Ground penetrating radar clutter suppression method based on graph signal processing

Abstract

The invention relates to a ground penetrating radar clutter suppression method based on graph signal processing. The method comprises the steps of collecting a ground penetrating radar B-type scanning image of a to-be-tested environment; converting the B-type scanning image of the ground penetrating radar into a vector, and performing image Fourier transform; inputting a result after the Fourier transform of the graph into a designed wiener filter variant based on a graph theory to carry out clutter suppression; inverse Fourier transform is carried out on the filtered image, and the filtered image is remodeled into the matrix size of the original image. Compared with the prior art, the method has the advantages of effectively realizing clutter suppression, adapting to a complex environment, improving the signal quality of a target and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of ground penetrating radar image clutter suppression, and in particular to a ground penetrating radar clutter suppression method based on image signal processing. Background Art

[0002] Ground-Penetrating Radar (GPR) has been widely used in many fields such as civil engineering, geological engineering, archaeology, mineral exploration and military. However, when identifying underground targets, GPR often faces significant background clutter interference, which mainly comes from reflections from the air-ground interface, the heterogeneity of the underground medium, and direct coupling between the transmitting and receiving antennas. The environmental clutter in the B-scan image of GPR poses a major challenge to the accurate detection and interpretation of targets.

[0003] Background subtraction type clutter suppression methods include mean subtraction (MS) and background matrix subtraction (BMS). These methods work well in homogeneous underground environments with flat surfaces because the average trajectory of the target-free area can accurately reflect the background clutter information. However, when the underground environment has a rough surface and is non-uniform, the performance of these methods will be greatly reduced because the background clutter will vary in different trajectories of the B-scan. Subspace-based methods such as singular value decomposition (SVD), principal component analysis (PCA) and independent component analysis (ICA) are also widely used in GPR clutter suppression. However, in order to achieve the best results, such methods must accurately estimate the clutter subspace. And such methods assume that there is a significant difference between clutter and target responses, so that the clutter information is mainly captured by the principal component. However, when the clutter and target signal strengths are similar, it becomes difficult to separate them, which may weaken the target signal during the clutter suppression process. Low-rank and sparse matrix factorization (LRSD) methods, such as robust PCA (RPCA), GoDec decomposition, robust non-negative matrix factorization (RNMF), robust orthogonal subspace learning (ROSL), and low-rank sparse factorization with total variation regularization (LRSD-TVR) are also used for GPR clutter suppression. The performance of these methods depends largely on the choice of hyperparameters to balance the low-rank and sparse components, which vary from case to case. It should be noted that the LRSD method focuses on strong low-rank and sparse components and may ignore the residual clutter in the sparse matrix, resulting in poor performance when the clutter in the sparse (target) matrix is ​​difficult to remove. In addition, the inconsistency of the LRSD method in different underground scenarios (such as the variation of target size and position) also poses additional challenges. In recent years, deep learning methods have also been tried for GPR clutter suppression in specific contexts, including autoencoder-based methods, convolutional autoencoder (CAE) network methods, and U-Net-based clutter suppression neural network (CR-Nets) architecture. Although deep learning has performed well in many fields, it usually requires a large amount of training data and specific conditions to achieve the best results. The generalization ability of deep learning methods to diverse and unknown scenes in GPR clutter suppression and target detection is still an active research area. In addition, it is also a challenge to obtain B-scan images of pure underground targets without clutter for training data in actual GPR experiments.

[0004] Graph signal processing (GSP) is a field that focuses on analyzing and processing signals defined on graphs. It is widely used in many image processing fields, such as brain imaging, SAR image analysis, and face image processing. These defined graphs are mathematical structures that represent the signal relationship between objects or entities. GSP performs well in clutter suppression and filtering. Given that the raw B-scan images of GPR need to be processed for clutter suppression, GSP technology is very suitable for application to such images. However, according to the latest literature survey, the application of GSP technology in the field of GPR imaging has not been reported.

[0005] How to achieve effective clutter suppression for ground penetrating radar B-scan images based on image signal processing has become a technical problem that needs to be solved. Summary of the invention

[0006] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a ground penetrating radar clutter suppression method based on image signal processing.

[0007] The purpose of the present invention can be achieved by the following technical solutions:

[0008] According to one aspect of the present invention, a method for suppressing clutter of a ground penetrating radar based on static image signal processing is provided, the method comprising:

[0009] Collecting ground penetrating radar B-scan images of the environment to be tested;

[0010] Convert the ground penetrating radar B-scan image into a vector and perform a graph Fourier transform;

[0011] The Fourier transformed result of the graph is input into the designed graph theory-based Wiener filter variant for clutter suppression;

[0012] The filtered image is inversely Fourier transformed and reshaped to the matrix size of the original image.

[0013] Preferably, the graph theory-based Wiener filter variant is specifically:

[0014]

[0015] in, is the frequency domain vector of the filtered image; p x represents the power spectral density of the B-scan image set containing a mixture of target and clutter, p y represents the power spectral density of the B-scan image set containing only clutter, is the frequency domain vector of the image after Fourier transformation corresponding to the B-scan image to be suppressed.

[0016] More preferably, the power spectral density p x The calculation formula is:

[0017]

[0018] Among them, x r is a vector sample converted from a single image in a B-scan image set containing a mixture of targets and clutter; m is x r The number of samples; For x r The sample mean of ; diag is a diagonal matrix function; X = [x1,…,x m ] covariance; V is the Fourier basis of the graph, that is, The feature vector of .

[0019] More preferably, the power spectral density p y The calculation formula is:

[0020]

[0021] Among them, V is the graph Fourier basis, that is, The eigenvector of is the sample covariance matrix of the B-scan image sample matrix Y containing only clutter, Y = [y1,…,y n ], n is the number of B-scan image samples containing only clutter, y n It is a vector sample converted from a single image in the B-scan image set containing only clutter.

[0022] More preferably, the construction process of the graph theory-based Wiener filter variant includes:

[0023] Step S101, obtaining a B-scan image set of a ground penetrating radar under a test environment, and dividing it into a B-scan image set containing a mixture of targets and clutter and a B-scan image set containing only clutter;

[0024] Step S102, respectively converting the B-type scan image containing a mixture of target and clutter and the B-type scan image containing only clutter into vectors, which are respectively recorded as B-type scan image sets and Where m is the number of B-scan images containing a mixture of targets and clutter, and n is the number of B-scan images containing only clutter;

[0025] Step S103, calculate The covariance of calculate The covariance of

[0026] Step S104, covariance Perform eigendecomposition to compute the graph Fourier basis V, compute the graph shift operator, and compute The power spectral density p x ; for covariance Perform static analysis. After the static analysis passes, calculate The power spectral density p y ;

[0027] Step S104: constructing a graph-theoretic Wiener filter variant based on the power spectral density.

[0028] More preferably, after the static analysis of the B-scan image vector containing only clutter is passed, the power spectrum density calculation of the clutter in step S104 is performed.

[0029] More preferably, the static analysis process includes:

[0030] Defining Metrics in‖‖ F represents the Frobenius norm, which calculates the square root of the sum of the squares of the matrix elements; diag is the diagonal matrix function;

[0031] When calculated When it is greater than 0.5, it indicates that the B-scan image vector containing only clutter is weakly static with respect to the image shift operator S, that is, the static analysis is passed.

[0032] More preferably, the graph shift operator is equal to , which is used to represent the graph Denoted as S, where is the number of nodes.

[0033] More preferably, assume that a signal x is weakly static with respect to the graph shift operator S, and the graph shift operator S and the signal covariance C x have the same set of orthogonal eigenvectors, the covariance C x From the sample covariance Approximately, where V and Λ c They are The eigenvectors and eigenvalues ​​of for covariance of , m is the number of B-scan images containing a mixture of target and clutter.

[0034] More preferably, an automatic gain control method is used to divide the ground penetrating radar B-type scanning image according to whether the B-type scanning image contains target features. If it contains target features, it is divided into a B-type scanning image containing a mixture of target and clutter; otherwise, it is divided into a B-type scanning image containing only clutter.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1) The present invention introduces a framework based on graph signal processing, converts the B-scan image of the ground penetrating radar into a vector, performs graph Fourier transform, and inputs the designed graph theory-based Wiener filter variant to suppress clutter, which effectively reduces the influence of clutter in the B-scan image. Compared with traditional time domain or frequency domain filtering methods, this method is more efficient in dealing with strong clutter interference in complex environments.

[0037] 2) The present invention suppresses clutter in the graph frequency domain by using a Wiener filter variant based on graph theory. The Wiener filter variant is implemented based on power spectral density and can better preserve the scattering characteristic signal of the target, thereby improving the signal quality of the target.

[0038] 3) The present invention uses a reference sample set to construct a highly adaptable graph shift operator, and dynamically estimates the graph Fourier basis through the eigendecomposition of the sample covariance matrix, thereby solving the non-stationary problem of clutter. This method is still effective when applied to targets of various sizes and positions, and can achieve adaptive filtering without relying on specific ground truth conditions, thereby enhancing adaptability to complex environments.

[0039] 4) The present invention uses simulation to compare the method of the present invention with multiple existing algorithms, verifying the excellent performance of the method in suppressing surface and underground clutter while maintaining the integrity of the target signal, thereby significantly improving the detection accuracy of the target signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic diagram of the framework of the clutter suppression method of the ground penetrating radar B-type scanning image in the present invention;

[0041] Figure 2 It is a schematic diagram of the B-scan image before and after the single target simulation clutter removal in the present invention;

[0042] Figure 3 It is a schematic diagram of the B-scan image before and after the multi-target simulation clutter removal in the present invention;

[0043] Figure 4 For the present invention Figure 2 Schematic diagram of the comparison of two A-scan results after different processing methods;

[0044] Figure 5 It is a schematic flow chart of the method for suppressing clutter of ground radar in the present invention. DETAILED DESCRIPTION

[0045] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0046] This embodiment relates to a ground penetrating radar clutter suppression method based on image signal processing, which aims to cope with the complexity of clutter suppression of B-scan images of ground penetrating radar (GPR) in actual scenarios, and proposes a new framework for GPR clutter suppression based on static image signal (SGS) processing technology to achieve the purpose of suppressing clutter signals and enhancing target signals. The method regards the GPR B-scan image as an image signal, converts it to the image frequency domain using the image Fourier transform (GFT), and subtracts the power density of the clutter part from the power density of the high-clutter image to achieve a B-scan radar image with a low signal-to-noise ratio.

[0047] The main steps of this method are as follows:

[0048] 1) First, a B-scan image containing a mixture of target and clutter and a B-scan image containing only clutter are collected under the same test environment, and the automatic gain control method (AGC) is used to distinguish the two types of images;

[0049] 2) Secondly, the B-scan image containing a mixture of target and clutter is used as the reference data sample for constructing the graph shift operator (GSO), and the clutter signal and the scattering signal of the target are used as the static graph signal (SGS).

[0050] 3) Finally, after establishing the weak static graph signal relative to GSO, a graph theory-based Wiener filter variant is designed for GPR to achieve clutter suppression in the B-scan image of ground penetrating radar (GPR).

[0051] With respect to step 1), an image containing only clutter is defined as a B-scan image that does not contain a hyperbola, while the B-scan image of the GPR to be subjected to clutter suppression contains hyperbolas of different sizes and positions generated by underground target scattering and background clutter. In order to distinguish whether the B-scan image contains a hyperbola (i.e., target features) or is composed only of clutter, the present invention adopts an automatic gain control (AGC) method, which is based on A-scan data, and a moving window is used to gradually adjust the gain during this process. The moving window moves downward along the time axis sample by sample, that is, each sampling point is processed one by one. In order to achieve gain adjustment, a scaling factor g(x,t) is used, which is defined as the inverse of the root mean square value of the signal in the time window. This scaling factor is used to scale the signal amplitude in the window to ensure that the dynamic range of the signal remains consistent at each position, so that weak signals can be more clearly displayed in the data, and is defined as follows:

[0052]

[0053] Where D represents the input 2D GPR B-scan image; N is the number of pixels in the moving window, and the window is weighted using a Gaussian kernel to ensure that only the pixels in the center of the window match the input data; i represents the row number.

[0054] By applying AGC to each A-scan in the original ground penetrating radar B-scan image, the weaker echo signal from the underground target can be enhanced to make it visible in the strong background clutter, while avoiding saturation of the strong clutter. In this application, the role of AGC is limited to enhancing the visibility of weak target features in the original B-scan and helping to determine whether the target features are contained in the B-scan image. In addition, AGC is also used to verify whether the underground target features in the B-scan image after the clutter suppression method is applied are consistent with the features in the original B-scan image. The image processed by AGC is only used as a reference to help distinguish whether the B-scan image contains hyperbolic features or is composed only of clutter.

[0055] Regarding step 2), a graph is used to represent the pairwise relationship between elements in a set and has intrinsic meaning. Suppose a weighted undirected graph in Represents a collection of nodes; represents the set of edges, and defines is the number of nodes. If there is a connection between node i and node j, it is recorded as (i,j)∈ε, and the weight of edge (i,j) is w ij In the context of graph signal processing (GSP), a graph signal is defined as a vector Each element x(i) represents the value of the signal at node i. In order to process these signals in the graph framework, this application introduces the concepts of graph shift operator (GSO) and graph Fourier transform (GFT). GSO is defined as a function of size equal to , which is used to represent the graph Denoted as S, the rank of the matrix is ​​S if and only if i = j or (i, j) ∈ ε ji ≠0. GSO captures the connection relationship between graph nodes. Usually, the graph Laplacian matrix or adjacency matrix is ​​selected as GSO. In this paper, we use the adjacency matrix as GSO, which is defined as follows:

[0056]

[0057] Here, we consider a fully connected graph (i.e. ), and w ij is the covariance of the signal at node i and node j, that is, w ij =cov(x[i],x[j]). By performing eigendecomposition on GSO, we can obtain a unit matrix V = [v1,…,v N] and a diagonal matrix Λ, satisfying S = VΛV H The graph Fourier transform (GFT) process can be expressed as:

[0058]

[0059] in, is the frequency domain signal after x transformation, and V is a unit matrix. In Graph Signal Processing (GSP), the vector set {v i} is regarded as a basis signal with different changing frequencies on the graph, similar to the trigonometric functions in the classical Fourier transform, so {v i} is also called the graph Fourier basis. Therefore, by performing a graph Fourier transform (GFT), we can decompose the signal x into the graph frequency domain so that the frequency components of the signal on the graph can be revealed.

[0060] Step 2) Specific implementation:

[0061] Take the sample set containing 100 reference original GPR B-scan images as an example, denoted as These images contain hyperbolic shapes corresponding to clutter, reflecting 10 cylindrical targets with different radii. Each target appears in 10 different locations in the image. In addition, an independent reference sample set containing only clutter contains 10 images, denoted as The size of each B-scan image is set to 250×41 and converted into a vector form. Specifically, each A-scan is connected one by one, the first time point of the second A-scan is connected to the last time point of the first A-scan, and so on, to complete the splicing of all A-scans. For simplicity, the B-scan image of the ground penetrating radar is still represented as x r or r , whose vector size is 10250×1 (10250=250×41). Then, all vectors representing the B-scan image set containing the target and clutter are horizontally spliced ​​to construct a matrix X of size 10250×100, where each column represents the vector representation of a B-scan image, denoted as Similarly, construct a matrix Y of size 10250×10, representing the B-scan image set containing only clutter, denoted as The above process corresponds to Figure 1 Second, the B-scan image set is regarded as 100 realizations of a zero-mean random graph signal x, and its covariance is recorded as

[0062] Assume that the signal x is weakly static with respect to an implicit graph shift operator S. Under this assumption, S and the signal covariance C xhave the same set of orthogonal eigenvectors. Therefore, even if we cannot obtain the true S, we can still obtain it by x Perform eigendecomposition to calculate the graph Fourier basis V. In this practice, the covariance (i.e., GSO) C x The sample covariance Approximately, where V and Λ c They are The eigenvectors and eigenvalues ​​of . The sample signal matrix X = [x1,…,x 100 ] is the covariance of . This process corresponds to Figure 1 Part 2 of .

[0063] For step 3), to verify the signal vector y r Regarding S being weakly static, we need to prove that the sample matrix of the B-scan image containing only noise is Y = [y1,…,y 10 The sample covariance matrix of ] It can be approximately diagonalized by V (graph Fourier basis). To verify this assumption, the matrix The block matrix of How is it approximately diagonalized by the graph Fourier basis V. To quantify this relationship more precisely, we use the following metric:

[0064]

[0065] in,‖‖ F represents the Frobenius norm, which calculates the square root of the sum of the squares of the matrix elements; diag is the diagonal matrix function. This indicator is used to evaluate The importance of diagonal elements in . The process of static analysis is shown in Figure 1 Section 3.1 of When it is greater than 0.5, it indicates that the graph signal vector y r Regarding S being weakly static, continue with Section 3.2.

[0066] Then, if Figure 1 As shown in Section 3.2 of , we estimate the power spectral density (PSD) of clutter by the following formula:

[0067]

[0068] Finally, we use the graph theory-based Wiener filter to suppress clutter from the original B-scan image containing the hyperbola. Here, the B-scan image containing the hyperbola and clutter is represented as a vector x rBy applying formula (3), we transform the B-scan image into the image frequency domain. In this image frequency domain, the clutter is filtered out by a graph-theoretic Wiener filter variant designed specifically for GPR clutter suppression:

[0069]

[0070] in, is the frequency domain vector representation of the filtered image; p x represents the PSD of the original (mixed with target and clutter) B-scan image sample set, p y Represents the PSD of a clutter image sample set.

[0071] The PSD of the mixed B-scan image sample set containing the target and clutter is set as

[0072]

[0073] Equal to the sample covariance matrix The characteristic value of is the vector x obtained by the Graph Fourier Transform (GFT) r The frequency domain representation of the graph; m is the vector x of the B-scan image containing the target and clutter. r The number of samples; is x r The multiplication and division in formula (7) are calculated element by element. Figure 1 This is shown in Section 4.1 of .

[0074] Then, the filtered vector From the inverse transformation of the frequency domain, we get Finally, in order to obtain the B-scan image after clutter suppression, we use the filtered vector Reshape to the original image matrix size, such as Figure 1 As shown in Section 4.2 of

[0075] In summary, the present invention has significant advantages over the prior art in terms of clutter suppression effect, adaptability, enhanced recognition of weak target features, and target signal recovery capability, and can be more effectively applied to complex underground target detection tasks.

[0076] The clutter suppression using the constructed graph-theoretic Wiener filter variant includes the following steps: Figure 5 :

[0077] Step S1, collecting a ground penetrating radar B-type scanning image of the environment to be tested;

[0078] Step S2, converting the ground penetrating radar B-scan image into a vector and performing a graph Fourier transform;

[0079] Step S3, inputting the result of the graph Fourier transform into the designed graph theory-based Wiener filter variant to perform clutter suppression;

[0080] Step S4: Perform inverse Fourier transform on the filtered image and reshape it into the matrix size of the original image.

[0081] This application conducts simulation experiments on this method, and quantitatively analyzes the effect of the designed algorithm in clutter suppression by using the open source gprMax simulation software experiment that is widely used in the GPR field, and compares it with the low-rank and sparse matrix decomposition method (RPCA), the background subtraction method (BMS), and the subspace-based method (SVD), reflecting the high signal-to-noise ratio of the results of the present invention.

[0082] Experimental setup: In this specific implementation, the modeling of heterogeneous soils uses a combination of semi-empirical models and fractal models. The soil model assumes specific values, including a sand fraction of 0.9, a clay fraction of 0.1, and a bulk density of 2 g / cm 3 , soil particle density is 2.66g / cm 3 . In order to reflect the heterogeneity, 50 different elements (water volume fraction from 3% to 10%) are randomly distributed in a random manner, using a fractal dimension of 1.5. The size of the heterogeneous soil environment space is 120×54×50cm. The underground target is represented by a cylindrical target, and cylinders of different sizes are placed at different positions in the soil. The minimum unit of the spatial discretization of the simulation is set to 0.002m.

[0083] The GPR system uses a bistatic model, where both the transmitter (TX) and receiver (RX) are polarized parallel to the longitudinal dimension of the cylinder. The TX and RX are 0.10 m apart and 0.01 m above the soil surface. A 1 GHz center frequency thunder pulse is used as the excitation waveform, the soil dielectric constant ranges from 6 to 8, and the wavelength is about 10 cm. To collect data, the TX and RX move with a step size of 0.025 m, and a total of 41 A-scans are collected, which are then merged into a B-scan. The above simulation parameters are shown in Table 1.

[0084] Table 1

[0085]

[0086] A single target result comparison and a multi-target result comparison are used for demonstration. The B-scan images before and after the single target and multi-target simulation clutter removal are shown as follows: Figure 2 and Figure 3As shown in the figure, the method proposed by the present invention is represented as "proposed", Raw A-scan is the original A-scan image, and Raw B-scan is the original B-scan image ( Figure 4 The figure also includes the signal clutter suppression parameters: I f , the clutter suppression effects of different methods are compared from a data perspective. The calculation method is to convert the signal-to-clutter ratio (SCR) of the processed B-scan image into processed ) and the signal-to-clutter ratio (SCR) of the original B-scan image raw ) to find the ratio:

[0087]

[0088] The signal-to-clutter ratio is determined by calculating the ratio between the average amplitude of the target scattered signal and the average amplitude of the clutter:

[0089]

[0090] Among them, N s and N n are the selected target signal pixels P s and clutter pixel P n The number of p is the amplitude of the pth pixel.

[0091] like Figure 2 As shown in the figure, in the existing RPCA, BMS and SVD methods, there is an obvious problem of fusing residual clutter with the object scattered signal. Although the SVD method performs relatively well in filtering out the air-ground interface clutter, it will produce obvious horizontal bar noise, as shown by the black arrow in the figure. This bar noise will interfere with the signal extraction of the target object, making it more difficult to accurately extract the response of the object.

[0092] Figure 3 The comparison of clutter suppression effects of different methods in multi-target scenarios is shown. RPCA and BMS fail to completely remove background clutter. These two methods tend to distort or weaken the tail of the target parabola (marked by the red arrow), reducing the recognition of the target shape. SVD performs well in suppressing clutter amplitude, but introduces residual strip noise. In contrast, the proposed method performs well in suppressing clutter, can clearly maintain the target response, and effectively retains the tail characteristics of the parabola. In the multi-target parabola intersection scenario, the tail below the intersection is still clearly discernible. The advantage of the proposed method is that it obtains the highest signal clutter suppression parameter (I f ) value, which further verifies the superiority of the proposed method.

[0093] from Figure 2 and Figure 3 It can be seen that the method applied for in the present invention performs excellently in suppressing clutter in both single-target and multi-target scenarios, compared with the prior art, and has strong scene adaptability.

[0094] To deeply analyze the performance of these four methods, Figure 4 Shown Figure 2 The two A-scan trajectories (A-scan) marked by white dotted lines I and II, the ellipse highlights the peak of the underground target, and Amplitude is the signal amplitude of the A-scan. In the original A-scan, since the amplitude of the surface signal is significantly higher than the target signal, it becomes difficult to identify the target. The method proposed in the present invention performs well in removing clutter and makes the target signal clearer. In contrast, although the traditional method can effectively enhance the signal of the underground target, it cannot fully remove the residual clutter noise underground. Figure 4 In part (b) of the above, the SVD method introduces significant stripe noise, whose amplitude is higher than that of other methods in the sample range (Samples) 125 to 145, corresponding to Figure 2 The strip noise of the B-scan affects the clarity of the target signal. The BMS method has the worst clutter suppression effect. Figure 4 As shown in part (b) of , the amplitude of some clutter even exceeds the value of the target signal, which seriously affects the recognition of the target signal. In contrast, the method proposed in the present invention suppresses surface and underground clutter while maintaining the integrity of the target signal, thereby significantly improving the detection accuracy of the target signal.

[0095] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A ground penetrating radar clutter suppression method based on graph signal processing, characterized in that: The method includes: Collecting ground penetrating radar B-scan images of the environment to be tested; Convert the ground penetrating radar B-scan image into a vector and perform a graph Fourier transform; The Fourier transformed result of the graph is input into the designed graph theory-based Wiener filter variant for clutter suppression; The filtered image is inversely Fourier transformed and reshaped to the matrix size of the original image.

2. The method for suppressing clutter of ground penetrating radar based on graph signal processing according to claim 1, characterized in that: The graph theory-based Wiener filter variant is specifically: in, is the frequency domain vector of the filtered image; p x represents the power spectral density of the B-scan image set containing a mixture of targets and clutter, py represents the power spectral density of the B-scan image set containing only clutter, is the frequency domain vector of the image after Fourier transformation corresponding to the B-scan image to be suppressed.

3. The method for suppressing clutter of ground penetrating radar based on image signal processing according to claim 2, characterized in that: The power spectral density p x The calculation formula is: Among them, x r is a vector sample converted from a single image in a B-scan image set containing a mixture of targets and clutter; m is x r The number of samples; For x r The sample mean of ; diag is a diagonal matrix function; X = [x1, ..., x m ] covariance; V is the Fourier basis of the graph, that is The feature vector of .

4. The method for suppressing clutter of ground penetrating radar based on image signal processing according to claim 2, characterized in that: The power spectral density p y The calculation formula is: Among them, V is the graph Fourier basis, that is, The eigenvector of is the sample covariance matrix of the B-scan image sample matrix V containing only clutter, Y = [y1, ..., y n ], n is the number of B-scan image samples containing only clutter, y n It is a vector sample converted from a single image in the B-scan image set containing only clutter.

5. A method for suppressing clutter of a ground penetrating radar based on image signal processing according to claim 1 or 2, characterized in that: The construction process of the graph theory-based Wiener filter variant includes: Step S101, obtaining a B-scan image set of a ground penetrating radar under a test environment, and dividing it into a B-scan image set containing a mixture of targets and clutter and a B-scan image set containing only clutter; Step S102, respectively converting the B-type scan image containing the target and the clutter and the B-type scan image containing only the clutter into vectors, which are respectively recorded as the B-type scan image set and Where m is the number of B-scan images containing a mixture of targets and clutter, and n is the number of B-scan images containing only clutter; Step S103, calculate The covariance of calculate The covariance of Step S104: covariance Perform eigendecomposition to compute the graph Fourier basis V, compute the graph shift operator, and compute The power spectral density p x ; for covariance Perform static analysis. After the static analysis passes, calculate The power spectral density p y ; Step S104: constructing a graph-theoretic Wiener filter variant based on the power spectral density.

6. The method for suppressing clutter of ground penetrating radar based on image signal processing according to claim 5, characterized in that: When the static analysis of the B-scan image vector containing only clutter passes, the power spectrum density calculation of the clutter is performed in step S104.

7. The method for suppressing clutter of ground penetrating radar based on image signal processing according to claim 6, characterized in that: The process of static analysis includes: Defining Metrics where |||| F represents the Frobenius norm, which calculates the square root of the sum of the squares of the matrix elements; diag is the diagonal matrix function; When calculated When it is greater than 0.5, it indicates that the B-scan image vector containing only clutter is weakly static with respect to the image shift operator S, that is, the static analysis is passed.

8. The method for suppressing clutter of ground penetrating radar based on image signal processing according to claim 5, characterized in that: The graph shift operator is equal to , which is used to represent the graph Denoted as S, where is the number of nodes.

9. The method for suppressing clutter of ground penetrating radar based on image signal processing according to claim 5, characterized in that: Assume that a signal x is weakly static with respect to the graph shift operator S, and the graph shift operator S and the signal covariance C x have the same set of orthogonal eigenvectors, the covariance C x From the sample covariance Approximately, where V and Λ c They are The eigenvectors and eigenvalues ​​of for covariance of , m is the number of B-scan images containing a mixture of target and clutter.

10. The method for suppressing clutter of ground penetrating radar based on image signal processing according to claim 5, characterized in that: The automatic gain control method is used to divide the ground penetrating radar B-scan image. According to whether the B-scan image contains target features, if it contains target features, it is divided into a B-scan image containing a mixture of target and clutter; otherwise, it is divided into a B-scan image containing only clutter.

Images