Ground penetrating radar clutter suppression method and system and medium

Abstract

The invention provides a ground penetrating radar clutter suppression method and system and a medium, the method highlights a texture change area through first-order difference, can accurately capture change information of echo intensity and suppress background fluctuation and noise, meanwhile, introduces graph regularization to ensure similarity of adjacent pixels, suppresses isolated noise points, and improves noise suppression efficiency. And the stability of the texture structure is enhanced. By means of the combination, clutters can be effectively removed when the method is used for processing complex backgrounds, and meanwhile key structure information is reserved. Complex echo texture data are projected to a singular value feature space through singular value decomposition, singular values with a large change rate are screened out for reconstruction, redundant interference can be effectively removed, and stable echo textures are reserved. According to the mechanism, when the method is used for processing multi-target, dispersed and irregular echo signals, target signals and clutters can be effectively separated, and the detection precision is influenced.

Description

Technical Field

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

[0002] Ground Penetrating Radar (GPR) is a geophysical exploration technology widely used in geological exploration, civil engineering, and environmental monitoring. It operates by emitting high-frequency electromagnetic waves and receiving reflected signals from the subsurface, thereby obtaining detailed information about underground structures. In recent years, GPR has been widely used to inspect the adhesion of building exterior insulation layers and to determine the distribution of rebar within walls and concrete. GPR accurately assesses the bond quality between the mortar and insulation board, and between the mortar and wall. It can promptly detect defects such as top voids between the mortar and insulation board, and base voids between the mortar and wall. This effectively prevents safety hazards such as insulation failure in exterior wall insulation systems, ensuring the proper use and durability of buildings. For steel bar distribution inspections within walls and concrete, precisely determining the location, spacing, and quantity of rebar helps assess the load-bearing capacity of building structures, providing critical information for the maintenance and reinforcement of existing buildings and quality control of new construction. Ground penetrating radar can provide non-destructive detection methods to help identify key information such as voids, cracks, and steel bar distribution.

[0003] However, during actual detection, ground-penetrating radar (GPR) echo signals are inevitably subject to interference from various types of clutter. This clutter, originating from a wide range of sources, significantly impacts detection accuracy and data availability. Firstly, direct coupling between the transmitting and receiving antennas generates clutter. This clutter is highly energetic and often manifests as strong reflections in the echo image, easily obscuring the target signal. Secondly, echoes reflected from the ground surface (i.e., surface clutter) and echoes reflected from interfaces with background media (i.e., underground clutter) can also be incorporated into the echo signal, severely impacting detection accuracy and reliability. When inspecting the adhesion of building exterior insulation, GPR can determine the adhesion status (e.g., proper adhesion versus gaps) between the insulation and the wall by analyzing echo signal intensity and texture variations. However, due to the small difference in dielectric constant between the insulation and the wall, the echo signal is weak and easily interfered with by clutter, resulting in inaccurate detection results. Furthermore, inspecting the insulation adhesion typically requires scanning a large area. The cumulative effect of clutter further reduces the signal-to-noise ratio and increases the difficulty of data analysis. In the detection of steel bar distribution in walls / concrete, ground penetrating radar can determine the location, depth and distribution of steel bars by identifying the reflected signals of steel bars.

[0004] However, the unevenness of concrete and the wall itself, as well as the complex reflection characteristics between rebar and concrete, lead to a large amount of clutter and noise in the echo signal. Clutter interference is particularly severe in areas with dense rebar, making it difficult to accurately extract the rebar reflection signal, affecting detection accuracy. Summary of the Invention

[0005] To address the existing problems, the present invention provides a ground-penetrating radar clutter suppression method, system, and medium. This approach addresses the existing technical issues of mixing a large amount of clutter and noise into echo signals due to the heterogeneity of concrete and wall materials, as well as the complex reflection characteristics between rebar and concrete. Clutter interference is particularly severe in areas with dense rebar density, making it difficult to accurately extract the rebar's reflected signal, thus affecting detection accuracy.

[0006] The present invention provides a ground penetrating radar clutter suppression method, comprising:

[0007] S1. Reshape the ground penetrating radar echo data according to the sampling frequency, perform first-order difference on the echo data to obtain a difference matrix and take the absolute value to obtain an absolute difference matrix, perform dimension completion on the absolute difference matrix, and calculate the image mean of the echo data;

[0008] S2. Calculate the standard deviation of the absolute difference matrix, design a first threshold based on the standard deviation, adjust the weighted image mean according to the first threshold, and generate a first data matrix;

[0009] S3. Construct a graph adjacency matrix, calculate a graph Laplacian matrix, and perform graph regularization on the first data matrix to obtain a second data matrix;

[0010] S4. Perform singular value decomposition on the second data matrix, perform first-order difference on the singular values, obtain the difference result and take the absolute value, design a second threshold based on the standard deviation of the absolute first-order difference, filter the index set based on the second threshold, and reconstruct and output the third data matrix based on the index set.

[0011] Optionally, reshaping the ground penetrating radar echo data according to the sampling frequency, performing first-order difference on the echo data to obtain a difference matrix and taking the absolute value to obtain an absolute difference matrix, performing dimension completion on the absolute difference matrix, and calculating the image mean of the echo data, includes:

[0012] The ground penetrating radar echo data data is reshaped into a two-dimensional matrix X according to the sampling frequency, which is expressed as:

[0013] X = reshape(data[m,n,1])

[0014] The dimension of the two-dimensional matrix X is (m, n), where m represents the number of sampling points and n represents the number of antenna positions. A first-order difference operation is performed on the two-dimensional matrix X to output the difference matrix ΔX∈R (m-1)×n , and take the absolute value of its difference result:

[0015] (|ΔX|) ij =|ΔX ij |=|X i+1，j -X i，j |,i=1,2,…,m-1; j=1,2,…,n

[0016] Fill the difference matrix with zeros to get the absolute difference matrix Its elements are:

[0017]

[0018] And calculate the mean of the two-dimensional matrix X according to the frequency dimension to obtain the mean vector

[0019]

[0020] Optionally, calculating a standard deviation of the absolute difference matrix and designing a first threshold based on the standard deviation, adjusting the weighted image mean according to the first threshold, and generating a first data matrix includes:

[0021] The standard deviation of the absolute difference matrix is ​​calculated, and the first threshold T is designed based on it, which is expressed as:

[0022]

[0023] Where α is the threshold weighting factor, σ(·) is the standard deviation, and the mask matrix M is created i，j ∈R m×n :

[0024]

[0025] Assume that the weighting coefficient is β, and replace the part of the absolute difference matrix that is smaller than the first threshold with the weighted mean to generate the first data matrix X1, which is expressed as:

[0026]

[0027] Optionally, constructing a graph adjacency matrix, calculating a graph Laplacian matrix, and performing graph regularization on the first data matrix to obtain a second data matrix includes:

[0028] Construct the graph adjacency matrix A ij , expressed as:

[0029]

[0030] Where n represents the number of columns of the first data matrix, and the degree matrix D∈R is calculated. m×n , whose diagonal elements are:

[0031]

[0032] Compute the graph Laplacian matrix, expressed as:

[0033] L=DA

[0034] Then, graph regularization is performed on the first data matrix to obtain a second data matrix X2, which is expressed as:

[0035] X2=X1-λ·X1·L

[0036] Here, λ represents the graph regularization parameter.

[0037] Optionally, performing singular value decomposition on the second data matrix, performing first-order differences on the singular values ​​to obtain difference results and taking absolute values, designing a second threshold based on a standard deviation of the absolute first-order differences, screening an index set using the second threshold, and reconstructing and outputting a third data matrix based on the index set, including:

[0038] The singular value decomposition of the second data matrix X2 is expressed as:

[0039] X2=USV T

[0040] Where U∈R mxr is a matrix The eigenvector of , V∈R nx r is a matrix The characteristic vector of S = diag(σ1, σ2, ...σ r ), r≤(min(M, N)) represents an M×N diagonal matrix composed of non-negative singular values ​​of a two-dimensional matrix, where all elements except the diagonal are 0 and the singular value σ i According to σ1≥σ2≥…≥σ r >0, sort in descending order, and then calculate the first-order difference of the singular value Δσ∈R r-1 , and take the absolute value of its difference result, which is expressed as:

[0041] (|Δσ|) k =|(Δσ) k |=|σ k+1 -σ k |,k=1,2,…,r-1

[0042] The second threshold is set according to the standard deviation of the absolute first-order differences, expressed as:

[0043] T σ =γ·σ(|Δσ|)

[0044] Among them, γ is the weighting factor, and the ones that satisfy (|Δσ|) are selected. k >T σ The index set I of the adjacent positions is retained to obtain the index set I′. The data is reconstructed using the index set I′ to obtain the third data matrix, which is expressed as:

[0045] X2=U I′ S I′ V I′ T

[0046] Among them, U I′ is the submatrix in U whose column index belongs to I′, S I′ is the submatrix in S whose row and column indices all belong to I′, V I′ is the submatrix in V whose column index belongs to I′.

[0047] The present invention also provides a ground penetrating radar clutter suppression system, comprising:

[0048] A differential processing module is used to reshape the ground penetrating radar echo data according to the sampling frequency, perform first-order difference on the echo data to obtain a difference matrix and take the absolute value to obtain an absolute difference matrix, perform dimension completion on the absolute difference matrix, and calculate the image mean of the echo data;

[0049] a threshold optimization module, configured to calculate a standard deviation of the absolute difference matrix, design a first threshold based on the standard deviation, adjust the weighted image mean according to the first threshold, and generate a first data matrix;

[0050] a graph regularization module, configured to construct a graph adjacency matrix, calculate a graph Laplacian matrix, and perform graph regularization on the first data matrix to obtain a second data matrix;

[0051] A singular value screening module is used to perform singular value decomposition on the second data matrix, perform first-order differences on the singular values, obtain the difference results and take the absolute values, design a second threshold based on the standard deviation of the absolute first-order differences, screen the index set using the second threshold, and reconstruct and output the third data matrix based on the index set.

[0052] The present invention also provides a computer-readable medium having executable instructions stored thereon, which, when executed by a processor, enables the processor to implement any of the aforementioned methods.

[0053] Compared with the prior art, the present invention:

[0054] First-order differences are used to accurately extract texture change information, highlight areas of echo intensity changes, and suppress background fluctuations; Laplace regularization of the nearest neighbor graph is used to enhance texture structural stability, effectively suppress clutter, and retain important structural information; singular value decomposition and screening are used to efficiently remove clutter, noise, and redundant information, highlighting key detection features; and this method has wide applicability and strong robustness, and can reliably suppress clutter in different detection scenarios and harsh environments. It performs well in detecting the bonding state of building exterior wall insulation layers and the distribution of steel bars in walls / concrete, can significantly improve image contrast and useful information content, effectively suppress complex clutter interference, and provide more reliable data support for subsequent detection and interpretation work. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0057] Figure 1 Schematic diagram of the method flow of the present invention;

[0058] Figure 2 A schematic diagram of an actual scenario in which the base layer between the bonding mortar and the wall is debonded in one embodiment of the present invention;

[0059] FIG3 is a schematic diagram showing a comparison of an original image B-scan echo map and a C-scan depth slice map in one embodiment of the present invention;

[0060] FIG4 is a schematic diagram showing a comparison of a B-scan echogram and a C-scan depth slice image screened in one embodiment of the present invention;

[0061] FIG5 is a schematic diagram of the distribution of steel bars at different positions in a ground penetrating radar concrete aggregate model simulated using gprMax in one embodiment of the present invention;

[0062] Figure 6 Schematic diagram of B-scan echo graph comparison after optimization of reinforcement simulation data in concrete aggregate according to one embodiment of the present invention;

[0063] Figure 7 Schematic diagram of echo diagram comparison after optimization of reinforcement simulation data in concrete aggregate according to one embodiment of the present invention. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other implementation cases obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. The functional units with the same labels in the examples of the present invention have the same and similar structures and functions.

[0065] See also Figure 1 The present invention provides a ground penetrating radar clutter suppression method, comprising:

[0066] S1. Reshape the ground penetrating radar echo data according to the sampling frequency, perform first-order difference on the echo data to obtain a difference matrix and take the absolute value to obtain an absolute difference matrix, perform dimension completion on the absolute difference matrix, and calculate the image mean of the echo data;

[0067] S2. Calculate the standard deviation of the absolute difference matrix, design a first threshold based on the standard deviation, adjust the weighted image mean according to the first threshold, and generate a first data matrix;

[0068] S3. Construct a graph adjacency matrix, calculate a graph Laplacian matrix, and perform graph regularization on the first data matrix to obtain a second data matrix;

[0069] S4. Perform singular value decomposition on the second data matrix, perform first-order difference on the singular values, obtain the difference result and take the absolute value, design a second threshold based on the standard deviation of the absolute first-order difference, filter the index set based on the second threshold, and reconstruct and output the third data matrix based on the index set.

[0070] In this embodiment, S1, the ground penetrating radar echo data is reshaped according to the sampling frequency, the first-order difference of the echo data is performed to obtain a difference matrix and the absolute value is taken to obtain an absolute difference matrix, the dimension of the absolute difference matrix is ​​completed, and the image mean of the echo data is calculated at the same time.

[0071] First, the GPR echo data data is reshaped into a two-dimensional matrix X according to the sampling frequency. The original GPR echo data is rearranged according to the frequency dimension to facilitate a more accurate analysis of the frequency variation characteristics of the data. The dimension of the reshaped two-dimensional matrix X is (m, n), where m represents the number of sampling points and n represents the number of antenna positions. The two-dimensional matrix X is expressed as:

[0072] X = reshape(Data[m,n,1])

[0073] Perform a first-order difference operation on the reshaped data along the frequency dimension (i.e., the first dimension of the two-dimensional matrix X) and output the difference matrix ΔX∈R (m-1)×n , and take the absolute value of its difference result:

[0074] (|ΔX|) ij =|ΔX ij |=|X i+1，j -X i，j |,i=1,2,…,m-1; j=1,2,…,n

[0075] In order to make |ΔX| match the dimension of the two-dimensional matrix X, a row of zeros is added in front of the first dimension to obtain the absolute difference matrix Its elements are:

[0076]

[0077] And calculate the mean of the two-dimensional matrix X according to the frequency dimension to obtain the mean vector

[0078]

[0079] First-order differences are performed on the GPR echo data, and the rate of change of the data is measured using the first-order derivative approximated by the first-order difference. In GPR echo data, different detection targets (such as areas with different bonding states in a building's exterior wall insulation layer, and areas between rebar in a wall or concrete and the surrounding medium) exhibit differences in echo intensity. These differences manifest as variations in echo intensity along the frequency dimension. By calculating the variations of adjacent data points along the frequency dimension through first-order differences, it is possible to accurately capture areas of drastic changes. For example, in a building's exterior wall insulation layer, the varying intensity values ​​of the hyperbolic texture changes between voided areas and properly bonded areas can be observed. Taking the absolute value operation further enhances these variations, thereby highlighting areas of texture change.

[0080] S2. Calculate the standard deviation of the absolute difference matrix and design a first threshold based on the standard deviation. Adjust the weighted image mean according to the first threshold to generate a first data matrix.

[0081] The standard deviation of the absolute difference matrix is ​​calculated and the first threshold T is designed based on it, which is expressed as:

[0082]

[0083] Among them, α is the threshold weighting factor set according to the actual data, which is used to adjust the threshold. It is generally taken as 0.8 to 2.4 in the building exterior wall insulation layer data, σ(·) is the standard deviation, and the mask matrix M is created. i，j ∈R m×n :

[0084]

[0085] Assume that the weighting coefficient is β, which is generally set to 0.4 to 0.8. Replace the part of the absolute difference matrix (change rate) that is less than the first threshold with the weighted mean to generate the first data matrix X1, which is expressed as:

[0086]

[0087] This step takes the image column mean Using the background as a reference, a threshold is set based on a weighted multiple of the standard deviation of the rate of change. Regions with low rates of change are then marked and replaced with the weighted mean, effectively suppressing clutter while preserving important information with significant texture variation. This approach suppresses weak, meaningless background clutter and noise fluctuations, preventing them from interfering with target recognition. It also filters out flat, featureless background textures in complex images, highlighting areas with prominent echo textures, such as those caused by insulation debonding and rebar embedded in walls or concrete. This improves image contrast and the proportion of effective information, laying the foundation for subsequent, precise analysis of target conditions.

[0088] S3. Construct a graph adjacency matrix, calculate the graph Laplacian matrix, and perform graph regularization on the first data matrix to obtain a second data matrix.

[0089] Let the number of columns n represent the number of data points to construct the graph adjacency matrix A ij , expressed as:

[0090]

[0091] Calculate the degree matrix D∈R m×n , which is a diagonal matrix whose diagonal elements are:

[0092]

[0093] Compute the graph Laplacian matrix, expressed as:

[0094] L=DA

[0095] Then, graph regularization is performed on the first data matrix to obtain a second data matrix X2, which is expressed as:

[0096] X2=X1-λ·X1·L

[0097] Here, λ represents the graph regularization parameter.

[0098] This step introduces spatial consistency in image processing by constructing a graph adjacency matrix and calculating a graph Laplacian matrix. This helps maintain the similarity of physically adjacent pixels in the image, thereby retaining important structural information while suppressing clutter. The graph regularization step adjusts the data based on the constructed graph adjacency matrix and Laplacian matrix. From the perspective of echo texture, since the echoes of adjacent detection points are physically correlated (affected by the uniformity of the medium and the law of propagation), this part strengthens the stability of the texture structure by constraining the data relationship of adjacent points. When there is noise interference or material inhomogeneity causing echo fluctuations, abnormal fluctuation points are suppressed to ensure that the texture remains coherent and consistent in spatial distribution, prevent isolated noise points from destroying the overall texture direction, make the image texture more consistent with physical reality, and improve the accuracy of target judgment.

[0099] S4. Perform singular value decomposition on the second data matrix, perform first-order difference on the singular values, obtain the difference result and take the absolute value, design a second threshold based on the standard deviation of the absolute first-order difference, filter the index set based on the second threshold, and reconstruct and output the third data matrix based on the index set.

[0100] The singular value decomposition of the second data matrix X2 after graph regularization is expressed as:

[0101] X2=USV T

[0102] Where U∈R mxr is a matrix The eigenvector of , V∈R nxr is a matrix The characteristic vector of S = diag(σ1, σ2, ...σ r ), r≤(min(M, N)) represents an M×N diagonal matrix composed of non-negative singular values ​​of a two-dimensional matrix, where all elements except the diagonal are 0 and the singular value σ i According to σ1≥σ2≥…≥σ r >0, sort in descending order, and then calculate the first-order difference of the singular value Δσ∈R r-1 , and take the absolute value of its difference result, which is expressed as:

[0103] (|Δσ|) k =|(Δσ) k |=|σ k+1 -σ k |,k=1,2,…,r-1

[0104] The second threshold is set according to the standard deviation of the absolute first-order differences, expressed as:

[0105] T σ =γ·σ(|Δσ|)

[0106] Among them, γ is a weighting factor, which is generally 0.6 to 0.8, and the ones that meet (|Δσ|) are selected. k >T σ The index set I of the adjacent positions is retained, and the retained singular values ​​and corresponding singular vectors are used to reconstruct the data to obtain the index set I′. The data is reconstructed using the index set I′ to obtain the third data matrix, which is expressed as:

[0107] X2=U I′ S I′ V I′ T

[0108] Among them, U I′ is the submatrix in U whose column index belongs to I′, S I′ is the submatrix in S whose row and column indices all belong to I′, V I′ is the submatrix in V whose column index belongs to I′.

[0109] In this step, singular value decomposition projects complex echo texture data into a singular value feature space, ranking each component by importance, based on the size of the singular values. In the B-scan image, large singular values ​​correspond to the main image structure and key bonding features (such as the common echo characteristics of large, stable bonding areas and the echo of rebar in walls / concrete), while small singular values ​​contain clutter or minor texture. By taking derivatives of the singular values ​​and reconstructing data with large differences between adjacent derivatives, the stable echo texture representing the key target information can be effectively separated from the minor texture reflecting shallow interference or minor medium differences. Reconstructing only the key characteristic components effectively removes clutter, noise, and redundant information while preserving the primary structure of the image, making the target-related texture features more prominent and easily discernible, thereby optimizing image entropy and information quality.

[0110] Finally, since qualitative visual inspections are difficult to fully evaluate the performance of the algorithm, quantitative evaluation indicators are needed to verify the effectiveness of clutter suppression. In this invention, target image entropy is used as a key indicator to measure the effectiveness of clutter suppression. Target image entropy reflects the richness and degree of confusion of information in the image by calculating the information content of the entire B-scan image. The value of image entropy describes the average number of bits in the image grayscale set, usually measured in bits / pixel. It not only measures the richness of the average information content in the image, but also reveals the concentration or dispersion characteristics of the image grayscale distribution. In clutter suppression, the level of target image entropy can indirectly reflect the effectiveness of the algorithm. Generally speaking, the lower the image entropy value, the more redundant information and noise in the image are effectively removed, the clutter interference is significantly reduced, and the image quality is improved. Conversely, a higher image entropy value means that there is more noise and clutter in the image, and the information distribution is more chaotic. Therefore, target image entropy can be used as an important indicator for evaluating image quality. During the image processing process, if the image entropy value remains within a reasonable range, it means that the effective information in the image is preserved, and the image clarity and target recognizability are high. Image entropy definition:

[0111]

[0112] Where p k is the probability of the k-th pixel, and L is the total number of gray levels in the image. However, when processing large-scale image data, this formula faces the problem of high computational complexity, especially when encountering the case where the pixel-level probability is zero, the formula cannot be directly applied. To solve this problem, an approximate calculation formula is used to ensure the effectiveness and feasibility of the calculation:

[0113]

[0114] The method first reshapes the GPR echo data according to the sampling frequency, performs a first-order difference on the data and takes its absolute value to highlight areas of texture variation and calculate the image mean. Next, a weighted threshold is set based on the rate of change, replacing regions with low rates of change with the weighted image mean to suppress meaningless background clutter and noise fluctuations. A graph adjacency matrix is ​​then constructed, and the graph Laplacian matrix is ​​calculated. Graph regularization is then applied to the data to suppress clutter, enhance texture structural stability, and ensure the similarity of adjacent pixels. Finally, singular value decomposition (SVD) is performed on the graph-regularized data to filter out singular values ​​with large rates of change and reconstruct the data, resulting in a clutter-suppressed image. This method effectively suppresses complex clutter interference in GPR echoes, such as direct waves, multiple reflections, and background noise, while accurately extracting key texture features of the target signal.

[0115] In another embodiment, to verify the effectiveness of the GPR clutter suppression method based on differential neighbor graph constraints and singular value screening proposed in the present invention, the method was applied to the processing of measured data of insulation layer bonding mortar using stepped frequency GPR and the processing of simulated data of concrete / wall aggregate reinforcement:

[0116] To simulate a scenario, see Figure 2 Figure 1. Actual scenario setup for bonding mortar with a gap between the base and the wall. "-2" and "-4" indicate the gap depth, while the remaining unmarked areas indicate normal bonding. Arrows indicate the direction of the B-scan lines. Six bonding mortars were tested. The experimental scenario included: insulation board thickness of 4.5 cm, mortar thickness of 1 cm, and random area. The frequency band was 4 GHz to 8 GHz, with 401 frequency points and a 10 kHz IF filter. Antenna spacing was 7.2 cm, with a 2 cm null coupling. The Y-axis point spacing was 3.5 cm, and the X-axis line spacing was 1 cm (scanning area 900 x 455 mm).

[0117] For comparison and implementation, first refer to Figure 3(a1), where Figure 3(a) is an original image B-scan echo map and C-scan depth slice map of a stepped frequency ground penetrating radar scanning band of 4-8 GHz, 401 sampling points, and no clutter suppression algorithm is used in one embodiment; Figure 3(b) is a B-scan echo map and C-scan depth slice map after using the mean cancellation method; Figure 3(c) is a B-scan echo map and C-scan depth slice map after using the SVD (singular value decomposition) method; Figure 3(d) is a B-scan echo map and C-scan depth slice map after using the PCA (principal component analysis) method based on SVD (singular value decomposition); Figure 3(e) is a B-scan echo map and C-scan depth slice map after using the local energy ratio 75% automatic selection method based on PCA; Figure 3(f) is a B-scan echo map and C-scan depth slice map after using the global energy ratio 99% automatic selection method based on PCA Figure 3(g) shows the B-scan echo and C-scan depth slices after using the ICA (independent component analysis) algorithm; Figure 3(h) shows the B-scan echo and C-scan depth slices after using the NMF (non-negative matrix factorization) algorithm; Figure 3(i) shows the B-scan echo and C-scan depth slices after using the RNMF (regularized non-negative matrix factorization) algorithm; Figure 3(j) shows the B-scan echo and C-scan depth slices after using the RPCA (robust principal component analysis) algorithm; Figure 3(k) shows the B-scan echo and C-scan depth slices after using the total variation denoising algorithm; Figure 3(l) shows the B-scan echo and C-scan depth slices after using the ground penetrating radar clutter suppression method based on differential neighbor graph constraints and singular value screening.

[0118] See Figure 4(a), which is a B-scan echo image and C-scan depth slice image in an ablation experiment using only graph regularization and singular value difference screening in one embodiment; Figure 4(b)(b) is a B-scan echo image and C-scan depth slice image in an ablation experiment using only first-order difference and singular value difference screening; Figure 4(c) is a B-scan echo image and C-scan depth slice image in an ablation experiment using only first-order difference and graph regularization.

[0119] 5( a ) shows the distribution of six steel bars at different positions in a GPR concrete (relative dielectric constant of 5) aggregate model simulated using gprMax in one embodiment. There are three types of aggregates in the concrete: quartz sand impurities (relative dielectric constant of 4, represented by a cuboid in the model), iron oxide impurities (relative dielectric constant of 4, represented by a sphere in the model), and mineral impurities (relative dielectric constant of 4, represented by a cylinder in the model). The positions of the impurities are randomly and evenly distributed. FIG5( b ) shows the distribution of six steel bars at different positions in an GPR wall (relative dielectric constant of 8) aggregate model simulated using gprMax. There are three types of aggregates in the concrete: quartz sand impurities (relative dielectric constant of 4, represented by a cuboid in the model), iron oxide impurities (relative dielectric constant of 4, represented by a sphere in the model), and mineral impurities (relative dielectric constant of 4, represented by a cylinder in the model). The positions of the impurities are randomly and evenly distributed.

[0120] See also Figure 6 middle Figure 6 (a) shows the B-scan echogram of the simulation data of steel bars in concrete aggregate in one embodiment, with a stepped frequency scan frequency range of 4-8 GHz and 41 sampling points, using the mean cancellation method; Figure 6 (b) is the B-scan echogram after using the SVD (singular value decomposition) method; Figure 6 (c) is the B-scan echogram after using the PCA (principal component analysis) method based on SVD (singular value decomposition); Figure 6 (d) is the B-scan echogram after using the PCA-based local energy ratio 75% automatic selection method; Figure 6 (e) is the B-scan echogram after using the PCA-based global energy ratio 99% automatic selection method; Figure 6 (f) is the B-scan echogram after using the ICA (independent component analysis) algorithm; Figure 6 (g) is the B-scan echo image after using the NMF (non-negative matrix factorization) algorithm; Figure 6 (h) is the B-scan echo image after using the RNMF (regularized non-negative matrix factorization) algorithm; Figure 6(i) is the B-scan echogram after using the RPCA (Robust Principal Component Analysis) algorithm; Figure 6 (j) is the B-scan echo image after using the total variation denoising algorithm; Figure 6 (k) is the B-scan echo image after using the ground penetrating radar clutter suppression method based on differential neighbor graph constraints and singular value screening.

[0121] See also Figure 7 middle Figure 7 (a) shows the B-scan echogram of the simulation data of steel bars in wall aggregate in one embodiment, with a stepped frequency scan frequency band of 4-8 GHz and 41 sampling points, using the mean cancellation method; Figure 7 (b) is the B-scan echogram after using the SVD (singular value decomposition) method; Figure 7 (c) is the B-scan echogram after using the PCA (principal component analysis) method based on SVD (singular value decomposition); Figure 7 (d) is the B-scan echogram after using the PCA-based local energy ratio 75% automatic selection method; Figure 7 (e) is the B-scan echogram after using the PCA-based global energy ratio 99% automatic selection method; Figure 7 (f) is the B-scan echogram after using the ICA (independent component analysis) algorithm; Figure 7 (g) is the B-scan echo image after using the NMF (non-negative matrix factorization) algorithm; Figure 7 (h) is the B-scan echo image after using the RNMF (regularized non-negative matrix factorization) algorithm; Figure 7 (i) is the B-scan echogram after using the RPCA (Robust Principal Component Analysis) algorithm; Figure 7 (j) is the B-scan echo image after using the total variation denoising algorithm; Figure 7 (k) is the B-scan echo image after using the ground penetrating radar clutter suppression method based on differential neighbor graph constraints and singular value screening.

[0122] Table 1(a) shows the quantitative comparison experimental results of different clutter suppression algorithms for building exterior wall insulation data in one embodiment; Table 1(b) shows the ablation experimental results of different clutter suppression algorithms for building exterior wall insulation data:

[0123] Table 1(a)

[0124]

[0125] Table 1(b)

[0126]

[0127] Table 2 shows the quantitative comparison test results of different clutter suppression algorithms for concrete aggregate and reinforcement simulation model data in one embodiment:

[0128] Table 2

[0129]

[0130] Table 3 shows the quantitative comparison experimental results of different clutter suppression algorithms for wall aggregate reinforcement simulation model data in one embodiment:

[0131] Table 3

[0132]

[0133] Through the above experimental comparisons, the GPR clutter suppression method proposed in this paper, based on differential neighbor graph constraints and singular value screening, demonstrates significant clutter suppression effectiveness in multiple scenarios. In B-scan images, while traditional clutter suppression methods (such as mean cancellation, SVD, and PCA) can reduce clutter around the target to a certain extent, they often fail to completely remove the uneven background interference. This interference can affect the determination of hyperbolic features, leading to inaccurate subsequent image interpretation. The method proposed in this paper, by combining first-order differences with graph regularization, effectively captures changes in echo intensity and, after singular value decomposition, selects key singular values ​​for reconstruction, resulting in clearer target echoes. The visualization results show a significant reduction in background clutter and a sharper target hyperbolic feature. As shown in Figure 3, in the detection of the bonding state of building exterior wall insulation, traditional clutter suppression methods often fail to completely remove background clutter and noise when processing gapped and properly bonded areas. This results in a lack of clear contrast between the echo features of the gapped areas and the normally bonded areas, making it impossible to distinguish between normal and gapped areas. After being processed by the ground penetrating radar clutter suppression method based on differential neighbor graph constraints and singular value screening, the background clutter can be effectively suppressed, making the echo characteristics of the void area more prominent and the background fluctuations greatly reduced. It can be clearly seen that the texture energy of the first two base void echoes in the X direction is stronger, and the texture energy characteristics of the third normal bonding echo are slightly weaker. At the same time, as shown in Figure 4, it can be seen that in the ablation experiment, only by combining first-order differences, graph regularization, and singular value decomposition with screening can the best effect be achieved. In the results of Figure 5, in the detection of steel bar distribution in concrete and walls, the traditional clutter suppression method has difficulty in effectively separating the steel bar reflection signal from the background and aggregate clutter when processing dense steel bar areas containing aggregates, resulting in interference with the position and distribution of the steel bars and making it difficult to accurately identify. After the clutter suppression of the present invention, the reflected echoes generated by the background and aggregates can be effectively removed, the echo texture is stabilized, and the steel bar reflection signal is clearer, which facilitates the subsequent steel bar position and distribution analysis.

[0134] Combined with Tables 1-3, the quantitative results of image entropy show that the method proposed in this paper exhibits the lowest image entropy values ​​in multiple experimental scenarios, indicating that this method can effectively retain key information in the image while removing clutter, significantly improving image clarity and target recognizability. In the building exterior wall insulation layer data, the image entropy can reach 17.795, far lower than other methods such as mean cancellation (27.809), SVD (24.179), and SVD-based PCA (28.442). In the concrete aggregate and steel bar simulation model, the image entropy of the method proposed in this paper is 33.602, which is lower than the mean cancellation method (36.173) and SVD (35.629). In the wall aggregate and steel bar simulation model, the image entropy of the method proposed in this paper is 33.102, which is also lower than the mean cancellation method (35.446) and SVD (34.911).

[0135] It can be seen that the reason why the method proposed in the present invention can show excellent clutter suppression effect in multiple scenarios is mainly due to the combination of first-order difference, graph regularization and singular value decomposition and screening mechanism. By highlighting the texture change area through first-order difference, it is possible to accurately capture the change information of echo intensity and suppress background fluctuations and noise. At the same time, the introduction of graph regularization ensures the similarity of adjacent pixels, suppresses isolated noise points, and enhances the stability of texture structure. This combination enables the method to effectively remove clutter while retaining key structural information when dealing with complex backgrounds. Complex echo texture data is projected into the singular value feature space through singular value decomposition, and singular values ​​with a larger change rate are screened out for reconstruction, which can effectively remove redundant interference and retain stable echo texture. This mechanism enables the method to effectively separate target signals from clutter when dealing with multi-target, scattered and irregular echo signals, thereby improving the overall detection accuracy.

[0136] The present invention also provides a ground penetrating radar clutter suppression system, comprising:

[0137] A differential processing module is used to reshape the ground penetrating radar echo data according to the sampling frequency, perform first-order difference on the echo data to obtain a difference matrix and take the absolute value to obtain an absolute difference matrix, perform dimension completion on the absolute difference matrix, and calculate the image mean of the echo data;

[0138] a threshold optimization module, configured to calculate a standard deviation of the absolute difference matrix, design a first threshold based on the standard deviation, adjust the weighted image mean according to the first threshold, and generate a first data matrix;

[0139] a graph regularization module, configured to construct a graph adjacency matrix, calculate a graph Laplacian matrix, and perform graph regularization on the first data matrix to obtain a second data matrix;

[0140] A singular value screening module is used to perform singular value decomposition on the second data matrix, perform first-order differences on the singular values, obtain the difference results and take the absolute values, design a second threshold based on the standard deviation of the absolute first-order differences, screen the index set using the second threshold, and reconstruct and output the third data matrix based on the index set.

[0141] The present invention also provides a computer-readable medium having executable instructions stored thereon, which, when executed by a processor, enables the processor to implement any of the aforementioned methods.

[0142] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0143] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is intended to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for suppressing clutter in a ground penetrating radar, characterized in that: include: S1. Reshape the ground penetrating radar echo data according to the sampling frequency, perform first-order difference on the echo data to obtain a difference matrix and take the absolute value to obtain an absolute difference matrix, perform dimension completion on the absolute difference matrix, and calculate the image mean of the echo data; S2. Calculate the standard deviation of the absolute difference matrix, design a first threshold based on the standard deviation, adjust the weighted image mean according to the first threshold, and generate a first data matrix; S3. Construct a graph adjacency matrix, calculate a graph Laplacian matrix, and perform graph regularization on the first data matrix to obtain a second data matrix; S4. Perform singular value decomposition on the second data matrix, perform first-order difference on the singular values, obtain the difference result and take the absolute value, design a second threshold based on the standard deviation of the absolute first-order difference, filter the index set based on the second threshold, and reconstruct and output the third data matrix based on the index set.

2. The method for suppressing clutter of a ground penetrating radar according to claim 1, wherein: The method of reshaping the ground penetrating radar echo data according to the sampling frequency, performing first-order difference on the echo data to obtain a difference matrix and taking the absolute value to obtain an absolute difference matrix, performing dimension completion on the absolute difference matrix, and calculating the image mean of the echo data includes: The ground penetrating radar echo data data is reshaped into a two-dimensional matrix X according to the sampling frequency, which is expressed as: X = reshape(data[m,n,1]) The dimension of the two-dimensional matrix X is (m, n), where m represents the number of sampling points and n represents the number of antenna positions. A first-order difference operation is performed on the two-dimensional matrix X to output the difference matrix ΔX∈R (m-1)×n , and take the absolute value of its difference result: (|ΔX|) ij =|ΔX ij |=|X i+1，j -X i，j |,i=1,2,…,m-1;j=1,2,…,n Fill the difference matrix with zeros to get the absolute difference matrix Its elements are: And calculate the mean of the two-dimensional matrix X according to the frequency dimension to obtain the mean vector 3. The method for suppressing clutter of a ground penetrating radar according to claim 2, wherein: The standard deviation of the absolute difference matrix is ​​calculated, and a first threshold is designed based on the standard deviation, and the weighted image mean is adjusted according to the first threshold to generate a first data matrix, including: The standard deviation of the absolute difference matrix is ​​calculated and the first threshold T is designed based on it, which is expressed as: Where α is the threshold weighting factor, σ(·) is the standard deviation, and the mask matrix M is created i，j ∈R m×n : Assume that the weighting coefficient is β, and replace the part of the absolute difference matrix that is smaller than the first threshold with the weighted mean to generate the first data matrix X1, which is expressed as:

4. The method for suppressing clutter of a ground penetrating radar according to claim 3, wherein: The step of constructing a graph adjacency matrix, calculating a graph Laplacian matrix, and performing graph regularization on the first data matrix to obtain a second data matrix includes: Construct the graph adjacency matrix A ij , expressed as: Where n represents the number of columns of the first data matrix, and the degree matrix D∈R is calculated. m×n , whose diagonal elements are: Compute the graph Laplacian matrix, expressed as: L=DA Then, graph regularization is performed on the first data matrix to obtain a second data matrix X2, which is expressed as: X2=X1-λ·X1·L Here, λ represents the graph regularization parameter.

5. The method for suppressing clutter of a ground penetrating radar according to claim 4, wherein: Performing singular value decomposition on the second data matrix, performing first-order differences on the singular values, obtaining difference results and taking absolute values, designing a second threshold value based on a standard deviation of the absolute first-order differences, screening an index set based on the second threshold value, and reconstructing and outputting a third data matrix based on the index set, including: The singular value decomposition of the second data matrix X2 is expressed as: X2=USV T Where U∈R mxr is a matrix The eigenvector of , V∈R nxr is a matrix The characteristic vector of S = diag(σ1, σ2, ...σ r ), r≤(min(M, N)) represents an M×N diagonal matrix composed of non-negative singular values ​​of a two-dimensional matrix, where all elements except the diagonal are 0 and the singular value σ i According to σ1≥σ2≥…≥σ r >0, sort in descending order, and then calculate the first-order difference of the singular value Δσ∈R r-1 , and take the absolute value of its difference result, which is expressed as: (|Ds|) k =|(Ds) k |=|s k+1 -s k |,k=1,2,…,r-1 The second threshold is set according to the standard deviation of the absolute first-order differences, expressed as: T σ =γ·σ(|Δσ|) Among them, γ is the weighting factor, and the ones that satisfy (|Δσ|) are selected. k >T σ The index set I of the adjacent positions is retained to obtain the index set I′. The data is reconstructed using the index set I′ to obtain the third data matrix, which is expressed as: X2=U I′ S I′ V I′ T Among them, U I′ is the submatrix in U whose column index belongs to I′, S I′ is the submatrix in S whose row and column indices all belong to I′, V I′ is the submatrix in V whose column index belongs to I′.

6. A ground penetrating radar clutter suppression system, characterized in that: include: A differential processing module is used to reshape the ground penetrating radar echo data according to the sampling frequency, perform first-order difference on the echo data to obtain a difference matrix and take the absolute value to obtain an absolute difference matrix, perform dimension completion on the absolute difference matrix, and calculate the image mean of the echo data; a threshold optimization module, configured to calculate a standard deviation of the absolute difference matrix, design a first threshold based on the standard deviation, adjust the weighted image mean according to the first threshold, and generate a first data matrix; a graph regularization module, configured to construct a graph adjacency matrix, calculate a graph Laplacian matrix, and perform graph regularization on the first data matrix to obtain a second data matrix; A singular value screening module is used to perform singular value decomposition on the second data matrix, perform first-order differences on the singular values, obtain the difference results and take the absolute values, design a second threshold based on the standard deviation of the absolute first-order differences, screen the index set using the second threshold, and reconstruct and output the third data matrix based on the index set.

7. A computer-readable medium having executable instructions stored thereon, which, when executed by a processor, causes the processor to implement the method according to any one of claims 1 to 5.

Images