A data visualization method for ground penetrating radar

Abstract

The present invention relates to a method for visualizing ground-penetrating radar (GPR) data, comprising the following steps: 1) establishing an underground medium model and performing GPR simulation to obtain multiple random simulation data of target parameters, thereby generating a corresponding B-scan image sample set; 2) performing GPR target detection on the B-scan image to obtain vertex coordinates of the target signal, extracting slices near the vertices, establishing corresponding data sets based on the material and dimensional parameter data corresponding to the slices, and respectively corresponding to pre-trained neural networks; 3) inputting the newly extracted slices into the pre-trained neural network, estimating the material and dimensional parameters of the target, and printing the corresponding parameter information on the corresponding vertices to complete data visualization. Compared with the prior art, the present invention adopts a GPR target detection method based on a row-linked clustering algorithm, which can more accurately detect, especially in complex situations where target signals intersect, and can identify GPR targets and perform visual display.

Description

Technical Field

[0001] The present invention relates to the field of ground penetrating radar target detection and target recognition, and in particular to a data visualization method for ground penetrating radar. Background Art

[0002] Ground-penetrating radar (GPR) is an important nondestructive assessment device. It uses the propagation and scattering characteristics of high-frequency electromagnetic (EM) waves and the varying electromagnetic properties of subsurface materials to detect subsurface areas. Originally used in the military industry, GPR has now become a ubiquitous tool in everyday life, such as engineering surveys. While GPR is highly attractive as a nondestructive subsurface exploration tool, the readability of GPR data is difficult for those without extensive prior knowledge. Manual interpretation by experts is far from timely and inefficient, and manual interpretation often relies on the expertise and experience of the interpreter, making it prone to misjudgments and unnecessary errors. Consequently, the need for GPR data visualization has become increasingly urgent. However, due to the physical properties of the subsurface and the complexity of electromagnetic wave propagation mechanisms, anomalies in the GPR scanning area are often difficult to automatically detect and identify.

[0003] In order to improve the readability of ground penetrating radar (GPR), data visualization methods for underground targets are of great significance in this research. Although existing methods can achieve the purpose of target detection and recognition for some simple targets, in the complex case of intersection of hyperbolic features, since general methods cannot effectively separate the intersection signals into independent target hyperbolic signals, direct curve fitting often leads to missed detection and false detection of targets, which greatly reduces the effect of GPR data visualization. Summary of the Invention

[0004] The purpose of the present invention is to provide a data visualization method for ground penetrating radar in order to overcome the defects of the above-mentioned prior art.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A method for visualizing ground penetrating radar data, comprising the following steps:

[0007] 1) Establish an underground medium model and perform ground penetrating radar simulation to obtain simulation data with random parameters of multiple targets, and generate corresponding B-scan image sample sets;

[0008] 2) Perform ground-penetrating radar target detection on the B-scan image to obtain the vertex coordinates of the target signal and extract slices near the vertex. Based on the material and size parameter data of the target corresponding to the slice, a corresponding data set is established and a corresponding pre-trained neural network is used;

[0009] 3) The newly extracted slices are input into the pre-trained neural network to estimate the material and size parameters of the target, and the corresponding parameter information is printed on the corresponding vertices to complete the data visualization.

[0010] In the step 1), a ground penetrating radar simulation method based on ray tracing is used to obtain simulation data with multiple random parameters, that is, a B-scan image sample set is generated by gprMax software.

[0011] In the step 2), the neural network is specifically ResNet, which is trained with slices near the vertex as input and the material and size parameters of the target corresponding to the target as output.

[0012] In the step 2), a ground penetrating radar target detection method based on a row connection clustering algorithm is used to implement ground penetrating radar target detection to obtain the target signal vertex coordinates.

[0013] The ground penetrating radar target detection method based on the row connection clustering algorithm specifically includes the following steps:

[0014] 21) Preprocessing each original B-scan image in the B-scan image sample set, including zero-time correction, ground clutter removal, amplitude enhancement, and adaptive binarization, to obtain two corresponding binary images;

[0015] 22) Target extraction is performed on each binary image. The intersecting hyperbolic target signals are separated using a row connection clustering algorithm. The separated hyperbolic signals are then fitted to obtain the target hyperbolic vertex position, completing target detection.

[0016] In the step 21), zero time is obtained according to the frequency of the radar signal to perform zero time correction;

[0017] Ground clutter removal is achieved by subtracting the mean value of each row, then:

[0018]

[0019] Where x(t) is the amplitude of a signal at time t, x j (t) is the amplitude of the jth B-scan signal at time t, n is the number of B-scan signals, and x1(t) is the amplitude of the B-scan signal at time t after ground clutter removal;

[0020] Considering the energy attenuation during the propagation process, a linear time-varying gain is used for amplitude enhancement, and then:

[0021]

[0022] Where x2(t) is the amplitude of the signal after amplitude enhancement at time t, t N is the total simulation time;

[0023] Adopting the adaptive threshold method for binarization, we have:

[0024] threshold=mean{I e |I e >ρ b ×Max Ie}

[0025] Among them, threshold is the threshold, mean{} is the function for calculating the average, I e is the intensity value of the edge pixel, is the maximum value of edge pixel intensity, ρ b A decimal number between 0 and 1.

[0026] In step 21), the amplitude-enhanced B-scan image is binarized twice to minimize information loss due to changes in reflection polarity. The first binarization is performed on the original amplitude-enhanced B-scan image, and the second binarization is performed on the inverted amplitude-enhanced B-scan image.

[0027] The step 22) specifically includes the following steps:

[0028] 221) Using cv2.findContours() function to separate potential blocks in the binary image, the potential blocks are defined as bright blocks that are not connected to each other in the binary image;

[0029] 222) A row connection clustering algorithm is used to separate the intersecting hyperbolas for the row segments in each potential block, wherein the row segments are defined as each set of continuous points in a row of the potential block, specifically including:

[0030] 2221) Determine whether the current row segment in the potential block meets the RCC condition, if so, proceed to step 2222), if not, proceed to step 2223);

[0031] 2222) extending and connecting the row fragment to the current target cluster;

[0032] 2223) determining the row segment as a new residual limb cluster, and transferring the determination target to the next target cluster, and returning to step 2221) to continue determining whether the RCC condition is met;

[0033] 2224) If the row segment is not connected to any existing target clusters, it is used as a new target cluster.

[0034] 223) Fitting is performed based on the separated hyperbola to obtain the vertex position coordinates of the corresponding hyperbola to complete the target detection.

[0035] The RCC condition is based on the geometric characteristics of hyperbolas and is used to determine whether a line segment is a new residual limb caused by the intersection of other hyperbolas. Specifically,

[0036] If the minimum column of the row segment is on the left side of the residual limb cluster to which it is connected, or the maximum column of the row segment is on the right side of the residual limb cluster to which it is connected, then the RCC condition is considered to be met. If the row segment does not meet the RCC condition, it is regarded as a new residual limb cluster.

[0037] In step 2), in order to further eliminate bad samples caused by false detection before target recognition, a positive and negative sample data set is established, and the neural network is trained after filtering out negative samples by training the positive and negative sample discrimination network.

[0038] Compared with the prior art, the present invention has the following advantages:

[0039] Ground-penetrating radar (GPR) is a sensor that uses electromagnetic waves to detect geological structures and / or underground targets. Interpreting GPR data is a challenging task. This paper proposes a full-process solution for GPR data interpretation and visualization, encompassing preprocessing, target detection, and target identification and classification. This solution enables the conversion of GPR data into target classification and parameter visualization. Furthermore, a row-link clustering algorithm is employed in target detection to improve target detection capabilities. Compared to existing column-link clustering algorithms, this paper directly captures hyperbolic vertices rather than feet, making it easy to separate overlapping hyperbolas, accurately identify targets, and intuitively display them. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is the effect diagram after target detection in an embodiment of the present invention. The pixels in the gray box will be input into the network for positive and negative sample discrimination, and only positive samples will be recognized by the target.

[0041] Figure 2 This is a rendering of the data visualization method for the ground penetrating radar of the present invention.

[0042] Figure 3 These are the relevant parameters of the materials involved in the model in the embodiment.

[0043] Figure 4 It is a flowchart of the method of the present invention. DETAILED DESCRIPTION

[0044] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0045] The present invention provides a data visualization method for ground penetrating radar, comprising the following steps:

[0046] 1) Establish an underground medium model and use the existing ray tracing-based ground-penetrating radar simulation method to obtain a large amount of simulation data with random parameters, laying the foundation for establishing a sample set for subsequent target recognition. That is, generate a B-scan image sample set using gprMax software;

[0047] 2) Ground-penetrating radar target detection is performed using a GPR target detection method based on a row-linked clustering algorithm to obtain the coordinates of the target signal vertices and extract slices near the vertices. Data sets are created based on the material and size of the corresponding targets, and neural networks (ResNet) are pre-trained for each. In fact, to further eliminate bad samples caused by false detection during the target detection step before target recognition, the present invention also establishes a positive and negative sample dataset for training a positive and negative sample discriminant network to filter out negative samples.

[0048] In the present invention, the specific implementation process of the ground penetrating radar target detection method based on the row connection clustering algorithm is as follows:

[0049] After converting the original B-SCAN image into two binary images, each binary image is sequentially subjected to target extraction.

[0050] First, the row connection clustering algorithm is used to separate all hyperbolic regions from each other. To better explain the row connection clustering algorithm, the present invention defines the following three concepts, including row segments, potential blocks, and RCC conditions, specifically:

[0051] Line segment: Each set of consecutive points in a line is called a line segment.

[0052] Latent blocks: There are some bright blocks in the binary image that are not connected to each other. These bright blocks are called latent blocks.

[0053] RCC condition: RCC is used to determine whether a row segment is a new residual limb caused by the intersection of other hyperbolas. This condition is proposed based on the geometric characteristics of the hyperbola. Specifically, if the minimum column of the row segment is on the left side of the cluster it is connected to, or the maximum column of the row segment is on the right side of the cluster it is connected to, then the RCC is considered to be satisfied. If the row segment does not satisfy the RCC, it is regarded as a new residual limb cluster.

[0054] The line segments in the potential block are then processed sequentially from top to bottom and from left to right. For each line segment, first determine whether it is connected to the current cluster. If it is connected, further determine whether it satisfies the RCC. If so, the line segment is expanded to the cluster; if it does not satisfy the RCC, it is determined to be a new residual cluster. If the line segment is not connected to the cluster, the above operation is performed on the next cluster until the last cluster is processed. If the line segment is not connected to any existing clusters, it is treated as a new cluster.

[0055] 3) Using the pre-trained neural network analysis and the GPR target detection method based on the row connection clustering algorithm, the newly extracted slices can estimate the material and size related parameters of the target, and print the relevant information on the relevant vertices of the GPR data after simple processing.

[0056] So far, the data visualization method of the ground penetrating radar involved in the present invention has been fully presented, thereby achieving a ground penetrating radar data visualization effect with a high recognition rate.

[0057] Example

[0058] The data visualization method of the ground penetrating radar provided in this embodiment includes the following steps:

[0059] 1) Build 850 models based on underground dielectric structures: The spatial dimensions of the model are 2 meters wide and 0.5 meters deep. The signal source is a Ricker pulse with a center frequency of 1.5 GHz. The background of each model is filled with soil. Three cylindrical objects are randomly placed in the soil. The underground objects have three material properties and three radius properties. Their selection is also random. The materials include perfect conductor (PEC), polyvinyl chloride (PVC) and rock. The size categories include 1 cm, 3 cm and 5 cm. The parameters of each material are as follows: Figure 3 shown.

[0060] 2) A ground-penetrating radar target detection method based on a row-linked clustering algorithm was used on 850 B-scan images generated from an underground medium structure model, generating 9,429 positive sample slices at the vertex of the hyperbola and 3,419 negative sample slices at the vertex of the hyperbola. Through data augmentation, the number of positive and negative samples was expanded to 10,000. The positive samples were then classified according to material and radius, and material and radius datasets were constructed. After these data were input into ResNet for training, the target detection accuracy reached 98.2%, and the average accuracy of material and radius recognition reached 97.6% and 92.2%, respectively.

[0061] 3) Using the pre-trained neural network to analyze the newly extracted slices using the GPR target detection method based on the row connection clustering algorithm, the material and size related parameters of the target can be estimated, and the relevant information can be printed on the relevant vertices of the GPR data after simple processing, such as Figure 2 As shown in the figure, several obvious hyperbolic features belong to the three targets in the figure, resulting in 8 judgment sentences. The second target has misjudgments in both material recognition and radius recognition. However, based on other judgment results of the same target, the above misjudgments can be easily discarded: the number of times it is judged to be the same category and its classification probability are both important references.

[0062] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A method for visualizing ground penetrating radar data, characterized in that: The following steps are involved: 1) Establish an underground medium model and perform ground penetrating radar simulation to obtain simulation data with random parameters of multiple targets, and generate corresponding B-scan image sample sets; 2) Perform ground-penetrating radar target detection on the B-scan image to obtain the vertex coordinates of the target signal and extract slices near the vertex. Based on the material and size parameter data of the target corresponding to the slice, a corresponding data set is established and a corresponding pre-trained neural network is used; 3) Input the newly extracted slices into the pre-trained neural network to estimate the material and size parameters of the target, and print the corresponding parameter information on the corresponding vertices to complete the data visualization; The ground penetrating radar target detection method based on row connection clustering algorithm is used to realize ground penetrating radar target detection and obtain the target signal vertex coordinates; The ground penetrating radar target detection method based on the row connection clustering algorithm specifically includes the following steps: 21) Preprocessing each original B-scan image in the B-scan image sample set, including zero-time correction, ground clutter removal, amplitude enhancement, and adaptive binarization, to obtain two corresponding binary images; 22) Target extraction is performed on each binary image. The intersecting hyperbolic target signals are separated using a row-connection clustering algorithm. The separated hyperbolic signals are then fitted to obtain the target hyperbolic vertex position, completing target detection. In the step 21), zero time is obtained according to the frequency of the radar signal to perform zero time correction; Ground clutter removal is achieved by subtracting the mean value of each row, then: Where x(t) is the amplitude of a signal at time t, x j (t) is the amplitude of the jth B-scan signal at time t, n is the number of B-scan signals, and x1(t) is the amplitude of the B-scan signal at time t after ground clutter removal; Considering the energy attenuation during the propagation process, a linear time-varying gain is used for amplitude enhancement, and then: Where x2(t) is the amplitude of the signal after amplitude enhancement at time t, t N is the total simulation time; Adopting the adaptive threshold method for binarization, we have: Among them, threshold is the threshold, mean{} is the function for calculating the average, I e is the intensity value of the edge pixel, is the maximum value of edge pixel intensity, ρ b is a decimal between 0 and 1; In step 21), the amplitude-enhanced B-scan image is binarized twice to minimize information loss due to changes in reflection polarity, wherein the first binarization process is performed on the original amplitude-enhanced B-scan image and the second binarization process is performed on the inverted amplitude-enhanced B-scan image; The step 22) specifically includes the following steps: 221) Using cv2.findContours() function to separate potential blocks in the binary image, the potential blocks are defined as bright blocks that are not connected to each other in the binary image; 222) A row connection clustering algorithm is used to separate the intersecting hyperbolas for the row segments in each potential block, wherein the row segments are defined as each set of continuous points in a row of the potential block, specifically including: 2221) Determine whether the current row segment in the potential block meets the RCC condition, if so, proceed to step 2222), if not, proceed to step 2223); 2222) extending and connecting the row fragment to the current target cluster; 2223) determining the row segment as a new residual limb cluster, and transferring the determination target to the next target cluster, and returning to step 2221) to continue determining whether the RCC condition is met; 2224) If the row segment is not connected to any existing target cluster, then take it as a new target cluster; 223) fitting the separated hyperbola to obtain the vertex position coordinates of the corresponding hyperbola to complete target detection; The RCC condition is based on the geometric characteristics of hyperbolas and is used to determine whether a line segment is a new residual limb caused by the intersection of other hyperbolas. Specifically: If the minimum column of the row segment is on the left side of the residual limb cluster to which it is connected, or the maximum column of the row segment is on the right side of the residual limb cluster to which it is connected, then the RCC condition is considered to be met. If the row segment does not meet the RCC condition, it is regarded as a new residual limb cluster.

2. The method for visualizing ground penetrating radar data according to claim 1, wherein: In the step 1), a ground penetrating radar simulation method based on ray tracing is used to obtain simulation data with multiple random parameters, that is, a B-scan image sample set is generated by gprMax software.

3. The method for visualizing ground penetrating radar data according to claim 1, wherein: In the step 2), the neural network is specifically ResNet, which is trained with slices near the vertex as input and the material and size parameters of the target corresponding to the target as output.

4. The method for visualizing ground penetrating radar data according to claim 1, wherein: In step 2), in order to further eliminate bad samples caused by false detection before target recognition, a positive and negative sample data set is established, and the neural network is trained after filtering out negative samples by training the positive and negative sample discrimination network.

Images