe how the state is converted into the output and how the input affects the output.
[0056] As Figure 3 (b) shows that in order to enable the state space model to process discrete sequences, a discrete time step \(\Delta t\) is introduced to discretize the continuous state space model to obtain a discrete state space model, which is described by the following two equations.
[0057]
[0058] Among them, \(x(n)\), \(h(n)\) and \(y(n)\) are the input sequence, system hidden state and output sequence after discretization respectively. represents the Hadamard product operation of matrices. and are the discrete state transition matrix and the discrete input matrix respectively. In addition, the first equation in Equation (2) can be expressed in the form of a change amount:
[0059]
[0060] Among them, \(\Delta h(n)\) is the change amount of the hidden state. In order to implement the deep learning form of the discrete state space model, the matrices \(A\), \(B\), \(C\) and \(D\) are modeled as learnable linear layers. At the same time, in order to selectively retain the required information, the discrete time step \(\Delta t\) is made to depend on the input and the current state. This content-aware operation is modeled as:
[0061] \(\Delta t=\text{Sig}(Eh(n - 1)+Fx(n))\ (4)\)
[0062] Among them, \(E\) and \(F\) are learnable linear layers, and \(\text{Sig}()\) is the Sigmoid activation function.
[0063]
[0064] Step 3: Use the time-dimensional decoder of the transposed convolution (ConvTranspose) operation to upsample and decode the feature data matrix obtained in Step 2 and output the detected result image; the final hyperbola object detection and localization result is obtained by post-processing the detected result image.
[0065] As Figure 5As shown in the figure, in this embodiment, the encoder is composed of a linear layer, three decoding modules each consisting of a residual neural network module and a transposed convolutional layer, and a Sigmoid activation function layer. Among them, each residual neural network module is composed of two two-dimensional convolutional layers and a residual connection. The convolutional layer is used to decode the features in the time dimension of the input data, and the residual connection is used to maintain the feature details and improve the gradient propagation. Between each residual module, a transposed convolutional layer is used to upsample and decode the feature data matrix and output the detected result image.
[0066] The parameters of each module of the decoder are shown in Table 2
[0067] Table 2 Structural parameters of the time dimension decoder
[0068]
[0069]
[0070] In this embodiment, in order to clearly explain how the state space model is embedded into the entire network model, the training process of the entire network model is given in the algorithm.
[0071] First, sample the data X (j) Input it into the encoder to obtain the output features
[0072] Secondly, through the two cyclic mechanisms of the state space model, the input features Are calculated as the output containing hyperbolic features Among them, the first cycle processes scanning the feature matrix from n = 1 to N To obtain the hidden state matrix H. The second cycle processes scanning the feature matrix in the opposite direction To obtain the hidden feature matrix H b .
[0073] Thirdly, Input it into the decoder to obtain the output of the entire network model
[0074] Finally, calculate the training loss and the network parameters are updated.
[0075]
[0076]
[0077]
[0078] Embodiment
[0079] In order to verify a ground penetrating radar hyperbolic target detection method based on the state space model proposed by the present invention, data verification was carried out.
[0080] First is the preparation of the dataset. The dataset includes 5332 simulated B-scan data and 283 measured B-scan data, a total of 5615 B-scan data. Among them, the simulated data is generated by the simulation software gprMax. In the simulation experiment, 5332 simulated data are obtained by randomly setting the simulation parameters in different ways. The simulation parameters are shown in Table 3. To enhance the generalization of the dataset, measured data is collected under different measured conditions and further screened. The measured experimental conditions are shown in Table 4. Figure 6 Examples of measured and simulated data images and corresponding label images are given.
[0081] Table 3 Simulation conditions of the dataset
[0082]
[0083] Table 4 Measured experimental conditions of the dataset
[0084]
[0085]
[0086] Secondly is network training. The network training parameters are shown in Table 5. The loss curve obtained from network training is as Figure 7 shown, and the training loss approaches convergence from 200 epochs.
[0087] Table 5 Network training parameters
[0088]
[0089] To verify the performance of the network model after training, simulation and measured data test verifications are carried out.
[0090] The simulated data is generated by the simulation software gprMax. The data image of the original data obtained by simulation after clutter suppression preprocessing is as Figure 8 (a) shown. The background medium of both groups of data is soil with a random moisture content between 2% and 25%. The targets are ideal metal cylinders. The depth range is 0.03 - 0.3 m. The transmitted waveform uses the first derivative of a Gaussian waveform with a center frequency of 1.5 GHz and a time window of 10 ns. For the first group of data, the length of the survey line in the simulation scenario is 1.5 m, the antenna step is 1 cm, the number of data scan channels obtained is 150, and the number of targets is 3. For the second group of data, the length of the survey line in the simulation scenario is 3.0 m, the antenna step is 1 cm, the number of data scan channels obtained is 300, and the number of targets is 10. Although the lengths of the survey lines of the two groups of data are different, the method of the present invention can handle them both, and the network output results obtained are as Figure 8(As shown in (b), the network output image data is a probability map, and each pixel value represents the probability that the pixel position is the position of the vertex region of the target hyperbola. From Figure 8 From the upper and lower two figures in (b), it can be seen that the method of the present invention can output a probability map representing the vertex region of the target hyperbola. After binarization and clustering processing, the final detection and positioning results are as Figure 8 (c) shown. It can be seen from the figure that for the two groups of data processed by the method of the present invention, the vertex region of the target hyperbola can be correctly detected and positioned.
[0091] The image of the original data obtained from the actual measurement experiment after clutter suppression preprocessing is as Figure 9 (a) shown. The background medium of the three groups of data is sandy soil, the targets are all metal cylinders, such as steel pipes and steel bars, the time window is 15 ns, and the antenna step is 1 cm. The length of the survey line in the experimental scenario of the first group of data is 1.6 m, the number of scanned channels of the obtained data is 161, the antenna is GSSI 2.6 GHz, and the diameters of the metal cylinder targets from left to right are 0.54 cm, 0.7 cm, 0.94 cm, and 1.13 cm respectively. The length of the survey line in the experimental scenario of the second group of data is 1.4 m, the number of scanned channels of the obtained data is 141, the antenna is LTD 900 MHz, and the target diameters from left to right are 4.8 cm and 2 cm respectively. The length of the survey line in the experimental scenario of the third group of data is 1.5 m, the number of scanned channels of the obtained data is 151, the antenna is LTD 1.5 GHz, and the target diameters from left to right are 2.15 cm and 3.31 cm respectively. Although the lengths of the survey lines of the three groups of data are different, the method of the present invention can process them all, and the network output results are as Figure 9 (b) shown. From Figure 9 From the three figures in (b), it can be seen that the method of the present invention can output a probability map representing the vertex region of the target hyperbola. After binarization and clustering processing, the final detection and positioning results are as Figure 9 (c) shown. It can be seen from the figure that for the three groups of actual measurement data processed by the method of the present invention, although there are different clutter interferences in the input data, the vertex region of the target hyperbola can be correctly detected and positioned, which proves the robustness and effectiveness of the method of the present invention.
[0092] In order to further explore the performance of the method, precision and recall are used as statistical indicators, and the indicators are statistically analyzed for a test data set composed of 52 pieces of simulation data and 29 pieces of actual measurement data. The statistical results are shown in Table 6.
[0093] Table 6 Confusion matrix
[0094]
[0095] Based on the statistical data in Table 6, the precision and recall are further calculated.
[0096]
[0097] It can be seen from the index statistics of the test data set that the method of the present invention not only has the ability to process B-scan data of any survey line length, but also has good target detection performance.
[0098] In summary, the above are only the preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.
Claims
1. A ground penetrating radar hyperbolic target detection method based on a state space model, characterized in that: The specific process is: Step 1: Using the time dimension encoder of the residual neural network, feature encoding is performed on the time dimension signal of each scanning track of the ground penetrating radar B-scan data to obtain a feature data matrix containing high-dimensional features of the time dimension signal; Step 2, using a deep neural network of a state space model, and using the state space model to extract bidirectional correlation features from the feature data matrix along the measurement line dimension, to obtain a feature data matrix containing hyperbolic target features in the measurement line dimension; Step three, using the time dimension decoder of the transposed convolution operation, upsample and decode the feature data matrix obtained in step two and output the detection result image; the final hyperbolic target detection and positioning result is obtained by post-processing the detection result image.
2. The method for detecting hyperbolic targets by ground penetrating radar based on a state space model according to claim 1, characterized in that: The state space model is: y(n)=Ch(n)+Dx(n) Among them, x(n), h(n) and y(n) are the discrete input sequence, system hidden state and output sequence respectively, matrices A, B, C and D are linear layers that can be learned, Δt is the discrete step length, Represents the Hadamard product operation of matrices.
3. The method for detecting hyperbolic targets by ground penetrating radar based on a state space model according to claim 2, characterized in that: The discrete step length is: Δt=Sig(Eh(n-1)+Fx(n)) Among them, E and F are learnable linear layers, and Sig() is the Sigmoid activation function.
4. The method for detecting hyperbolic targets by ground penetrating radar based on a state space model according to claim 3, characterized in that: When the entire network model consisting of the encoder, deep neural network and decoder is jointly trained, the state space model training parameters include θ S ={A,B,C,D,E,F}, the state space model uses two loop mechanisms to transform the input features Computed as output containing hyperbolic features The first loop processes the feature matrix from n=1 to N. Get the hidden state matrix H, and the second loop processes the feature matrix from the opposite direction Get the hidden feature matrix H b ; Using the hidden state matrix H and the hidden feature matrix H b Compute the output sequence.
5. The method for detecting hyperbolic targets by ground penetrating radar based on a state space model according to claim 4, characterized in that: The first loop processes the scanning feature matrix from n=1 to N. Get the hidden state matrix H, specifically: h(n)=tanh(Δh(n)+h(n-1)) Among them, h(n) is the element of the hidden state matrix H.
6. The method for detecting hyperbolic targets by ground penetrating radar based on a state space model according to claim 4, characterized in that: The second loop processes the feature matrix from the opposite direction Get the hidden feature matrix H b , specifically: h b (n)=tanh(Δh b (n)+h b (n-1)) Among them, in the second cycle h b The expression of h(n) is the same as h(n), h b (n) is the hidden feature matrix H b elements.
7. The method for detecting hyperbolic targets by ground penetrating radar based on a state space model according to claim 4, characterized in that: The hidden state matrix H and the hidden feature matrix H b The calculated output sequence is:
8. The method for detecting hyperbolic targets by ground penetrating radar based on a state space model according to claim 1, characterized in that: The encoder consists of a two-dimensional convolutional layer that performs convolution along the time dimension of the B-scan data, two feature extraction modules consisting of residual neural network modules and maximum pooling layers, and a module consisting of flattening and linear layers.
9. The method for detecting hyperbolic targets by ground penetrating radar based on a state space model according to claim 1, characterized in that: The encoder consists of a linear layer, three decoding modules consisting of residual neural network modules and transposed convolutional layers, and a Sigmoid activation function layer.
10. The method for detecting hyperbolic targets by ground penetrating radar based on a state space model according to claim 8 or 9, characterized in that: The residual neural network modules in the decoder and encoder consist of two two-dimensional convolutional layers and a residual connection.