Ground penetrating radar clutter suppression method based on principal component analysis and deep neural network

Abstract

The invention discloses a ground penetrating radar clutter suppression method based on principal component analysis combined with a deep neural network. The method comprises the following steps: firstly, acquiring original radar data received by a ground penetrating radar GPR system; preprocessing the original radar data to obtain preprocessed radar data; then, inputting the preprocessed radar data into a pre-trained target convolutional neural network PGCNN model, and outputting an initial clutter suppression result; wherein a principal component analysis (PCA) algorithm is fused into the target convolutional neural network PGCNN model; and finally, inputting the preprocessed radar data and the initial clutter suppression result into a pre-trained target convolutional auto-encoder CAE model, and outputting a target clutter suppression result. According to the method, complex clutters are effectively suppressed in a nonlinear environment, real target information can be reserved to the maximum extent, and a solution with high precision and real-time performance is provided for clutter suppression of a GPR system.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing technology, specifically relating to a ground-penetrating radar clutter suppression method based on principal component analysis combined with deep neural networks. Background Technology

[0002] Ground-penetrating radar (GPR) is an advanced non-destructive testing device widely used in engineering surveys, facility assessments, and other fields. When GPR systems detect data, coherent clutter, composed of direct-coupled waves and surface-reflected waves, is the primary factor affecting imaging quality and detection accuracy because its energy is far greater than that of underground target echoes. However, traditional clutter suppression algorithms based on linear prior models struggle to effectively handle complex nonlinear clutter information.

[0003] Existing clutter suppression techniques have developed along two main paths. One path is based on traditional algorithms using prior signal models, such as the MS algorithm, median suppression algorithm, SVD algorithm, and PCA algorithm. The core of these algorithms is the assumption that clutter and signals are separable in a linear space. However, these linear model-based algorithms have fundamental limitations. The effectiveness of MS and median methods drops sharply when dealing with non-flat surfaces; while subspace methods such as SVD and PCA, when facing horizontally layered targets with similar clutter characteristics, suffer irreversible damage to the effective signal due to the linear aliasing of the signal and clutter subspaces, even while suppressing clutter.

[0004] To overcome the limitations of linear models, another approach is to introduce neural networks with nonlinear expressive capabilities. For example, applying convolutional autoencoders to clutter suppression of 2D GPR data can effectively suppress clutter in B-scan data. However, for the robustness of the network model, a significant amount of time is required to generate simulated datasets, and the generalization ability of the network model is limited when the measured dataset differs greatly from the simulated dataset.

[0005] To maximize the nonlinear expressive power of neural network models while minimizing the algorithm's dependence on the GPR dataset, an autoencoder (AE)-guided GPR clutter suppression algorithm is employed. This algorithm utilizes the AE as a tool to perform nonlinear decomposition of the clutter and target subspaces, thus achieving effective suppression of GPR clutter without requiring training the network model on the GPR dataset. However, this method still faces some challenges: (1) Real-time issue. Each iteration requires backpropagation calculation of the network weights inside AE. The clutter suppression algorithm is time-consuming and has almost no real-time capability in engineering applications.

[0006] (2) Scalability bottleneck of architecture. Traditional AE mostly uses fully connected linear layers, and the number of network parameters will increase quadratically or even higher with the increase of the input image size. This makes it difficult for the network to directly process high-resolution GPR data. If downsampling is used to adapt to small input size, the detailed information in the image will inevitably be lost, affecting the suppression accuracy.

[0007] (3) Structure awareness problem. Traditional AE architecture lacks the ability to perceive the spatial structure of images. It is better at separating global, statistical content, but it is difficult to understand the geometric information in GPR images. This structure insensitivity is the underlying reason for its insufficient ability to protect weak and complex targets.

[0008] The limitations of existing technologies restrict the detection performance of GPR systems in complex environments. Therefore, there is an urgent need for a clutter suppression method for GPR that can effectively solve the above problems, achieve effective suppression of complex clutter in nonlinear environments, and preserve the true target information to the maximum extent. Summary of the Invention

[0009] To address the aforementioned problems in the existing technology, this invention provides a ground-penetrating radar clutter suppression method based on principal component analysis combined with a deep neural network.

[0010] The technical problem to be solved by this invention is achieved through the following technical solution: This invention provides a ground-penetrating radar clutter suppression method based on principal component analysis and deep neural networks, comprising: Acquire raw radar data received by the ground-penetrating radar (GPR) system; The raw radar data is preprocessed to obtain preprocessed radar data; Preprocessed radar data is input into a pre-trained target convolutional neural network (PGCNN) model, which outputs initial clutter suppression results; principal component analysis (PCA) algorithm is incorporated into the target convolutional neural network (PGCNN) model. The preprocessed radar data and initial clutter suppression results are input into the pre-trained target convolutional autoencoder (CAE) model, which outputs the target clutter suppression results.

[0011] This invention provides a ground-penetrating radar (GPR) clutter suppression method based on principal component analysis (PCA) combined with a deep neural network. By integrating the linear dimensionality reduction capability of PCA with the nonlinear mapping capability of convolutional neural networks, it achieves precise decoupling between the clutter and effective signal subspaces. Furthermore, a lightweight convolutional autoencoder (CAE) model is introduced, and the processing mode is solidified through unsupervised learning to improve the algorithm's computational efficiency. This invention effectively overcomes the performance bottleneck of traditional subspace projection methods, providing a high-precision and real-time solution for clutter suppression in GPR systems.

[0012] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating a ground-penetrating radar clutter suppression method based on principal component analysis and deep neural network provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of the target PGCNN model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the processing procedure of the target PGCNN model provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the processing results of the target PGCNN model provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of the target CAE model provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the processing procedure for the target CAE model provided in an embodiment of the present invention; Figures 7A to 7B This is a simulation experiment diagram of a ground-penetrating radar clutter suppression method based on principal component analysis and deep neural network provided in an embodiment of the present invention. Detailed Implementation

[0014] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0015] This invention provides a method for suppressing clutter in ground-penetrating radar (GPR) based on principal component analysis combined with a deep neural network. See also... Figure 1 The method includes the following steps: S10. Acquire the raw radar data received by the ground-penetrating radar (GPR) system.

[0016] For example, the raw radar data includes target radar data and clutter.

[0017] S20. Preprocess the raw radar data to obtain preprocessed radar data.

[0018] Optionally, step S20 may specifically include: centralizing the raw radar data to obtain preprocessed radar data.

[0019] By centralizing the raw radar data, the impact of differences in the mean values ​​of different features in the raw radar data can be eliminated.

[0020] S30. Input the pre-processed radar data into the pre-trained target convolutional neural network PGCNN model and output the initial clutter suppression result.

[0021] Principal Component Analysis (PCA) algorithm is incorporated into the target convolutional neural network (PGCNN) model.

[0022] Optionally, the target convolutional neural network PGCNN model includes a principal component analysis (PCA) guiding module and a convolutional neural network module.

[0023] Step S30 may specifically include: S301. Input the pre-processed radar data into the pre-trained target convolutional neural network PGCNN model, and process the pre-processed radar data through the principal component analysis (PCA) guidance module to output the target dimensionality reduction features.

[0024] For example, the Principal Component Analysis (PCA) guidance module is used to process preprocessed radar data using the PCA algorithm. The PCA algorithm performs operations such as eigenvalue decomposition on the covariance matrix to extract the principal components that best represent the variance in the data, which are then used for dimensionality reduction or feature extraction. The PCA algorithm is an unsupervised linear dimensionality reduction algorithm. Its core objective is to map high-dimensional data to a low-dimensional space while preserving the main information of the data, thus solving problems such as sparsity, high computational cost, and difficulty in visualization of high-dimensional data.

[0025] Since the PCA algorithm can effectively reduce the impact of mean shift between different A-scans in the measured data on the clutter suppression effect and has stronger robustness, the PCA algorithm with background noise subspace is used to guide the weight iteration process of the network model in the target convolutional neural network PGCNN model.

[0026] S302. The target dimensionality reduction features are processed by the convolutional neural network module to output the initial clutter suppression result.

[0027] Optionally, the convolutional neural network module includes an input module, a data dimensionality reduction module, and a data reconstruction module.

[0028] Step S302 may specifically include: S3021. The target dimensionality reduction features are enhanced by the input module, and the first feature is output.

[0029] S3022. The clutter feature is extracted from the first feature through the data dimensionality reduction module, and the second feature is output.

[0030] S3023. The second feature is processed by the data reconstruction module to output the initial clutter suppression result.

[0031] For example, because convolutional layers share weights through local connections, they can avoid the problem of network parameters growing quadratically with the input size. Convolutional layers have two main characteristics: translation invariance and locality. Translation invariance means that the detected object remains within the input data... Translation in the middle only results in hidden representation The translation within the convolutional layer, based on this characteristic, significantly reduces the number of parameters in the convolutional layer. Locality, on the other hand, refers to the ability to collect parameters for training. The relevant information does not need to be offset to distance. A very distant location, that is or hour, In summary, a convolutional layer can be represented as: ; For a two-dimensional data Let its hidden representation be... It is a two-dimensional tensor, and and They have the same shape. Let and These represent the positions in the input two-dimensional data and the hidden representation, respectively. Data at the location, index and By moving between positive and negative offsets, the entire two-dimensional data is covered. . Represents the weight tensor. It is called a convolution kernel. This represents the offset parameter of the convolutional layer.

[0032] Since the same convolutional kernel shares weights at different locations, the number of parameters in a convolutional layer is independent of the input data size, depending only on the kernel size and the number of output channels. The calculation method for the number of parameters in a convolutional layer is shown in the following formula: ; in Indicates the kernel width. Indicates the height of the convolution kernel. and These represent the number of input and output channels, respectively.

[0033] As shown in the above formula, when using convolutional layers to design the network model, the number of parameters in the model is significantly reduced compared to a network model built with linear layers. Although the reduction in the number of parameters in convolutional layers is based on the requirement that features have translation invariance and local properties, for clutter information in GPR measured B-scan data, clutter information caused by subsurface inhomogeneity, surface interference, and non-target reflections can appear at different locations in the B-scan image, and usually has similarity across range cells or across azimuths, which is consistent with the modeling ability of convolutional layers for local correlations. By using a convolutional structure to design the network model, it is possible to increase the input data size while slowing down the increase in the number of network parameters, effectively reducing the hardware requirements for training the network model.

[0034] It's important to note that both linear and convolutional layers perform linear operations when used individually. To prevent the target PGCNN model from degenerating into a linear model, a non-linear activation function needs to be introduced into the network. Assume a hidden layer has weights... With bias An activation function is introduced after this hidden layer. For a given input data The hidden representation after passing through the hidden layer and activation function It can be written as: ; The target PGCNN network structure mainly uses two activation functions: ReLU and Leaky ReLU. These two activation functions can be expressed as follows: ; ; in, ReLU represents the scalar multiplier of negative input values. The piecewise nature of the ReLU activation function allows it to create complex nonlinear decision boundaries. Furthermore, since the gradient is always 1 in the positive region and 0 in the negative region, ReLU effectively avoids the gradient approaching zero problem in the saturation region of traditional activation functions like Sigmoid and Tanh, while maintaining network sparsity. However, if some neurons consistently output negative values ​​during training, using ReLU to activate the hidden layers may prevent weight updates, leading to neuron death. To mitigate the negative impact of neuron death on PGCNN, the Leaky ReLU activation function introduces a small gradient into the negative region, ensuring continuous gradient flow during backpropagation of the target PGCNN model and improving overall model stability.

[0035] Reference Figure 2The diagram illustrates the structural block diagram of the convolutional neural network module in the target PGCNN model. Based on their functions, the convolutional neural network module can be divided into an input module, a data dimensionality reduction module, and a data reconstruction module. The input module includes a convolutional layer (Conv-1), a ReLU activation function, and a pooling layer; the data dimensionality reduction module includes a convolutional block (Conv Block I), a linear connection block (Linear Block I), a convolutional block (Conv Block II), a pooling layer, and a convolutional layer (Conv-2); the data reconstruction module includes an upsampling layer, a convolutional block (DEConv Block I), a linear connection block (Linear Block II), a convolutional block (DEConv Block II), and a convolutional layer (Conv-3).

[0036] In the data dimensionality reduction section, clutter features in the original B-scan data are extracted step by step through two convolutional blocks, Conv Block I and Conv Block II. The structures of these two convolutional blocks used for data dimensionality reduction are as follows: Figure 2 As shown in (b). In the data reconstruction section, convolutional blocks DEConv Block I and DEConv Block II are used to reconstruct clutter information in the original B-scan data (i.e., the original radar data). The structures of the two convolutional blocks used to perform the data reconstruction are as follows: Figure 2 As shown in (c), to further improve the training speed of the target PGCNN model and alleviate the gradient vanishing and gradient exploding problems during training, a batch normalization function is introduced into the convolutional blocks to standardize the feature space. Furthermore, by introducing a series of nonlinear activation functions into the target PGCNN model, the nonlinear expressive power of the network model is further enhanced.

[0037] For example, the network parameters of each layer in the target PGCNN model are shown in Table 1: Table 1 Network structure parameters of the target PGCNN model

[0038] Furthermore, to improve the computational efficiency of the algorithm, this embodiment redesigns the loss function solution part of the network model, binding the judgment of clutter information reconstruction quality to the loss in the neural network structure, so that the network weights iterate according to the quality of clutter information reconstruction. The improved network model, when guiding GPR clutter suppression, no longer requires repeated training of the target PGCNN model, thereby improving the algorithm's efficiency.

[0039] Optionally, during the training of the target convolutional neural network PGCNN model, the first target loss function is expressed as: ; In the formula, Denotes the first objective loss function. This represents preprocessed radar data, where n is the channel index, indicating that this is the nth radar signal collected along the survey line; t is the time index, indicating the t-th time sampling point in a radar signal. This represents the prediction result output by the target convolutional neural network PGCNN model during training. This represents the network weights of the target PGCNN model. and These represent the principal spatial direction matrix formed by the eigenvectors corresponding to several largest eigenvalues ​​selected by adaptive thresholding after the preprocessed radar data has been processed by Principal Component Analysis (PCA) algorithm, and the secondary spatial direction matrix formed by the eigenvectors corresponding to several smallest eigenvalues. Denotes the square of the Frobenius norm; in, and They are represented as follows: ; ; In the formula, The eigenvector matrix represents the principal components. The eigenvector matrix represents the subcomponent.

[0040] After obtaining the value of the loss function, the Adaptive Moment Estimation (Adam) algorithm is applied to iteratively refine the weights of the neural network model, as follows: Optionally, during the training of the target convolutional neural network (PGCNN) model, the iterative update process of the network weights of the PGCNN model includes: Definition of the first During round training, the preprocessed radar data input to the target PGCNN model is: The network weights of the corresponding target PGCNN model are represented as follows: Then the gradient factor of the target PGCNN model Represented as: ; ; ; In the formula, Indicates the first The function value corresponding to the first objective loss function of the target PGCNN model during the first training round. Indicates the first The first moment during round training, Indicates the first Second moment during round training, and This represents different non-negative weighted parameters.

[0041] First moment in the configuration initialization phase and second moment When both are 0, a fairly large initial deviation is usually obtained. To reduce the impact of this phenomenon on the efficiency of weight iteration, this problem can be solved by the following application formula: The corresponding number The standardized state variables after each round of training can be represented as follows: ; ; ; In the formula, Represents the momentum component. Indicates the adaptive learning rate. ; Then the first The corresponding network weights of the PGCNN model Represented as: ; in, This represents the initial learning rate. This represents the numerical stability constant.

[0042] For example, typically, can be taken , As the default value for the parameter. Usually, it is taken as... . Used to accelerate model convergence. Used to control the update magnitude of network weights.

[0043] The overall processing flow of the target PGCNN model is as follows: Figure 3 As shown. Since the loss function of this part of the network structure reflects the approximation between the forward propagation result and the clutter space located by PCA, the network does not require the GPR dataset to train the network model. It can achieve the convergence of the network model by using only a single GPR dataset for weight iteration, and obtain the nonlinear clutter information extraction work based on PCA guidance.

[0044] Reference Figure 4The images show a before-and-after comparison of the target PGCNN model processing the original radar data. (a) shows the original radar data, and (b) shows the processing result of the target PGCNN model. Before inputting the original radar data into the target PGCNN model for clutter suppression, the original data needs to be standardized to ensure it can be effectively input into the network model. This method effectively suppresses clutter and background noise when processing target echoes with significant impact.

[0045] S40. Input the preprocessed radar data and initial clutter suppression results into the pre-trained target convolutional autoencoder (CAE) model, and output the target clutter suppression results.

[0046] For example, the target clutter suppression result is the clutter suppression result optimized from the initial clutter suppression result.

[0047] While the PGCNN model alone can achieve nonlinear decomposition of clutter and target signals in raw radar data, thus suppressing clutter and background noise in GPR data, this neural network structure requires iterative calculation of network weights for each different GPR dataset. This limitation leads to a long processing time. For example, on a computer with an AMD Ryzen 7 5700X CPU, the average time to suppress clutter in a single set of raw B-scan data using this algorithm is 95.4 seconds.

[0048] To further improve the computational efficiency of the clutter suppression algorithm, this embodiment introduces a target CAE model after the target PGCNN model. The target CAE model takes the preprocessed radar data and the initial clutter suppression results output by the target PGCNN model as network inputs, and directly learns the clutter suppression processing mode of the target PGCNN model through unsupervised learning. After training, the model can suppress clutter with only one forward propagation operation of the network model, eliminating the need for iterating the network weights for each different B-scan dataset. Compared to using only the target PGCNN model for clutter suppression, the introduction of the target CAE model enables the algorithm to achieve higher overall computational efficiency and real-time performance.

[0049] Reference Figure 5This paper illustrates the structure of a target CAE model, which consists of an encoder with several convolutional layers and a decoder with several convolutional layers. The number of convolutional layers in both the encoder and decoder is the same, and their overall structures are symmetrical. For the target CAE model, the loss function is typically the root mean square error (MSE) between the reconstructed data and the original data. The main function of the target CAE model is to achieve dimensionality reduction and reconstruction of the original data, making the network's output as similar to its input as possible.

[0050] This embodiment designs the network structure parameters of the target CAE model, as detailed in Table 2.

[0051]

[0052] Table 2. Network structure parameters of each layer of the target CAE model Simultaneously, the loss function of the target CAE model is adjusted. Utilizing the nonlinear expressive power of the target CAE model, while learning the clutter suppression mode of the target PGCNN model, local errors in the output of the target PGCNN model are further corrected, thereby further improving the computational speed and processing performance of the target PGCNN model.

[0053] Optionally, during the training of the target convolutional autoencoder CAE model, the second objective loss function is expressed as: ; ; in, Denotes the second objective loss function. This indicates preprocessing of radar data. This represents the prediction result output by the target convolutional autoencoder CAE model during training. This represents the prediction result output by the trained target convolutional neural network PGCNN model. This represents the network weights of the trained target convolutional neural network PGCNN model. Represents the network weights of the target CAE model. This indicates the initial clutter suppression result. This is the function for solving the root mean square error.

[0054] The weights of the target CAE model are also iterated using the Adam optimizer. The training process of the target CAE model is as follows: Figure 6 As shown.

[0055] This embodiment of the ground-penetrating radar clutter suppression method based on principal component analysis and deep neural networks divides the clutter suppression work in GPR data into two steps. First, the upper-layer neural network is guided by the PCA algorithm, combining the dimensionality reduction capability of the PCA algorithm with the nonlinear learning capability of the autoencoder to achieve nonlinear decomposition of the GPR data. Subsequently, by integrating the output results of a series of upper-layer neural networks and using the target CAE model to correct the local errors in the output of the target PGCNN network, the clutter suppression effect of the algorithm is further improved. At the same time, through learning from the clutter suppression results of a series of upper-layer networks, the generalization ability of the convolutional neural network will be continuously improved. After training, the target PGCNN model can suppress GPR clutter through a single forward propagation operation of the neural network. When the clutter suppression algorithm optimized by the target CAE model is executed on a computer with an AMD Ryzen 7 5700X CPU, it can output the clutter suppression result within 10ms, which significantly improves the calculation speed.

[0056] The following simulation experiment further illustrates the ground-penetrating radar clutter suppression method based on principal component analysis and deep neural network provided by this invention.

[0057] This embodiment will compare and analyze the performance of clutter suppression algorithms based on SVD, PCA, and the present invention when performing clutter suppression processing on measured data. First, the measured data is acquired using a GPR system, and the target region is extracted. The original B-scan data containing the target region is shown below. Figure 7A As shown in the figure, the area circled in blue is the target echo region, while the area circled in red contains clutter signals caused by uneven road surfaces. The results of processing the above measured data based on SVD, PCA, and the clutter suppression algorithm of this invention are shown below. Figure 7B As shown in (a), (b), and (c) of the table, the improvement of each clutter suppression algorithm on the measured data can be reflected in Table 3: Table 3. Measured data before and after implementing the clutter suppression algorithm Comparison table with image entropy

[0058] By comparing the experimental data, we can conclude that... Figure 7B While the classic SVD clutter suppression algorithm corresponding to (a) can effectively suppress direct reflection clutter from the ground, its effect on suppressing background noise is very limited. In addition, since it uses a fixed threshold for clutter suppression, the SVD clutter suppression algorithm also damages the signal integrity of the target echo while suppressing direct reflection clutter. Figure 7BThe PCA clutter suppression algorithm corresponding to (b) also has limited suppression of background noise. However, since this type of algorithm performs a series of signal preprocessing operations such as centering before execution, it reduces the negative impact of mean fluctuation on signal processing during the data acquisition process of the GPR system. Therefore, the background noise after PCA clutter suppression is more uniform than that of the SVD algorithm.

[0059] In comparison, this method achieves an improvement factor (IF) of 11.35 and an image entropy of 3.22. Compared to traditional algorithms, the IF value is increased by 5.22 to 6.14, and the image entropy is reduced by 1.96 to 2.56. The clutter suppression algorithm of this invention effectively suppresses most clutter and background noise while preserving the integrity of the target echo to the greatest extent. Furthermore, for complex clutter signals present within the red rectangular region, the clutter suppression algorithm of this invention also achieves more effective suppression than the other two algorithms.

[0060] It should be noted that the terms "first," "second," etc., are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention.

[0061] In summary, compared with existing technologies, the beneficial effects of this invention are as follows: 1. Effectively overcomes the limitations of traditional subspace methods: Traditional subspace projection algorithms often suffer from aliasing between the signal and clutter subspaces when processing signals with similar characteristics to clutter, due to their inherent linear model assumptions. This results in irreversible damage to the valid target information while suppressing clutter. This invention achieves precise decoupling between the clutter and valid signal subspaces by deeply integrating the linear dimensionality reduction capability of PCA with the nonlinear expression capability of neural networks. This ensures that the integrity of the true target signal is maximized while efficiently suppressing clutter.

[0062] 2. Balancing Structure Awareness and Detail Preservation: Compared to the shortcomings of AE (Advanced Image Processing) models in lacking the ability to perceive the spatial structure of images, this invention can better understand the geometric and spatial structural information in ground-penetrating radar (GPR) images. This allows the algorithm to more effectively protect weak and complex target signals while separating global clutter, solving the deep-seated problems of traditional AE models being insensitive to structure and insufficient in detail preservation.

[0063] 3. Addressing Real-Time Performance and Scalability Issues: This invention effectively reduces the number of model parameters and computational complexity through PCA guidance and a convolutional layer-based main structure, enabling it to efficiently process high-resolution GPR data and reducing reliance on hardware resources. Simultaneously, a target CAE model is introduced after the target PGCNN model, directly learning the clutter suppression processing mode of the target PGCNN model through unsupervised learning, improving the algorithm's computational efficiency in practical deployments and giving it real-time processing capabilities for engineering applications.

[0064] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0065] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings and the disclosure in carrying out the claimed invention. In the description of the invention, the word "comprising" does not exclude other components or steps, "a" or "an" does not exclude a plurality, and "a plurality" means two or more, unless otherwise explicitly specified. Furthermore, while different embodiments may describe certain measures, this does not mean that these measures cannot be combined to produce good results.

[0066] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A ground penetrating radar clutter suppression method based on principal component analysis combined with deep neural network, characterized in that, include: Acquire raw radar data received by the ground-penetrating radar (GPR) system; The raw radar data is preprocessed to obtain preprocessed radar data; The preprocessed radar data is input into a pre-trained target convolutional neural network (PGCNN) model, which outputs initial clutter suppression results; wherein, the target convolutional neural network (PGCNN) model incorporates principal component analysis (PCA) algorithm. The preprocessed radar data and the initial clutter suppression results are input into a pre-trained target convolutional autoencoder (CAE) model, which outputs the target clutter suppression results. 2.The ground penetrating radar clutter suppression method based on principal component analysis combined with deep neural network according to claim 1, wherein, The preprocessing of the original radar data to obtain preprocessed radar data includes: The raw radar data is centrally processed to obtain preprocessed radar data. 3.The ground penetrating radar clutter suppression method based on principal component analysis combined with deep neural network according to claim 2, characterized in that, The target convolutional neural network (PGCNN) model includes a principal component analysis (PCA) guidance module and a convolutional neural network module. The process of inputting the pre-processed radar data into the pre-trained PGCNN model and outputting initial clutter suppression results includes: The preprocessed radar data is input into a pre-trained target convolutional neural network (PGCNN) model, and the preprocessed radar data is processed by the principal component analysis (PCA) guidance module to output target dimensionality reduction features. The target dimensionality reduction features are processed by the convolutional neural network module to output the initial clutter suppression result.

4. The ground penetrating radar clutter suppression method based on principal component analysis combined with deep neural network according to claim 3, characterized in that, During the training of the target convolutional neural network PGCNN model, the first target loss function is expressed as: ； In the formula, Denotes the first objective loss function. This indicates preprocessing of radar data. This represents the prediction result output by the target convolutional neural network PGCNN model during training. This represents the network weights of the target PGCNN model. and These represent the principal spatial direction matrix formed by the eigenvectors corresponding to several largest eigenvalues ​​selected by adaptive thresholding after the preprocessed radar data has been processed by Principal Component Analysis (PCA) algorithm, and the secondary spatial direction matrix formed by the eigenvectors corresponding to several smallest eigenvalues. Denotes the square of the Frobenius norm; in, and They are represented as follows: ； ； In the formula, The eigenvector matrix represents the principal components. The eigenvector matrix represents the subcomponent.

5. The ground-penetrating radar clutter suppression method based on principal component analysis and deep neural network according to claim 4, characterized in that, During the training process of the target convolutional neural network (PGCNN) model, the iterative update process of the network weights of the PGCNN model includes: Definition of the first During round training, the preprocessed radar data input to the target PGCNN model is: The network weights of the corresponding target PGCNN model are represented as follows: Then the gradient factor of the target PGCNN model Represented as: ； ； ； In the formula, Indicates the first The function value corresponding to the first objective loss function of the target PGCNN model during the first training round. Indicates the first The first moment during round training, Indicates the first Second moment during round training, and Represents different non-negative weighted parameters; The corresponding number The standardized state variables after each round of training can be represented as follows: ； ； ； In the formula, Represents the momentum component. Indicates the adaptive learning rate. ; Then the first The corresponding network weights of the PGCNN model Represented as: ； in, This represents the initial learning rate. This represents the numerical stability constant.

6. The ground-penetrating radar clutter suppression method based on principal component analysis and deep neural network according to any one of claims 1-5, characterized in that, During the training of the target convolutional autoencoder CAE model, the second objective loss function is expressed as follows: ； ； in, Denotes the second objective loss function. This indicates preprocessing of radar data. This represents the prediction result output by the target convolutional autoencoder CAE model during training. This represents the prediction result output by the trained target convolutional neural network PGCNN model. This represents the network weights of the trained target convolutional neural network PGCNN model. Represents the network weights of the target CAE model. This indicates the initial clutter suppression result. This is the function for solving the root mean square error.

7. The ground-penetrating radar clutter suppression method based on principal component analysis and deep neural network according to claim 3, characterized in that, The convolutional neural network module includes an input module, a data dimensionality reduction module, and a data reconstruction module. The process of processing the target dimensionality reduction features through the convolutional neural network module to output initial clutter suppression results includes: The input module performs feature enhancement on the target dimensionality reduction features and outputs the first feature. The data dimensionality reduction module extracts clutter features from the first feature and outputs the second feature. The data reconstruction module processes the second feature and outputs the initial clutter suppression result.

Images