Quantitative recognition method for subgrade hidden diseases based on ground penetrating radar

Abstract

The invention discloses a subgrade concealment disease quantitative identification method based on a ground penetrating radar, relates to the technical field of image processing, and aims to respectively establish ground penetrating radar forward simulation models for highway subgrade cavity, crack, looseness and void diseases, simulate the propagation process of radar emission electromagnetic waves in a medium, and identify the subgrade concealment disease. A B-scan radar image with clear disease types and sizes is obtained, a data set is obtained by expanding a simulation radar image, a high-quality standard data set of highway roadbed diseases is created, and quantitative recognition of the roadbed diseases is facilitated; according to the method, highway subgrade disease characteristics of different forms, different development degrees and different fillers are analyzed, the disease characteristics at least comprise cavity diseases, crack diseases, loosening diseases and void diseases, the relation between the preset disease size and the width and area in a GPR image is analyzed, and quantitative analysis of subgrade diseases is achieved. And intelligent recognition of highway subgrade diseases is realized by using a deep learning algorithm.

Description

Technical Field

[0001] The invention relates to the technical field of image processing, and in particular to a method for quantitatively identifying hidden roadbed diseases based on ground penetrating radar. Background Art

[0002] The challenges faced in the field of highway subgrade inspection are significant, so there is an urgent need to seek an efficient and accurate diagnostic method to intelligently detect the distribution and type of subgrade defects. The ground penetrating radar (GPR) nondestructive testing method has developed rapidly in recent years. It has the advantages of convenient detection, wide application range, and intuitive test results. It has become the main method for highway subgrade disease detection. Deep learning (DL)-based technology has been widely used in structural damage detection in civil engineering and has achieved excellent results. Manual processing of GPR images has also been gradually replaced by deep learning-based methods.

[0003] In actual engineering, underground media are distributed in three dimensions. After using GPRmax software to generate radar images of highway subgrade diseases and forming a high-quality standard data set of B-scan images of highway subgrade diseases with different forms, different development levels and different filling materials, simple two-dimensional B-scan images have limitations: they cannot fully reflect the characteristics of crack objects. For example, affected by the detection direction and image selection method, the key information in the B-scan image is sometimes difficult to be accurately identified, which may lead to erroneous detection results. By performing detailed feature analysis and quantitative evaluation of B-scan images, GPR data can be interpreted more accurately. In view of this, it is particularly necessary to analyze the characteristic laws of highway subgrade diseases reflected in ground penetrating radar signals and images, and to conduct quantitative research on the radar image characteristics of subgrade diseases. In addition, the use of continuously improving artificial intelligence technology can realize the intelligent identification of highway subgrade diseases, thereby reducing dependence on experience, reducing the time and economic cost required for detection, and improving the accuracy and efficiency of disease detection.

[0004] In summary, there is an urgent need for a quantitative identification method of hidden roadbed defects based on ground penetrating radar to solve the problems in the existing technology. Summary of the invention

[0005] In view of the above technical problems to be solved, the present invention provides a method for quantitatively identifying hidden roadbed diseases based on ground penetrating radar, which solves the problem in the prior art that simple two-dimensional B-scan images cannot fully reflect the characteristics of crack objects.

[0006] In order to solve the above technical problems, the technical solution proposed by the present invention is: A method for quantitatively identifying hidden roadbed diseases based on ground penetrating radar comprises the following steps: Step S1, constructing a data set: using a ground penetrating radar system to detect the actual roadbed, collecting ground penetrating radar B-scan actual image data, establishing a highway roadbed disease model, generating a ground penetrating radar B-scan simulation image, performing image processing and image annotation on the diseases therein, and forming a high-quality standard data set of highway roadbed B-scan images that clearly identify roadbed diseases; Step S2, quantitative analysis of radar images: analyzing the disease characteristics of the highway subgrade and the relationship between the preset disease size and the waveform width and area in the GPR image through the B-scan simulation image, wherein the disease characteristics at least include void disease, crack disease, loose disease, and void disease; Step S3, intelligent recognition of roadbed disease radar images: first build a highway roadbed disease detection model, including building a target detection neural network model with a detection frame and classification probability; then train the model, and adjust and configure the corresponding training parameters.

[0007] As a further improvement of the above technical solution: Preferably, in step S1, the process of generating a ground penetrating radar B-scan simulation image is as follows: firstly, a waveform of an A-scan signal and a B-scan image are generated, wherein the A-scan signal determines the location of the earthquake source through the ground penetrating radar and performs a forward calculation; the B-scan image is moved in a specific direction, and the transmitting and receiving sources move in a spatial position to generate forward simulation data of a certain section of the model; then, a forward simulation is performed on the type of damage in the roadbed to generate a ground penetrating radar B-scan simulation image.

[0008] Preferably, the image processing includes direct wave removal, noise superposition, brightness enhancement, contrast enhancement, image flipping, rotation, affine transformation, shear transformation, HSV data enhancement and translation expansion.

[0009] Preferably, the image annotation is to identify and select the target signal of the image, determine the position and type of each highway subgrade disease target signal, reflect it on the coordinates of the marking rectangular frame, and define the type information by the content of the mark.

[0010] Preferably, in step S2, there is a positive correlation between the cavity radius of the cavity disease and the reflection waveform area.

[0011] Preferably, in step S2, the amplitude of the reflected electromagnetic wave increases with the increase of the crack width in the crack disease.

[0012] Preferably, in step S2, the porosity disease is composed of a plurality of tiny pores, the range of reflected electromagnetic waves expands as the number of pores increases, and the coverage of the waveform in the radar image increases as the number of pores increases.

[0013] Preferably, in step S2, the reflected waveform area and the hollow defect size are positively correlated.

[0014] Preferably, in step S3, a disease detection module is constructed based on the YOLOv11 optimization algorithm, specifically: S3-1-1, integrate low-level features, high-level features and top-level features into the target detection process, extract features from the backbone network, input the features into AFPN for processing, and obtain features at different levels for fusion; S3-1-2, the backbone network combines triple attention to improve the C2PSA attention module; the triple attention consists of three parallel branches, two of which are responsible for capturing the cross-dimensional interaction between the channel dimension C and the spatial dimension H or W, and the other parallel branch is used to establish spatial attention and construct the interdependence between input channels or spatial positions.

[0015] Preferably, given an input tensor in the three parallel branches, the input tensor is first passed to the three branches, each module is designed to process a feature map C×H×W, where C is the number of channels, H is the height, and W is the width; the modules calculate the attention weights in different ways.

[0016] Compared with the prior art, the method for quantitatively identifying hidden roadbed defects based on ground penetrating radar provided by the present invention has the following advantages: The method for quantitatively identifying hidden roadbed defects based on ground penetrating radar of the present invention first constructs a high-quality standard data set for roadbed defects; then performs quantitative analysis on B-scan images; and then constructs a defect detection module based on the YOLOv11 optimization algorithm: AFPN is used to optimize the detection head, and low-level features, high-level features and top-level features are integrated into the target detection process. After extracting features from the backbone network, the features are input into AFPN for processing, and features of different levels are obtained for fusion, which can avoid large semantic gaps between non-adjacent Levels; C2PSA is combined with triple attention in the backbone network to improve the C2PSA attention module, and the triple attention consists of three parallel branches, two of which are responsible for capturing the cross-dimensional interaction between the channel C and the space H or W, and the last branch is used to construct the spatial attention. Finally, the outputs of the three branches are aggregated using average to construct the interdependence between the input channels or spatial positions, with low computational cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1is a schematic diagram of expanding the disease image database in an embodiment of the present invention; Figure 2 is a diagrammatic annotation of a disease image database in an embodiment of the present invention; Figure 3 is a comparison of radar images of holes of different radii on the highway roadbed in an embodiment of the present invention; Figure 4 is the regression relationship between the holes of different radii and the reflected waveform area of ​​the highway roadbed in the embodiment of the present invention; Figure 5 is a comparison of radar images of cracks of different widths on a highway roadbed in an embodiment of the present invention; Figure 6 The present invention is a comparison of radar images of different degrees of looseness of the highway roadbed in an embodiment of the present invention; Figure 7 It is a comparison of radar images of different sizes of highway roadbeds in an embodiment of the present invention; Figure 8 is a structural schematic diagram of a disease detection module in an embodiment of the present invention; Fig. 9 is an architectural diagram of AFPN in an embodiment of the present invention; Fig.10 is an architectural diagram of triple attention in an embodiment of the present invention; Fig.11 It is the training result of the target detection neural network model with detection box and classification probability in the embodiment of the present invention. DETAILED DESCRIPTION

[0018] The specific embodiments of the present invention are described in detail below. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0019] like Figure 1-Figure 10 As shown, the method for quantitatively identifying hidden roadbed defects based on ground penetrating radar of the present application comprises the following steps: Step S1, constructing a data set.

[0020] The ground penetrating radar system is used to detect the actual roadbed, and the actual ground penetrating radar B-scan image data is collected. A highway roadbed disease model is established, and the FDTD-based GPRmax software is used to generate ground penetrating radar B-scan simulation images. The images are then processed and labeled to form a high-quality standard dataset of highway roadbed B-scan images that clearly show roadbed diseases.

[0021] Specifically include the following: S1-1, establish a highway subgrade disease model.

[0022] In this embodiment, the roadbed disease model includes four disease models of voids, voids, looseness, and cracks, and the layered model parameters are set as shown in Table 1. According to the actual roadbed structure data, a roadbed forward simulation model is established, which is mainly divided into four layers: the first layer is the atmosphere layer, with a thickness of 0.3m and a relative dielectric constant of 1; the second layer is the surface layer, which mainly includes asphalt and cement, etc., with a thickness of 0.3m, a relative dielectric constant of 4, and a conductivity of 0.001S / m; the third layer is the road base, which uses materials including graded crushed stone, cement and lime-stabilized crushed stone, with a thickness of 0.4m, a relative dielectric constant of 9, and a conductivity of 0.05S / m; the bottom layer is the roadbed part, which is mainly composed of gravel, crushed stone, slag, and cement or lime-stabilized soil, with a thickness of 0.4m, a relative dielectric constant of 10 to 30, and a conductivity of 0.1S / m.

[0023] Table 1: Parameters of highway subgrade and pavement layer model

[0024] S1-2, generate a ground penetrating radar B-scan simulation image.

[0025] In this embodiment, the waveform of the A-scan signal and the B-scan image can be generated by the GPRmax software. The A-scan signal is a measurement of a single point. The positions of the two seismic sources are determined by the ground penetrating radar, and forward calculations are performed accordingly. The B-scan image involves line measurement. Along a specific direction, the transmitting and receiving sources move in a series of spatial positions, thereby generating forward simulation data for a certain section of the forward simulation model. Forward simulation is performed on the four main types of defects in the roadbed to generate ground penetrating radar B-scan simulation images.

[0026] The forward simulation model of the roadbed is established by forward simulation software. In this embodiment, GPRmax3.0 is used to perform ground penetrating radar forward numerical simulation. The main steps include: ① Establish the main space of the model and determine the spatial discretization and time window size; ② According to the size and burial depth of the shallow hidden disease model of the roadbed, select the appropriate wave source type, determine the spatial position and detection route of the electromagnetic wave transmitting and receiving sources; ③ Determine the time step and space step of the transmitting source and the receiving source; ④ Set the electrical parameters of the medium material of the shallow hidden disease model of the roadbed, establish the spatial model information of the roadbed disease, and give the boundary absorption conditions of the forward simulation; ⑤ Save the model parameter information in a file with an extension of .in. After simulation using the command prompt window, you will eventually get an electromagnetic wave simulation data file with extensions of .out and .vti.

[0027] The size of the GPR forward simulation model is set to 2.0m×6.0m×0.002m to simulate the propagation process of radar-transmitted electromagnetic waves in the medium. The radar excitation source is located on the surface of the model, and the free space above it is a 30cm air layer to reduce the interference of electromagnetic waves on detection. The specific parameters are shown in Table 2.

[0028] Table 2: GPRmax forward modeling parameters

[0029] S1-3, construct a high-quality standard dataset of B-scan images of highway subgrades with clear defects.

[0030] In this embodiment, the ground penetrating radar B-scan simulation image can clearly show the waveform of the defect, facilitate the analysis of the waveform characteristics of each defect, and facilitate the analysis of the relationship between the preset defect size and the width and area in the GPR image.

[0031] In this embodiment, the image processing is to perform operations such as removing direct waves, noise superposition, brightness enhancement, contrast enhancement, image flipping, rotation, affine transformation, shear transformation, hue-saturation-value (HSV) data enhancement and translation expansion on the simulated B-scan simulation image.

[0032] In this embodiment, the high-quality standard dataset of highway roadbed B-scan images is constructed according to the standard format of the YOLO dataset and is divided into a training set, a validation set, and a test set in a ratio of 7:2:1. In this embodiment, LabelImg software is preferably used to manually identify and select target signals in the image. LabelImg is a free open source image annotation tool suitable for image classification and target detection tasks.

[0033] In this embodiment, image annotation is to identify and select target signals in the image, determine the location and type of target signals of each highway roadbed disease, reflect them on the coordinates of the marked rectangular frame, and define the type information by the marked content. Use LabelImg software to manually identify and select target signals in radar images. LabelImg is a free open source image annotation tool suitable for image classification and target detection tasks.

[0034] In this embodiment, if Figure 1 As shown in the figure, the simulated radar images are subjected to operations such as removing direct waves, noise superposition, brightness enhancement, contrast enhancement, image flipping, rotation, affine transformation, shear transformation, hue-saturation-value (HSV) data enhancement and translation expansion. Finally, 1260 highway roadbed disease images are obtained to form a high-quality data set. The data set construction plan is shown in Table 3.

[0035] Table 3: Implementation plan for constructing highway subgrade disease dataset

[0036] In this embodiment, the simulated radar simulation image is marked with position and category, and the marking example is as follows: Figure 2 The red rectangle represents the void signal, the green rectangle represents the crack signal, the blue rectangle represents the loose signal, and the brown rectangle represents the void signal.

[0037] Step S2: quantitative analysis of radar images.

[0038] The B-scan simulation image is used to analyze the characteristics of road subgrade voids, cracks, looseness, and voids in the GPR image, as well as the relationship between the preset defect size and the waveform width and area in the GPR image. Specifically: 1) The reflection wave of the cavity disease presents a hyperbolic shape as a whole, with the opening downward. In order to realize the quantitative identification of the B-scan simulation image of the roadbed disease, the top of the waveform hyperbola is set to ( , ), the two symmetrical points at the bottom of the hyperbola are ( , )、( , ), the pixel area of ​​the reflection area of ​​the cavity disease image is obtained by formula (1), that is, the reflection waveform area S: (1) like Figure 3 Shows the comparison of radar images of cavity damage with different radii ( Figure 3 In the figure, from left to right are: the radius of the cavity is 0.05m, the radius is 0.1m, and the radius is 0.2m. In the forward simulation model of the roadbed cavity, the center position is kept unchanged, the medium type is the same, and the radius is 0.05m, 0.1m, and 0.2m respectively. Through analysis, it is found that when the cavity radius increases, the reflected waveform area also increases. This phenomenon verifies that there is a positive correlation between the cavity radius R and the reflected waveform area S. The regression relationship between different cavity radii and the reflected waveform area and the corresponding fitting relationship are shown in Figure 4 and formula (2):

[0039] (2)

[0040] 2) The shapes on both sides of the waveform of the crack disease are not curves, but closer to straight lines, and the electromagnetic wave signal presents a spike shape. When the crack is filled with water, its dielectric constant increases significantly, resulting in a significant increase in the intensity of the reflected wave in the forward simulation, and the waveform presents the characteristics of an isosceles triangle.

[0041] Figure 5 The radar waveforms of vertical cracks of different widths in radar images are shown. The cracks penetrate the base and the roadbed, and the medium is air. In order to distinguish them from rectangular voids, the width of the cracks is small, and two sizes of 3cm and 5cm are selected ( Figure 5 The middle picture (a) is a radar image of a crack with a width of 3 cm, and the picture (b) is a radar image of a crack with a width of 5 cm). Figure 5 It can be seen that as the crack width increases, the amplitude of the reflected electromagnetic wave also increases, and the downward waveform opening in the image also expands accordingly. The hyperbola at the top of the radar image is very sensitive to the change in crack width. At the same time, since the hyperbola characteristics at the bottom of the crack are obvious, the horizontal distance of the top hyperbola is selected as an indicator for quantitative analysis of crack width. The analysis results show that the horizontal distance at the top of the crack hyperbola increases with the increase of crack width.

[0042] 3) The loose disease is composed of many tiny pores. When the electromagnetic wave passes through the pavement layer and the base layer to reach the diseased area, it encounters different dielectric constant changes during the propagation due to multiple refractions and reflections. This causes the radar image of the loose disease of the roadbed to show a superposition of multiple curves, and the overall waveform appears complex and disordered.

[0043] The looseness in the roadbed is assumed to be composed of air as the medium. The radius of the tiny pores is set to 0.04m or 0.08m, and they are arranged in rows of 5 or 10 at a fixed center position, or 0.04m pores are arranged in rows of 5. The radar images of different degrees of looseness are shown in Figure 1. Figure 6 As shown. Figure 6 It can be seen that as the number of pores increases, the range of reflected electromagnetic waves expands, and the coverage of the waveform in the radar image also increases accordingly; in addition, as the pore radius increases, the opening of the downward waveform in the image will also become larger accordingly.

[0044] 4) In the hollowing disease image, a waveform is divided into two curves, the upper and lower parts, which differ in shape and energy distribution. The top of the upper curve is relatively flat, but as the distance from the starting point increases, the curve becomes more curved and gradually weakens. The lower curve presents a triangular outline, with concentrated energy at the top and more prominent in the image, and the curve decreases on both sides.

[0045] The reflection waveform of the hollowing disease generally shows a downward-opening hyperbola feature. Let the top of the hyperbola be ( , ), the two symmetrical points at the bottom of the hyperbola are ( , )、( , ), the pixel area S of the reflection area of ​​the void disease image is obtained by formula (1), and the void disease can be quantitatively analyzed. Figure 7 Comparison of radar images of air gap damage of different sizes ( Figure 7 (From left to right in the figure are: 25cm*4cm, 25cm*8cm, 50cm*4cm void radar images). In the forward simulation model of roadbed voids, under the same medium conditions, when the length and width increase, the reflected waveform area increases. This proves that the reflected waveform area is positively correlated with the void disease size.

[0046] Step S3: Intelligent recognition of roadbed damage radar images.

[0047] First, a highway subgrade disease detection model was constructed, including a target detection neural network model with a detection box and classification probability. Second, the 1,260 highway subgrade disease images in step S1 were randomly divided into a training set, a validation set, and a test set in a ratio of 7:2:1. The corresponding training parameters were adjusted and configured, including setting the number of training rounds to 200 to make it more suitable for the detection task. Specifically: S3-1, build disease detection module.

[0048] like Figure 8 As shown in the figure, a disease detection module is constructed based on the YOLOv11 optimization algorithm, specifically: S3-1-1, uses AFPN (Asymptotic Feature Pyramid Network) to optimize the detection head. By introducing an asymptotic feature fusion strategy, low-level features, high-level features, and top-level features are gradually integrated into the detection process of the detection head. This asymptotic fusion method helps to reduce the semantic gap between features at different levels, improve the feature fusion effect, and enable the detection model to better adapt to semantic information at different levels. After extracting features from the backbone network, the features are input into AFPN for processing, and then it can obtain features at different levels for fusion.

[0049] The architecture of AFPN is as follows Fig. 9 As shown in the figure, the black arrows represent convolution, the sea blue arrows represent adaptive spatial fusion, and the first and second layers from bottom to top are low-level features, the third layer is high-level features, and the fourth layer is top-level features. In the bottom-up feature extraction process of the backbone network, AFPN gradually integrates low-level features, high-level features, and top-level features by fusing two adjacent low-level features and gradually incorporating high-level features into the fusion process.

[0050] The architecture of AFPN is progressive, which will bring the semantic information of features at different levels closer in the process of progressive fusion. AFPN uses 1×1 convolution and bilinear interpolation methods to upsample features. On the other hand, downsampling is performed using different convolution kernels and strides according to the required downsampling rate, for example: 2×2 convolution with a stride of 2 is applied to achieve 2x downsampling; 4×4 convolution with a stride of 4 is applied to achieve 4x downsampling; and 8×8 convolution with a stride of 8 is applied to achieve 8x downsampling. After feature fusion, AFPN learns features using four residual units, each of which includes two 3×3 convolutions.

[0051] S3-1-2, in the backbone network part, the C2PSA attention module is improved by combining triplet attention. Triplet attention consists of three parallel branches, two of which are responsible for capturing the cross-dimensional interaction between the channel dimension C and the spatial dimension H or W, and the remaining last parallel branch is used to establish spatial attention, which can construct the interdependence between input channels or spatial positions, and has low computational cost.

[0052] Going a step further, the triple attention architecture is as follows Fig.10 As shown. The triple attention consists of three parallel branches, each of which is responsible for capturing the interaction features between the spatial dimension H or W and the channel dimension C. Given an input tensor , passing it to three parallel branches in the triple attention. Each module in the three parallel branches is designed to process feature maps (C×H×W), where C is the number of channels, H is the height, and W is the width. These modules calculate attention weights by interacting with different dimensions, enhancing the network's attention to important parts of the features, thereby improving performance in various visual tasks. Fig.10 In the example, the symbol ⊕ represents element-wise addition.

[0053] In the first parallel branch, an interaction is established between the H dimension and the C dimension. Input tensor Rotate 90° counterclockwise along the H axis. This rotation tensor is recorded as , whose shape is (W×H×C), and then the tensor after passing through the Z-Pool (virtual storage pool) The shape of is (2×H×C), and then passes through a standard convolutional layer with a kernel size of k×k×k, and then through a batch normalization layer, which provides an intermediate output of dimension (1×H×C). Finally, the tensor The resulting attention weight is generated by sigmoid (S-type activation function), which is rotated 90° clockwise along the H axis to keep the same shape as the input.

[0054] In the second parallel branch, an interaction is established between the C dimension and the W dimension. Input tensor Rotate 90° counterclockwise along the W axis. This rotation tensor is recorded as , whose shape is (H×C×W), and then the tensor after Z-Pool The shape of is (2×C×W), and then passes through a standard convolutional layer with a kernel size of k×k×k, and then through a batch normalization layer, which provides an intermediate output of dimension (1×C×W). Finally, the tensor The resulting attention weights are generated by sigmoid, which is rotated 90° clockwise along the W axis to keep the same shape as the input.

[0055] In the third parallel branch, an interaction is established between the H dimension and the W dimension. Input tensor The channels of the Z-pool are simplified to 2 variables. The tensor after Z-Pool The shape of is (2×H×W), and then passes through a standard convolution layer with a kernel size of k×k×k, and a batch normalization layer. The output passes through a sigmoid activation layer to generate an attention weight of shape (1×H×W), which is applied to the input , and get the result The refined tensors (C×H×W) produced by the three branches are then aggregated together by simple averaging.

[0056] The final output tensor: (3) Among them, σ represents sigmoid; ψ1, ψ2, and ψ3 represent standard 2D convolutional layers defined by the kernel size k in the three parallel branches of triple attention.

[0057] Finally, the results of the three parallel branches are average pooled and then aggregated to generate the final attention weights.

[0058] S3-2, uses the optimized YOLO model to perform intelligent detection on high-quality standard datasets.

[0059] In this embodiment, the experimental environment is run on a computer with an operating system of Windows 11, a processor of Intel i5-13500H, and a memory size of 32GB. The development tool is PyCharm, the development language is Python 3.11.9, and the deep learning framework is Pytorch 2.0.0. The specific experimental training parameters are shown in Table 4, and the training results of the target detection neural network model with detection boxes and classification probabilities are shown in Fig.11 .

[0060] Table 4: Experimental parameters

[0061] The above implementation cases are only preferred embodiments of the present invention and are not intended to limit the present invention in any form. Although the present invention has been disclosed as above with preferred embodiments, they are not intended to limit the present invention. Therefore, any simple modification, equivalent changes and modifications made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the scope of protection of the technical solution of the present invention.

Claims

1. A method for quantitatively identifying hidden roadbed defects based on ground penetrating radar, characterized in that: The specific steps include: Step S1, constructing a data set: using a ground penetrating radar system to detect the actual roadbed, collecting ground penetrating radar B-scan actual image data, establishing a highway roadbed disease model, generating a ground penetrating radar B-scan simulation image, performing image processing and image annotation on the diseases therein, and forming a high-quality standard data set of highway roadbed B-scan images that clearly identify roadbed diseases; Step S2, quantitative analysis of radar images: analyzing the disease characteristics of the highway subgrade and the relationship between the preset disease size and the waveform width and area in the GPR image through the B-scan simulation image, wherein the disease characteristics at least include void disease, crack disease, loose disease, and void disease; Step S3, intelligent recognition of roadbed disease radar images: first build a highway roadbed disease detection model, including building a target detection neural network model with a detection frame and classification probability; then train the model, and adjust and configure the corresponding training parameters.

2. The method for quantitatively identifying hidden roadbed defects based on ground penetrating radar according to claim 1 is characterized in that: In step S1, the process of generating the ground penetrating radar B-scan simulation image is as follows: firstly, the waveform of the A-scan signal and the B-scan image are generated, wherein the A-scan signal determines the location of the earthquake source through the ground penetrating radar and performs forward calculation; the B-scan image is moved in a specific direction, and the transmitting and receiving sources are moved in the spatial position to generate forward simulation data of a certain section of the model; then, the type of disease in the roadbed is forward simulated to generate the ground penetrating radar B-scan simulation image.

3. The method for quantitatively identifying hidden roadbed defects based on ground penetrating radar according to claim 1 is characterized in that: The image processing includes direct wave removal, noise superposition, brightness enhancement, contrast enhancement, image flipping, rotation, affine transformation, shear transformation, HSV data enhancement and translation expansion.

4. The method for quantitatively identifying hidden roadbed defects based on ground penetrating radar according to claim 1 is characterized in that: The image annotation is to identify and select the target signal of the image, determine the position and type of each highway roadbed disease target signal, reflect it on the coordinates of the marking rectangular frame, and define the type information by the marked content.

5. The method for quantitatively identifying hidden roadbed defects based on ground penetrating radar according to claim 1 is characterized in that: In step S2, there is a positive correlation between the cavity radius of the cavity disease and the reflection waveform area.

6. The method for quantitatively identifying hidden roadbed defects based on ground penetrating radar according to claim 1 is characterized in that: In the step S2, the amplitude of the reflected electromagnetic wave increases as the crack width in the crack disease increases.

7. The method for quantitatively identifying hidden roadbed defects based on ground penetrating radar according to claim 1 is characterized in that: In step S2, the porosity disease is composed of a plurality of tiny pores, the range of reflected electromagnetic waves expands as the number of pores increases, and the coverage range of the waveform in the radar image increases as the number of pores increases.

8. The method for quantitatively identifying hidden roadbed defects based on ground penetrating radar according to claim 1 is characterized in that: In step S2, the reflected waveform area and the size of the hollow defect are positively correlated.

9. The method for quantitatively identifying hidden roadbed defects based on ground penetrating radar according to claim 1 is characterized in that: In step S3, a disease detection module is constructed based on the YOLOv11 optimization algorithm, specifically: S3-1-1, integrate low-level features, high-level features and top-level features into the target detection process, extract features from the backbone network, input the features into AFPN for processing, and obtain features at different levels for fusion; S3-1-2, the backbone network combines triple attention to improve the C2PSA attention module; the triple attention consists of three parallel branches, two of which are responsible for capturing the cross-dimensional interaction between the channel dimension C and the spatial dimension H or W, and the other parallel branch is used to establish spatial attention and construct the interdependence between input channels or spatial positions.

10. The method for quantitatively identifying hidden roadbed defects based on ground penetrating radar according to claim 9, characterized in that: Given an input tensor in the three parallel branches, the input tensor is first passed to the three branches. Each module is designed to process a feature map C×H×W, where C is the number of channels, H is the height, and W is the width; the modules calculate the attention weights in different ways.

Images