GPR image recognition and three-dimensional reconstruction method for loose disease of semi-rigid base

Abstract

The invention is suitable for the technical field of road detection, and relates to a semi-rigid base layer loose disease GPR image recognition and three-dimensional reconstruction method, which comprises the following steps: constructing a semi-rigid base layer loose disease random medium model, carrying out forward modeling on a road structure, generating simulation data, and expanding a data set in combination with a data enhancement technology; a DN module fusing a dynamic deformation convolution module and a RepNCSPELAN4 module is designed, meanwhile, a WIoU loss function is introduced, training is conducted through an obtained data set, and identification of the semi-rigid base layer loose disease is achieved; calculating the area of a disease region through pixel and physical size conversion based on an instance segmentation result obtained through identification; a three-dimensional point cloud is generated by adopting inverse distance weighted interpolation, and the disease volume is calculated in combination with a Delaunay triangulation method. According to the method, accurate disease recognition is achieved, meanwhile, two-dimensional recognition is converted into three-dimensional reconstruction, and a quantitative basis is provided for maintenance priority division and repair scheme formulation.

Description

Technical Field

[0001] This invention belongs to the field of road inspection technology, and in particular relates to a method for GPR image recognition and three-dimensional reconstruction of semi-rigid base loose disease. Background Technology

[0002] Semi-rigid base courses are widely used in road engineering due to their high strength and good stability. However, they are prone to loosening due to factors such as material shrinkage, vehicle loads, and environmental erosion. If these defects are not identified and addressed in time, they will accelerate road surface damage. Traditional manual inspection is inefficient and highly subjective. Although ground-penetrating radar (GPR) non-destructive testing technology has become mainstream due to its speed and non-invasiveness, existing applications still have the following shortcomings: Insufficient realism of simulation data: Existing GPR forward modeling often uses homogeneous medium models, which do not match the "random distribution and non-uniformity" structural characteristics of actual semi-rigid base layers. This leads to significant differences between simulated and real detection data, resulting in weak generalization ability of the algorithm after training and limited accuracy in disease identification; Low accuracy of disease identification algorithms: Mainstream target detection algorithms (such as the YOLO series) have poor adaptability to targets like road diseases that are "irregularly shaped and have ambiguous features." Feature extraction modules (such as the C2f module) do not consider the multi-scale feature differences of diseases, making it easy to miss small diseases and inaccurate boundary identification. Loss functions (such as CIoU) are insufficient for optimizing bounding box regression in complex scenarios, resulting in large identification errors and difficulty in meeting engineering accuracy requirements; Lack of three-dimensional quantitative analysis: Existing technologies mostly remain at two-dimensional image recognition, without three-dimensional reconstruction and quantification of diseases. This makes it impossible to accurately obtain key parameters such as disease area and volume, resulting in a lack of quantitative basis for maintenance decisions and insufficient targeting of repair solutions.

[0003] Patent application CN114170480A discloses a multi-threshold semi-supervised intelligent detection method for road anomalies based on YOLO V5. The method includes collecting road anomaly images to form a teacher sample set, preprocessing the teacher sample set to obtain an initial training dataset; training a YOLO V5 object detection network to obtain an initial teacher model; further collecting real-world road image data and forming a student sample set based on a multi-threshold approach; performing strong data augmentation on the teacher sample set and weak data augmentation on the student sample set, and training the model based on the initial teacher model to obtain an evolved teacher model; continuously collecting real-world road image data and continuously optimizing the evolved teacher model to finally obtain a target road anomaly recognition and detection network; and performing intelligent detection of road anomalies. This patent application also uses the mainstream object detection algorithm YOLO for detection and recognition, and suffers from the same drawbacks as existing technologies.

[0004] Therefore, how to improve the accuracy of identifying loose defects in semi-rigid base layers and expand three-dimensional quantitative analysis is a problem that urgently needs to be solved by researchers in this field. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method for GPR image recognition and three-dimensional reconstruction of loose diseases in semi-rigid base layers, thereby solving the problems of low recognition accuracy of loose diseases in semi-rigid base layers and the inability to perform three-dimensional reconstruction and quantification of diseases in existing technologies.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] This invention provides a method for GPR image recognition and three-dimensional reconstruction of semi-rigid base loose disease, comprising the following steps:

[0008] S10. Construct a stochastic medium model of semi-rigid base loose disease, and perform forward modeling of road structures at different depths, different looseness levels and different radar antenna frequencies using the finite-difference time-domain method to generate simulation data and expand the dataset by combining data augmentation techniques.

[0009] S20. Design a DN module that integrates the dynamic deformation convolution module and the RepNCSPELAN4 module to replace the C2f module in the YOLOv8 algorithm. At the same time, introduce the WIoU loss function and use the dataset obtained in S10 to train the improved YOLOv8 algorithm to identify loose defects in semi-rigid base layers.

[0010] S30. Based on the instance segmentation results identified in S20, the area of ​​the diseased region is calculated by converting pixels to physical dimensions; a three-dimensional point cloud is generated by inverse distance weighted interpolation, and the volume of the disease is calculated by combining the Delaunay triangulation method.

[0011] Furthermore, the road structure layer parameters of the semi-rigid base loose disease random medium model in S10 are set as follows: the thickness of the asphalt concrete surface layer is 0.15m, the thickness of the semi-rigid base layer is 0.4m, the thickness of the semi-rigid subbase layer is 0.2m, and the thickness of the subgrade is 0.25m, wherein the dielectric constant of the base layer is randomly distributed in the range of 7.0 to 11.0.

[0012] Furthermore, the semi-rigid base loose disease random medium model in S10 has a size of 1.8m×1.0m, a spatial grid size of 0.002m×0.002m, a time window of 25ns, an excitation source type of Ricker, and an antenna frequency of 600MHz, 1000MHz, or 1600MHz.

[0013] Furthermore, the DN module in S20 is constructed as follows: the four hierarchical feature fusion paths of the RepNCSPELAN4 module are retained, the initial segmentation operation is improved, and the segmented features are processed by the joint operation of the DBB module and the RepNCSP module, the DBB module is processed separately, and finally the features are fused by the Concat operation.

[0014] Furthermore, the WIoU loss function in S20 adopts WIoU, and a gradient dynamic adjustment mechanism is constructed to configure advantageous weights for samples, focusing on ordinary quality detection boxes and suppressing the interference of low-quality samples on model parameter updates.

[0015] Furthermore, in step S30, let P be the weight decay index of the inverse distance weighted interpolation, and calculate the predicted value of the interpolation point using the following expression:

[0016]

[0017] Among them, D i This is the distance between the interpolation point and the known point.

[0018] Furthermore, in S30, the specific steps for calculating the volume using the Delaunay triangulation method are as follows:

[0019] S301. Triangulate the 3D point cloud data generated by inverse distance weighted interpolation to construct a triangular mesh that conforms to the empty circumcircle criterion and the minimum angle maximization criterion.

[0020] S302. Project each triangle onto the upper and lower boundaries of the point cloud slice to form a triangular prism, and calculate the volume of a single triangular prism.

[0021] S303. Sum the volumes of all triangular prisms to obtain the total volume of the diseased area.

[0022] Furthermore, in step S302, the expression for calculating the volume of a single triangular prism is as follows:

[0023]

[0024] in, Let h be the area of ​​the triangle and h be the height of the triangular prism.

[0025] The method for GPR image recognition and three-dimensional reconstruction of semi-rigid base loose disease provided by the present invention has at least the following advantages compared with the prior art:

[0026] Existing technologies suffer from low accuracy in identifying loose diseases in semi-rigid substrates and cannot perform 3D reconstruction and quantification of the diseases. This invention features a simple and convenient process. It restores the non-uniformity of semi-rigid substrates using a random medium model, and combines multi-condition simulation and data augmentation to provide a more comprehensive dataset covering various scenarios, offering high-quality samples for algorithm training. The DN module adapts to the multi-scale characteristics of the diseases, and the WIoU-optimized boundary regression algorithm achieves an improved mAP of 97.25% and an FPS of 32.05, ensuring both accurate disease identification and real-time performance. Furthermore, this invention achieves a transition from 2D identification to 3D reconstruction, quantifying the degree of disease development through area and volume calculations, providing a quantitative basis for prioritizing maintenance and developing repair plans. Attached Figure Description

[0027] To more clearly illustrate the solution of the present invention, a brief introduction will be given to the drawings used in the description of the embodiments below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0028] Figure 1 A flowchart of a method for GPR image recognition and three-dimensional reconstruction of semi-rigid base loose disease provided in an embodiment of the present invention;

[0029] Figure 2 This is a schematic diagram of a random medium model of a semi-rigid base layer in a method for GPR image recognition and three-dimensional reconstruction of loose defects in a semi-rigid base layer provided in an embodiment of the present invention.

[0030] Figure 3 This is a structural diagram of the C2f module in the YOLOv8 algorithm;

[0031] Figure 4 This is a structural diagram of the DN module constructed in a method for GPR image recognition and three-dimensional reconstruction of semi-rigid base loose disease provided in an embodiment of the present invention;

[0032] Figure 5 This is a schematic diagram illustrating the three-dimensional reconstruction effect of a semi-rigid base loose disease GPR image recognition and three-dimensional reconstruction method provided in an embodiment of the present invention. Detailed Implementation

[0033] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0034] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The invention provides:

[0035] This invention provides a method for GPR image recognition and 3D reconstruction of semi-rigid base course loose defects, which is applied to the accurate identification, quantitative analysis, and 3D visualization of road defects in semi-rigid base courses, providing technical support for road maintenance decision-making. The method for GPR image recognition and 3D reconstruction of semi-rigid base course loose defects includes the following steps:

[0036] S10. Construct a stochastic medium model for semi-rigid base loose disease. Perform forward modeling of road structures at different depths, looseness levels, and radar antenna frequencies using the finite-difference time-domain method to generate simulation data and expand the dataset using data augmentation techniques. S20. Design a DN module that integrates a dynamic deformation convolution module and a RepNCSPELAN4 module to replace the C2f module in the YOLOv8 algorithm. Introduce the WIoU loss function. Use the dataset obtained in S10 to train the improved YOLOv8 algorithm to identify semi-rigid base loose disease. S30. Based on the instance segmentation results identified in S20, calculate the disease area area by converting pixels to physical dimensions. Use inverse distance weighted interpolation to generate a 3D point cloud and combine it with the Delaunay triangulation method to calculate the disease volume.

[0037] This invention restores the non-uniformity of semi-rigid base layers through a random medium model, and combines multi-condition simulation and data augmentation to provide a more comprehensive dataset covering various scenarios, thus providing high-quality samples for algorithm training. The DN module adapts to the multi-scale characteristics of diseases, and the WIoU optimization algorithm boundary regression accurately identifies diseases while ensuring real-time performance. From two-dimensional identification to three-dimensional reconstruction, the invention quantifies the degree of disease development through area and volume calculations, providing a quantitative basis for prioritizing maintenance and formulating repair plans.

[0038] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0039] This invention provides a method for GPR image recognition and 3D reconstruction of semi-rigid base course loose defects, which is applied to the accurate identification, quantitative analysis, and 3D visualization of road defects in semi-rigid base courses, providing technical support for road maintenance decision-making. Figures 1 to 5 In this embodiment, the method for GPR image recognition and three-dimensional reconstruction of semi-rigid base loose disease includes the following steps:

[0040] S10. Construct a stochastic medium model of semi-rigid base loose disease, and perform forward modeling of road structures at different depths, different looseness levels and different radar antenna frequencies using the finite difference time-domain method to generate simulation data and expand the dataset by combining data augmentation techniques.

[0041] Specifically, in this embodiment, the specific steps for constructing the stochastic medium model of loose disease in semi-rigid base layers are as follows:

[0042] Random medium model construction: Based on the characteristics of semi-rigid base materials, a road structure consisting of surface layer, base layer, subbase layer, and soil layer is constructed. The base layer adopts a random medium model. The road structure layer parameters of the random medium model for loose defects in the semi-rigid base layer are set as follows: asphalt concrete surface layer thickness is 0.15m, semi-rigid base layer thickness is 0.4m, semi-rigid subbase layer thickness is 0.2m, and soil layer thickness is 0.25m. The dielectric constant of the base layer is randomly distributed in the range of 7.0~11.0 to simulate the non-uniformity of actual base materials. The dielectric constant of the loose defect area is reduced by 10%~30% to simulate defects with different degrees of looseness.

[0043] GPR forward modeling: The finite-difference time-domain (FDTD) method was used. The random medium model of loose disease in semi-rigid base layer was 1.8m×1.0m in size, with a spatial grid size of 0.002m×0.002m, a time window of 25ns, and a rakeer excitation source type. The model simulated antenna frequencies of 600MHz, 1000MHz, and 1600MHz, as well as multiple working condition combinations with disease depths of 0.2m to 0.6m and looseness of 10% to 40%, generating 3845 sets of initial simulation data.

[0044] Dataset augmentation and construction: The simulated data was augmented with adjustments to brightness and contrast, random rotation from -15° to 15°, cropping of diseased areas to ≥70%, addition of Gaussian noise with variance of 0.01 to 0.05, and local occlusion of ≤10% of the area, expanding the dataset to 7000+ samples; the training set, validation set, and test set were divided into three sets in an 8:1:1 ratio for algorithm training and validation.

[0045] S20. Design a DN module that integrates the dynamic deformation convolution module and the RepNCSPELAN4 module to replace the C2f module in the YOLOv8 algorithm. At the same time, introduce the WIoU loss function and use the dataset obtained in step S10 to train the improved YOLOv8 algorithm to identify loose defects in semi-rigid base layers.

[0046] Specifically, in this embodiment, the DN module in step S20 is constructed as follows: the four hierarchical feature fusion paths of the RepNCSPELAN4 module are retained, and the initial segmentation operation is improved so that the segmented features are processed by the joint operation of the DBB module and the RepNCSP module, and by the DBB module alone, and finally fused by the Concat operation. The DDB module dynamically adjusts the shape of the convolution kernel to adapt to the "irregular, multi-scale" features of diseases, enhancing the feature extraction capability for small diseases and blurred boundaries. The RepNCSPELAN4 module retains the four hierarchical feature fusion paths and improves the initial segmentation operation so that the features are processed by "DBB+RepNCSP (DBBNCSP) joint processing" and "DBB alone processing," and then fused by the Concat operation, improving the utilization rate of multi-scale features.

[0047] Specifically, in this embodiment, the WIoU loss function in step S20 uses WIoU. A gradient dynamic adjustment mechanism is constructed to assign advantageous weights to samples, focusing on ordinary quality bounding boxes and suppressing the interference of low-quality samples on model parameter updates. The WIoU loss function constructs a gradient dynamic adjustment mechanism: assigning advantageous weights to different quality bounding boxes, focusing on optimizing ordinary quality boxes (suppressing interference from low-quality boxes), improving bounding box regression accuracy, and solving the problem of inaccurate lesion localization in complex scenarios.

[0048] S30. Based on the instance segmentation results identified in step S20, the two-dimensional area of ​​the diseased region is calculated by converting pixels to physical dimensions. Inverse distance weighted interpolation (IDW) is used to interpolate the point cloud data of the diseased region to generate a dense three-dimensional point cloud. A triangular network is constructed by combining the Delaunay triangulation method to satisfy the criteria of empty circumcircle and minimum angle maximization. The volume of the disease is calculated. The triangular mesh and the upper and lower boundaries of the point cloud slice are projected to form a triangular prism. The volume of each individual part is calculated by formula, and the total volume of the disease is obtained by summing them up to achieve three-dimensional quantitative assessment.

[0049] Specifically, in this embodiment, in step S30, the weight decay index of the inverse distance weighted interpolation is set to P, and the predicted value of the interpolation point is calculated using the following expression:

[0050]

[0051] Among them, D i This is the distance between the interpolation point and the known point.

[0052] Specifically, in this embodiment, the specific steps for calculating the volume using the Delaunay triangulation method in step S30 are as follows:

[0053] S301. Triangulate the 3D point cloud data generated by inverse distance weighted interpolation to construct a triangular mesh that conforms to the empty circumcircle criterion and the minimum angle maximization criterion.

[0054] S302. Project each triangle onto the upper and lower boundaries of the point cloud slice to form a triangular prism, and calculate the volume of a single triangular prism.

[0055] S303. Sum the volumes of all triangular prisms to obtain the total volume of the diseased area.

[0056] Furthermore, in this embodiment, in step S302, the volume expression for a single triangular prism is calculated as follows:

[0057]

[0058] in, Let h be the area of ​​the triangle and h be the height of the triangular prism.

[0059] Example 1

[0060] Step 1: Environmental Preparation

[0061] Hardware: Equipped with an NVIDIA RTX 3090 GPU and an AMD Ryzen 9 7950X CPU to ensure computing power for forward modeling and algorithm training;

[0062] Software: Based on Python 3.8, using GprMax software, PyTorch, Open3D and other libraries to build the development environment.

[0063] Step 2: Simulation Data Generation and Implementation

[0064] Model parameter settings:

[0065] The structural layer thickness settings include 0.15m for the asphalt surface layer, 0.4m for the semi-rigid base layer, 0.2m for the subbase layer, and 0.25m for the subgrade. The dielectric constants are: 5.0 for the surface layer, 7.0~11.0 (random) for the base layer, 6.0 for the subbase layer, and 4.0 for the subgrade. The dielectric constant of the loose diseased area is achieved by randomly distributed unit air cells (dielectric constant of 1).

[0066] Forward simulation:

[0067] Write the GprMax input file, define the model structure, material parameters, and antenna parameters (600 / 1000 / 1600MHz), run the FDTD simulation, and output multi-condition GPR image data.

[0068] Data augmentation:

[0069] The OpenCV library is used to perform batch operations on simulated images, including brightness adjustment (±20%), rotation (-15°~15°), cropping (retaining ≥70% of the disease area), noise addition (variance 0.03), and occlusion (5% area), to generate an enhanced dataset.

[0070] Step 3: Algorithm Improvement and Training Implementation

[0071] DN module code implementation:

[0072] In the YOLOv8 network architecture, replace the C2f module with the DN module.

[0073] WIoU loss function ensemble:

[0074] Replace the original loss function with WIoU.

[0075] Algorithm training process:

[0076] Load the augmented dataset, divide it into 8:1:1 ratios and construct the dataset; initialize the improved YOLOv8 model, set the learning rate to 0.001 and batch_size to 16; use the Adam optimizer, combined with the WIoU loss function, train for 200 epochs, and validate and save the optimal model every 10 epochs.

[0077] Step 4: 3D Reconstruction and Quantitative Implementation

[0078] 2D recognition result conversion:

[0079] The trained model is called, the GPR detection image is input, and the disease instance segmentation mask is output. The pixel coordinates are converted into physical coordinates by calibration parameters (such as GPR antenna step size and sampling interval) and the two-dimensional area of ​​the disease is calculated.

[0080] 3D point cloud and reconstruction:

[0081] Extract point cloud data from the diseased area and generate a dense point cloud using inverse distance weighted interpolation (IDW); perform Delaunay triangulation on the generated point cloud to construct a triangular mesh.

[0082] Volume calculation and output:

[0083] Traverse the triangular mesh and calculate the volume of the triangular prism corresponding to each triangle.

[0084] Step 5: Verification and Application

[0085] Algorithm verification:

[0086] Validated on the test set, the improved algorithm achieved an mAP of 97.25% (higher than the original YOLOv8's 92.13%) and an FPS of 32.05, meeting the real-time and accuracy requirements of engineering detection.

[0087] Engineering applications:

[0088] By 3D modeling and simulating different ground-penetrating radar survey lines, the 3D model is sliced ​​to generate a profile radar map. The semi-rigid base loose disease GPR image recognition and 3D reconstruction method provided in this embodiment of the invention is used to detect and identify loose disease. The average accuracy of volume calculation after 3D reconstruction is 78.07%.

[0089] The GPR image recognition and 3D reconstruction method for semi-rigid base loose disease described in the above embodiments has advantages over existing technologies. Existing technologies have low accuracy in recognizing semi-rigid base loose disease and cannot perform 3D reconstruction and quantification of the disease. This invention features a simple process and convenient operation. It restores the non-uniformity of the semi-rigid base through a random medium model, and combines multi-condition simulation and data augmentation to provide a more comprehensive dataset covering various scenarios, offering high-quality samples for algorithm training. The DN module adapts to the multi-scale characteristics of the disease, and the WIoU-optimized boundary regression algorithm achieves an improved mAP of 97.25% and an FPS of 32.05, ensuring both accurate disease identification and real-time performance. Furthermore, this invention achieves a transition from 2D recognition to 3D reconstruction, quantifying the degree of disease development through area and volume calculations, providing a quantitative basis for maintenance priority allocation and repair plan formulation.

[0090] Obviously, the embodiments described above are merely preferred embodiments of the present invention, and not all embodiments. The accompanying drawings illustrate preferred embodiments of the present invention, but do not limit the scope of the patent. The present invention can be implemented in many different forms; rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this invention.

Claims

1. A method for GPR image recognition and three-dimensional reconstruction of semi-rigid base layer loose disease, characterized in that, Includes the following steps: S10. Construct a stochastic medium model of semi-rigid base loose disease, and perform forward modeling of road structures at different depths, different looseness levels and different radar antenna frequencies using the finite-difference time-domain method to generate simulation data and expand the dataset by combining data augmentation techniques. S20. Design a DN module that integrates the dynamic deformation convolution module and the RepNCSPELAN4 module to replace the C2f module in the YOLOv8 algorithm. At the same time, introduce the WIoU loss function and use the dataset obtained in S10 to train the improved YOLOv8 algorithm to identify loose defects in semi-rigid base layers. S30. Based on the instance segmentation results identified in S20, the area of ​​the diseased region is calculated by converting pixels to physical dimensions; a three-dimensional point cloud is generated by inverse distance weighted interpolation, and the volume of the disease is calculated by combining the Delaunay triangulation method.

2. The method for GPR image recognition and three-dimensional reconstruction of semi-rigid base loose disease according to claim 1, characterized in that, The road structure layer parameters of the semi-rigid base loose disease random medium model in S10 are set as follows: the thickness of the asphalt concrete surface layer is 0.15m, the thickness of the semi-rigid base layer is 0.4m, the thickness of the semi-rigid subbase layer is 0.2m, and the thickness of the subgrade is 0.25m, wherein the dielectric constant of the base layer is randomly distributed in the range of 7.0~11.

0.

3. The method for GPR image recognition and three-dimensional reconstruction of semi-rigid base loose disease according to claim 1, characterized in that, The semi-rigid base loose disease random medium model in S10 has a size of 1.8m×1.0m, a spatial grid size of 0.002m×0.002m, a time window of 25ns, an excitation source type of Ricker, and an antenna frequency of 600MHz, 1000MHz, or 1600MHz.

4. The method for GPR image recognition and three-dimensional reconstruction of semi-rigid base loose disease according to claim 1, characterized in that, The DN module in S20 is constructed as follows: the four hierarchical feature fusion paths of the RepNCSPELAN4 module are retained, the initial segmentation operation is improved, and the segmented features are processed by the joint operation of the DBB module and the RepNCSP module, the DBB module is processed separately, and finally the features are fused by the Concat operation.

5. The method for GPR image recognition and three-dimensional reconstruction of semi-rigid base loose disease according to claim 1, characterized in that, The WIoU loss function in S20 uses WIoU and assigns advantageous weights to samples by constructing a gradient dynamic adjustment mechanism, focusing on ordinary quality detection boxes and suppressing the interference of low-quality samples on model parameter updates.

6. The method for GPR image recognition and three-dimensional reconstruction of semi-rigid base loose disease according to claim 1, characterized in that, In step S30, let P be the weight decay index of the inverse distance weighted interpolation. The predicted value of the interpolation point is calculated using the following expression: ； Among them, D i This is the distance between the interpolation point and the known point.

7. The method for GPR image recognition and three-dimensional reconstruction of semi-rigid base loose disease according to claim 1, characterized in that, In S30, the specific steps for calculating the volume using the Delaunay triangulation method are as follows: S301. Triangulate the 3D point cloud data generated by inverse distance weighted interpolation to construct a triangular mesh that conforms to the empty circumcircle criterion and the minimum angle maximization criterion. S302. Project each triangle onto the upper and lower boundaries of the point cloud slice to form a triangular prism, and calculate the volume of a single triangular prism. S303. Sum the volumes of all triangular prisms to obtain the total volume of the diseased area.

8. A method for GPR image recognition and three-dimensional analysis of semi-rigid base layer loosening defects according to claim 7. The reconstruction method is characterized by, In step S302, the expression for calculating the volume of a single triangular prism is as follows: ； in, Let h be the area of ​​the triangle and h be the height of the triangular prism.

Images