Roadbed disease detection method based on yolov11 and reverse-time migration imaging algorithm

Abstract

The application relates to a subgrade disease detection method based on YOLOv11 and reverse time migration imaging algorithm, detection data of a subgrade is acquired by using a ground penetrating radar, a YOLOv11 network is improved, a CGAFusion module is introduced at the neck of the YOLOv11 network, a LSDECD detection head is used to replace a detection head of the YOLOv11 network, an MPDIoU loss function is used to replace a CIoU loss function, the improved YOLOv11 network is trained, and a subgrade hidden disease recognition model is obtained; the detection data is input into the subgrade hidden disease recognition model, hidden diseases in the subgrade are recognized, and the types of the hidden diseases are obtained; a reverse time migration imaging algorithm based on a GPU parallel sine algorithm is used to process the detection data, high-precision profile images of the subgrade are generated, and accurate positioning is carried out; the application can realize recognition and rapid imaging positioning of various subgrade hidden diseases, has high recognition precision, is accurate in positioning, and greatly improves the efficiency.

Description

Technical Field

[0001] This invention relates to a method for detecting roadbed defects, and more particularly to a method for detecting roadbed defects based on YOLOv11 and reverse time migration imaging algorithm. Background Technology

[0002] Hidden defects in urban roadbeds cause frequent road surface collapses, becoming a "city disease" threatening public safety. Hidden defects in roadbeds are characterized by their early concealment and high degree of suddenness. The traditional "firefighting" excavation and emergency repair model is a reactive measure that cannot meet the needs of modern urban road management.

[0003] Ground-penetrating radar (GPR) has advantages such as high efficiency and accurate positioning, and is widely used in road detection. However, when faced with massive and complex data, manual interpretation has low accuracy and cannot achieve automated real-time interpretation.

[0004] Deep learning combined with ground-penetrating radar, especially algorithms such as YOLO, has been widely used in anomaly identification. However, existing technologies are not very accurate in identifying various hidden defects in roadbeds, such as voids, loose bodies, and underground pipelines, and lack the ability to quantify physical information such as the location and size of hidden defects in the roadbed.

[0005] Although migration imaging technology can convert measured data into real structures and has the ability to quantify physical information such as the location and size of hidden roadbed defects, reverse time migration (RTM) imaging technology has high accuracy and strong adaptability, but it is computationally intensive and inefficient, making it difficult to meet the needs of real-time imaging. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide a roadbed defect detection method based on YOLOv11 and reverse time-shift imaging algorithm. This method can identify and quickly image and locate various hidden roadbed defects, with high identification accuracy, accurate positioning, saving calculation time and greatly improving efficiency.

[0007] The technical solution of the present invention:

[0008] A method for detecting roadbed defects based on YOLOv11 and reverse time migration imaging algorithm includes the following steps:

[0009] Step 1: Use ground-penetrating radar to detect the roadbed and obtain radar detection data of the roadbed.

[0010] Step 2: Enhance and amplify the radar detection data;

[0011] Step 3: Construct the training dataset for the YOLOv11 network using the amplified radar detection data;

[0012] Step 4: Improve the YOLOv11 network: Introduce the CGAFusion (Content-Guided Attention Fusion) module at the neck of the YOLOv11 network, replace the detection head of the YOLOv11 network with the LSDECD (Lightweight Shared Detail Enhanced Convolutional Detection Head) detection head, and use the MPDIoU loss function instead of the CIoU loss function of the YOLOv11 network; the improved YOLOv11 network is called the YOLOv11-CL network;

[0013] The CGAFusion module is used to fuse multi-scale features to improve the completeness of target feature representation. The fusion of high-level and low-level features is as follows: The CGAFusion module combines the multi-scale outputs of different layers of the network. High-level features generated by the backbone network are added element-wise with low-level features in the feature pyramid and fused. The fused feature map is then sent to the CGAFusion module for further processing. The CGA-processed feature map is multiplied element-wise with the original high-level and low-level features, and these two multiplication results are added to their respective original feature maps to generate enhanced feature representations. Finally, a 1×1 convolutional layer transforms the fused feature map into the final output. This design strategy effectively utilizes the complementarity between high-level and low-level features, enhancing feature completeness and expressiveness, and improving the network's detection performance. This, in turn, improves the model's robustness and accuracy in multimodal scenarios.

[0014] The CGAFusion module is a content-guided attention mechanism composed of spatial attention and channel attention. It guides the model to focus on key feature regions, enhancing the expression and interaction of multi-scale features and improving the accuracy and stability of object detection. Specifically, the CGAFusion module quickly captures globally significant features of the image by generating a coarse spatial attention map. Then, it refines the attention map channel by channel based on each channel of the input feature map, focusing on unique parts of the feature map. The weight Wc is calculated through the channel attention mechanism, and the weight Ws is calculated through the spatial attention mechanism. The two attention mechanisms work together to optimize the model's response to different features, improving the accuracy and efficiency of feature extraction.

[0015] The detection head of the YOLOv11 network is replaced with the LSDECD detection head to achieve lightweight shared detail enhancement convolution and improve detection accuracy. LSDECD first applies 1×1 convolution and group normalization to the P3, P4, and P5 feature maps of the neck input to achieve channel compression and standardization, reducing channel dimensionality and enhancing feature representation. Then, the features are input into the shared convolution module to achieve cross-scale parameter sharing and unified feature extraction, enhancing the interactivity between detection heads and the transitivity of features. The features output by the shared convolution are respectively entered into the classification branch and the regression branch, which output target category and location predictions through independent convolution operations.

[0016] The DEConv module of the LSDECD detection head incorporates vertical differential convolution, horizontal differential convolution, angular differential convolution, and central differential convolution. In the DEConv module, the original convolutions are used for baseline feature extraction. The outputs of each branch are fused element-wise, and then transformed into an equivalent single convolutional kernel using a reparameterization method. Combined with a group normalization detail enhancement strategy, it aggregates cross-level spatial and semantic information, improving feature representation capabilities and model generalization ability, reducing the computational burden on the model, and playing a crucial role in maintaining detection performance, thus ensuring accuracy and stability in multi-scale object detection tasks.

[0017] The MPDIoU loss function is used to replace the CIoU loss function of the YOLOv11 network to optimize the convergence and localization accuracy of bounding box regression. The CIoU loss function in the YOLOv11 network has some ambiguity in describing the relative aspect ratio. When the predicted box and the ground truth box have the same aspect ratio but different dimensions, the bounding box regression cannot be effectively optimized. In addition, the CIoU loss function relies on the overlapping area of ​​bounding boxes to measure similarity. When the target is partially occluded, the overlap area decreases, which leads to a decrease in the IoU value and affects the network detection accuracy. The MPDIoU loss function solves the occlusion problem by introducing the minimum point distance between bounding boxes and combines geometric information such as center point distance and width-to-height difference to improve the spatial distribution description of the target, thereby improving localization accuracy and robustness.

[0018] Step 5: Use the training dataset from Step 3 to train the YOLOv11-CL network to obtain a trained roadbed hidden disease identification model.

[0019] Step 6: Input the radar detection data of the roadbed by the ground penetrating radar into the roadbed hidden disease identification model to identify the hidden diseases in the roadbed and determine the type of hidden disease;

[0020] Step 7: Use the GPU-based parallel symplectic algorithm to process the radar detection data of the roadbed by the ground penetrating radar, generate a high-precision profile image of the roadbed, and accurately locate the hidden defects in the roadbed identified in Step 6.

[0021] In step 2, geometric transformation and cropping operations are used to enhance and augment the radar detection data.

[0022] In step 4, the MPDIoU loss function is defined as follows:

[0023]

[0024] In the formula, d1 and d2 represent the two Euclidean distances between the ground truth bounding box and the predicted bounding box, respectively; IoU represents the cross ratio between the predicted bounding box and the ground truth bounding box; w and h are the width and height of the input image, respectively; and L... MPDIoU Let MPDIoU be the loss function.

[0025] In step 6, the types of hidden defects include voids, loose bodies, and underground pipelines.

[0026] Step 7, the reverse time-shift imaging algorithm based on GPU parallel symplectic algorithm, includes the following steps:

[0027] Step 7.1: Initialize the model parameters, including dielectric constant, conductivity, and permeability;

[0028] Step 7.2: Discretize Maxwell's equations in isotropic lossy media using the symplectic algorithm to construct forward modeling equations;

[0029] Maxwell's equations for isotropic lossy media are expressed as follows:

[0030]

[0031] In the formula, E and H represent the electromagnetic field vector and the vector magnetomotive force, respectively, and ε, σ and μ are the permittivity, conductivity and permeability, respectively.

[0032] The vector magnetic potential H = ∇ × A, and given E = -U, the two-dimensional TM wave canonical equation is expressed as:

[0033]

[0034] In the formula, A z and U z Let A and U be the components of field components A and U along the z-axis, respectively, and ∇² be the Laplace operator;

[0035] Discretizing the two-dimensional TM wave canonical equations using the symplectic algorithm yields:

[0036]

[0037] In the formula, A n i,j and Un i,j The discrete value of the electromagnetic field vector at grid (i,j) at time step n;

[0038] Step 7.3: Calculate the forward propagation wave field using GPU parallel acceleration technology. The field component update process of each grid node is independent and synchronous.

[0039] Step 7.4: Store the boundary wavefield and the wavefield at the last moment;

[0040] Step 7.5: Use GPU parallel acceleration technology to calculate the backpropagation wave field, propagating backwards from the last moment;

[0041] Step 7.6: Generate profile images by applying cross-correlation imaging conditions.

[0042] Compared to the commonly used finite-difference time-domain method, which requires three equations to describe the entire electromagnetic field, the symplectic algorithm only requires the aforementioned two equations, necessitating less storage space and higher computational efficiency. The main computational task of ground-penetrating radar forward modeling based on the symplectic algorithm is updating the field components A and U. The field component update process for each grid node is independent and synchronous, exhibiting high parallelism, making it particularly suitable for GPU parallel acceleration computation.

[0043] Ground-penetrating radar migration imaging (RTM) works by propagating the recorded electromagnetic wave field backward along the time axis. When the wave field is pushed back to time zero, all reflected and diffracted wave energy returns to its initial position. Finally, imaging conditions are applied to obtain an imaging profile of the underground structure. The process consists of three steps:

[0044] Step 1: Solve the two-way wave equation and calculate the forward propagation wave field S(x,z,t) at the source point at different times;

[0045] Step 2: Extrapolate the electromagnetic wave at the receiving point in reverse time to calculate the corresponding back propagation wave field R(x,z,t);

[0046] Step 3: Apply imaging conditions to the forward propagation wavefield at the source point and the reverse propagation wavefield at the receiver point at the same time to obtain the imaging profile.

[0047] The cross-correlation imaging condition for ground-penetrating radar (RTM) imaging is expressed as:

[0048]

[0049] In the formula, I(x,z) represents the imaging result. Forward propagation wave field from the source point, The wave field is the reverse propagation field at the receiving point.

[0050] The beneficial effects of this invention are:

[0051] 1. This invention improves the YOLOv11 network by introducing a CGAFusion module at the neck of the YOLOv11 network, which enhances the robustness and accuracy of the detection model in multimodal scenarios; it replaces the detection head with an LSDECD detection head, realizing lightweight shared detail enhancement convolution, which improves the detection accuracy of the detection model; and it uses the MPDIoU loss function instead of the CIoU loss function, which improves the localization accuracy of the detection model.

[0052] 2. The reverse time migration imaging algorithm of this invention adopts the symplectic algorithm. Compared with the commonly used finite difference time-domain method, which requires three equations to describe the entire electromagnetic field, the symplectic algorithm only requires two equations, requires less storage space, and has higher computational efficiency.

[0053] 3. The main computational task of the ground-penetrating radar forward modeling based on the symplectic algorithm in this invention is to update the field components A and U. The field component update process of each grid node is independent and synchronous, which has high parallelism. Using GPU parallel acceleration technology can save computation time and greatly improve computational efficiency.

[0054] 4. This invention is based on reverse time-shift imaging using an improved YOLOv11 network and a GPU parallel symplectic algorithm, which can identify hidden defects such as road voids with high precision and locate them quickly and accurately. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of the YOLOv11-CL network structure;

[0056] Figure 2 This is a schematic diagram of the CGAFusion module structure;

[0057] Figure 3 A schematic diagram of the attention mechanism of the CGAFusion module;

[0058] Figure 4 This is a schematic diagram of the LSDECD detection head structure;

[0059] Figure 5 This is a schematic diagram of the DEConv module structure;

[0060] Figure 6 This is a flowchart of the GPU parallel symplectic algorithm for reverse time migration (RTM) imaging.

[0061] Figure 7 This is a schematic diagram of the urban road surface structure.

[0062] Figure 8 This is a forward modeling cross-section of a three-layer urban road surface structure detected using ground-penetrating radar.

[0063] Figure 9This is a cross-sectional view of the reverse time migration (RTM) imaging results of a three-layer road surface structure in the city.

[0064] Figure 10 A profile of the results of ground-penetrating radar detection of hidden defects on urban roads;

[0065] Figure 11 A cross-sectional view of the identification results of hidden defects on urban roads using the YOLOv11-CL network;

[0066] Figure 12 A cross-sectional view showing the identification and location results of hidden defects on urban roads using this invention.

[0067] In the diagram, 1 represents the surface layer, 2 the base layer, 3 the subbase layer, 4 the concrete pipe, 5 the void, 6 the loose body, and 7 the irregular cavity. Detailed Implementation

[0068] The method for detecting roadbed defects based on YOLOv11 and reverse time migration imaging algorithm includes the following steps:

[0069] Step 1: Use ground-penetrating radar to detect the roadbed of the municipal road and obtain radar detection data of the roadbed. A total of 837 detection data images were obtained.

[0070] Step 2: The radar detection data is enhanced and augmented using geometric transformation and cropping operations. The original images are augmented to 2046 images after random cropping, random flipping, translation and scaling.

[0071] Step 3: Use the amplified radar detection data to construct the training dataset for the YOLOv11 network. The training dataset is divided into training set, validation set and test set in a 7:2:1 ratio.

[0072] Step 4: Improve the YOLOv11 network: Introduce the CGAFusion (Content-Guided Attention Fusion) module at the neck of the YOLOv11 network, replace the YOLOv11 network's detection head with an LSDECD (Lightweight Shared Detail Enhanced Convolutional Detection Head) detection head, and replace the YOLOv11 network's CIoU loss function with the MPDIoU loss function; the improved YOLOv11 network is called the YOLOv11-CL network, and its structure is as follows. Figure 1 As shown in the figure, the locations of the CGAFusion module and the LSDECD detection head in the network are illustrated.

[0073] The CGAFusion module is used to fuse multi-scale features, improving the completeness of target feature representation; the structure of the CGAFusion module is as follows: Figure 2 As shown, the fusion method of high-level and low-level features is described in detail: The CGAFusion module combines the multi-scale outputs of different layers of the network. The high-level features generated by the backbone network are fused element-wise with the low-level features in the feature pyramid. The fused feature map is then sent to the CGAFusion module for further processing. The CGA-processed feature map is multiplied element-wise with the original high-level and low-level features, and the results of these two multiplications are added to their respective original feature maps to generate enhanced feature representations. Finally, a 1×1 convolutional layer is used to convert the fused feature map into the final output. This design strategy effectively utilizes the complementarity between high-level and low-level features, enhances feature integrity and expressiveness, and improves the detection performance of the network, thereby improving the robustness and accuracy of the model in multimodal scenarios.

[0074] The CGAFusion module is a content-guided attention mechanism composed of two attention mechanisms: spatial attention and channel attention. The attention mechanism of the CGAFusion module is as follows: Figure 3 As shown, by guiding the model to focus on key feature regions, the expression and interaction of multi-scale features are strengthened, thereby improving the accuracy and stability of object detection. The CGAFusion module works by quickly capturing globally significant features of the image through a generated coarse spatial attention map, and then refining the attention map channel by channel based on each channel of the input feature map, focusing on unique parts of the feature map. The weight Wc is calculated through a channel attention mechanism, and the weight Ws is calculated through a spatial attention mechanism. The two attention mechanisms are used together to optimize the model's response to different features, improving the accuracy and efficiency of feature extraction.

[0075] The detection head of the YOLOv11 network is replaced with an LSDECD detection head to achieve lightweight shared detail augmentation convolution, thereby improving detection accuracy. The structure of the LSDECD detection head is as follows: Figure 4 As shown, LSDECD first applies 1×1 convolution and group normalization to the P3, P4, and P5 feature maps input from the neck, achieving channel compression and standardization, reducing channel dimensions and enhancing feature expressive power. Then, the features are input into a shared convolution module to achieve cross-scale parameter sharing and unified feature extraction, enhancing the interactivity between detection heads and the transitivity of features. The features output from the shared convolution are respectively entered into the classification branch and the regression branch, which output target category and location predictions through independent convolution operations.

[0076] The DEConv module of the LSDECD detector head incorporates vertical differential convolution, horizontal differential convolution, angular differential convolution, and center differential convolution. The structure of the DEConv module is as follows: Figure 5 As shown, the original convolution in the DEConv module is used for baseline feature extraction. The outputs of each branch are fused element-wise and then transformed into an equivalent single convolution kernel using a reparameterization method. Combined with the group normalization detail enhancement strategy, it aggregates cross-level spatial and semantic information, improves feature representation ability and model generalization ability, reduces the computational burden of the model, and plays an important role in maintaining detection performance, ensuring accuracy and stability in multi-scale object detection tasks.

[0077] The MPDIoU loss function is used to replace the CIoU loss function of the YOLOv11 network to optimize the convergence and localization accuracy of bounding box regression. The CIoU loss function in the YOLOv11 network has some ambiguity in describing the relative aspect ratio. When the predicted box and the ground truth box have the same aspect ratio but different dimensions, the bounding box regression cannot be effectively optimized. In addition, the CIoU loss function relies on the overlapping area of ​​bounding boxes to measure similarity. When the target is partially occluded, the overlap area decreases, which leads to a decrease in the IoU value and affects the network detection accuracy. The MPDIoU loss function solves the occlusion problem by introducing the minimum point distance between bounding boxes and combines geometric information such as center point distance and width-to-height difference to improve the spatial distribution description of the target, thereby improving localization accuracy and robustness.

[0078] Step 5: Use the training dataset from Step 3 to train the YOLOv11-CL network to obtain a trained roadbed hidden disease identification model.

[0079] The network training environment was Windows 11, with an Intel i9-14900KF CPU and an NVIDIA GeForce RTX 5070 Ti GPU. It was developed based on the PyTorch 2.5.1 framework and Python 3.10. The network training input image resolution was set to 640×640, the optimizer was stochastic gradient descent, the initial learning rate was 0.01, the weight decay rate was 0.0005, the momentum parameter was 0.937, the total number of training epochs was 300, and the batch size was 32.

[0080] To verify the improvement effect of the YOLOv11-CL network, ablation experiments were conducted. The comparison of ablation experiment results for different models is shown in Table 1. Compared with the YOLOv11 network, the YOLOv11-CL network improved Precision, F1 and mAP@0.5 by 3.32%, 0.27% and 1.2% respectively. Among them, Precision and mAP@0.5 reached 88.17% and 88.01% respectively, while maintaining a certain lightweight computational cost of FLOPs.

[0081] Table 1

[0082] Model Precision (%) F1(%) mAP@0.5(%) FLOPs(G) YOLOv6n 78.92 80.52 84.63 11.5 YOLOv8n 84.61 82.59 86.24 6.8 YOLOv9s 85.56 83.35 85.89 22.1 YOLOv10n 84.00 79.84 85.10 6.5 YOLOv11n 84.85 83.18 86.81 6.3 YOLOv11-CL 88.17 83.45 88.01 7.5

[0083] Step 6: Input the radar detection data of the roadbed by the ground penetrating radar into the roadbed hidden disease identification model to identify the hidden diseases in the roadbed and determine the type of hidden disease;

[0084] Step 7: Use the GPU-based parallel symplectic algorithm to process the radar detection data of the roadbed by the ground penetrating radar, generate a high-precision profile image of the roadbed, and accurately locate the hidden defects in the roadbed identified in Step 6.

[0085] In step 4, the MPDIoU loss function is defined as follows:

[0086]

[0087] In the formula, d1 and d2 represent the two Euclidean distances between the ground truth bounding box and the predicted bounding box, respectively; IoU represents the cross ratio between the predicted bounding box and the ground truth bounding box; w and h are the width and height of the input image, respectively; and L... MPDIoU Let MPDIoU be the loss function.

[0088] In step 6, the types of hidden defects include voids, loose bodies, and underground pipelines.

[0089] In step 7, the flowchart of the reverse time-shift imaging algorithm based on GPU parallel symplectic algorithm is as follows: Figure 6 As shown in the figure, the complete calculation process from forward modeling to imaging localization includes the following steps:

[0090] Step 7.1: Initialize the model parameters, including dielectric constant, conductivity, and permeability;

[0091] Step 7.2: Discretize Maxwell's equations in isotropic lossy media using the symplectic algorithm to construct forward modeling equations;

[0092] Maxwell's equations for isotropic lossy media are expressed as follows:

[0093]

[0094] In the formula, E and H represent the electromagnetic field vector and the vector magnetomotive force, respectively, and ε, σ and μ are the permittivity, conductivity and permeability, respectively.

[0095] The vector magnetic potential H = ∇ × A, and given E = -U, the two-dimensional TM wave canonical equation is expressed as:

[0096]

[0097] In the formula, A z and U z Let A and U be the components of field components A and U along the z-axis, respectively, and ∇² be the Laplace operator;

[0098] Discretizing the two-dimensional TM wave canonical equations using the symplectic algorithm yields:

[0099]

[0100] In the formula, A n i,j and U n i,j The discrete value of the electromagnetic field vector at grid (i,j) at time step n;

[0101] Step 7.3: Calculate the forward propagation wave field using GPU parallel acceleration technology. The field component update process of each grid node is independent and synchronous.

[0102] Step 7.4: Store the boundary wavefield and the wavefield at the last moment;

[0103] Step 7.5: Use GPU parallel acceleration technology to calculate the backpropagation wave field, propagating backwards from the last moment;

[0104] Step 7.6: Generate profile images by applying cross-correlation imaging conditions.

[0105] Compared to the commonly used finite-difference time-domain method, which requires three equations to describe the entire electromagnetic field, the symplectic algorithm only requires the aforementioned two equations, necessitating less storage space and higher computational efficiency. The main computational task of ground-penetrating radar forward modeling based on the symplectic algorithm is updating the field components A and U. The field component update process for each grid node is independent and synchronous, exhibiting high parallelism, making it particularly suitable for GPU parallel acceleration computation.

[0106] Ground-penetrating radar migration imaging (RTM) works by propagating the recorded electromagnetic wave field backward along the time axis. When the wave field is pushed back to time zero, all reflected and diffracted wave energy returns to its initial position. Finally, imaging conditions are applied to obtain an imaging profile of the underground structure. The process consists of three steps:

[0107] Step 1: Solve the two-way wave equation and calculate the forward propagation wave field S(x,z,t) at the source point at different times;

[0108] Step 2: Extrapolate the electromagnetic wave at the receiving point in reverse time to calculate the corresponding back propagation wave field R(x,z,t);

[0109] Step 3: Apply imaging conditions to the forward propagation wavefield at the source point and the reverse propagation wavefield at the receiver point at the same time to obtain the imaging profile.

[0110] The cross-correlation imaging condition for ground-penetrating radar (RTM) imaging is expressed as:

[0111]

[0112] In the formula, I(x,z) represents the imaging result. Forward propagation wave field from the source point, The wave field is the reverse propagation field at the receiving point.

[0113] To verify the reliability and efficiency of the GPU-parallel symplectic algorithm-based reverse time migration imaging algorithm, a forward modeling simulation of ground-penetrating radar detection of urban pavement structures with various hidden defects was conducted. The urban pavement structure is a three-layer pavement structure, with surface layer 1, base layer 2, subbase layer 3, and defects arranged as follows: Figure 7 As shown, the relative permittivity of pavement surface layer 1 is 4, base layer 2 is 6, and subbase layer 3 is 10. A concrete pipe 4 with a relative permittivity of 8 is installed in pavement surface layer 1. In base layer 2, a void 5 and a loose body 6, each with a relative permittivity of 1, are sequentially arranged from left to right. An irregular void 7 is located in subbase layer 3. The excitation source was a 600MHz Ricker wavelet, and 400 data points were recorded.

[0114] Forward modeling profile of road surface structure detected by ground penetrating radar, such as Figure 8 As shown, the waveforms of the pavement structure layer interface, pipes, cavities and loose defects are clearly visible, but the waveforms of the layer interface and pipes and various hidden defects are coupled with each other, while the waveforms of deep cavities are dispersed.

[0115] Reverse time migration (RTM) imaging results are as follows Figure 9 As shown, the pavement layer interface is clearly visible, and the locations of pipes, voids, and loose defects are well-positioned, with their shapes and model settings being basically consistent.

[0116] In the reverse time migration (RTM) imaging localization of urban road surface structures, the RTM algorithm based on the finite-difference time-domain (FDTD) method takes 19556.57 s, while the RTM algorithm based on GPU parallel symplectic algorithm takes 965.28 s. Compared with commonly used migration imaging algorithms, the imaging localization algorithm proposed in this invention can save 95.06% of the computation time, indicating that this method can achieve rapid and accurate imaging localization of hidden defects in urban roadbeds.

[0117] To verify the applicability and reliability of the identification and positioning technology proposed in this invention in the detection of hidden defects on urban roads, a hidden defect detection, identification, and positioning exercise was conducted on a certain urban road. This engineering application used a Raptor-45 ground-penetrating radar with an antenna frequency of 450MHz for hidden defect detection.

[0118] Ground-penetrating radar (GPR) profile images were used to detect hidden defects in urban roads, such as... Figure 10 As shown; YOLOv11-CL network was used for disease detection and identification, and the results are as follows. Figure 11 As shown in the figure, the multi-hidden defect identification model based on YOLOv11-CL network proposed in this invention can accurately identify hidden defects such as road voids, realizing rapid and intelligent detection and evaluation of hidden defects in urban roadbeds.

[0119] Based on the YOLOv11-CL network recognition, reverse time migration (RTM) imaging localization processing was carried out. The imaging localization results are as follows: Figure 12 As shown in the figure, the waveforms of the pipeline and the cavity defects have been well repositioned. The cavity defects are elongated and are located at 0.92m.

[0120] To verify the applicability and reliability of the identification and positioning technology, borehole verification was performed on Void 5 disease, and the internal condition of Void 5 disease was observed using an endoscope. The endoscope revealed that the cavity was elongated, located at approximately 0.90m, which was basically consistent with the cavity location in the reverse time migration (RTM) imaging profile, with an error within 5%. The results indicate that the roadbed disease detection method based on YOLOv11 and the reverse time migration imaging algorithm proposed in this invention can achieve rapid and accurate detection and positioning of hidden diseases on urban roads.

Claims

1. A method for detecting roadbed defects based on YOLOv11 and reverse time migration imaging algorithm, comprising the following steps: Step 1: Use ground-penetrating radar to detect the roadbed and obtain radar detection data of the roadbed. Step 2: Enhance and amplify the radar detection data; Step 3: Construct the training dataset for the YOLOv11 network using the amplified radar detection data; Step 4: Improve the YOLOv11 network: Introduce the CGAFusion module into the neck of the YOLOv11 network, replace the detection head of the YOLOv11 network with the LSDECD detection head, and replace the CIoU loss function of the YOLOv11 network with the MPDIoU loss function; the improved YOLOv11 network is called the YOLOv11-CL network. Step 5: Use the training dataset from Step 3 to train the YOLOv11-CL network to obtain a trained roadbed hidden disease identification model. Step 6: Input the radar detection data of the roadbed by the ground penetrating radar into the roadbed hidden disease identification model to identify the hidden diseases in the roadbed and determine the type of hidden disease; Step 7: Use the GPU-based parallel symplectic algorithm to process the radar detection data of the roadbed by the ground penetrating radar, generate a high-precision profile image of the roadbed, and accurately locate the hidden defects in the roadbed identified in Step 6.

2. The method for detecting roadbed defects based on YOLOv11 and reverse time migration imaging algorithm according to claim 1, characterized in that: In step 2, geometric transformation and cropping operations are used to enhance and amplify the radar detection data.

3. The method for detecting roadbed defects based on YOLOv11 and reverse time migration imaging algorithm according to claim 1, characterized in that: In step 4, the MPDIoU loss function is defined as follows: In the formula, d1 and d2 represent the two Euclidean distances between the ground truth bounding box and the predicted bounding box, respectively; IoU represents the cross ratio between the predicted bounding box and the ground truth bounding box; w and h are the width and height of the input image, respectively; and L... MPDIoU Let MPDIoU be the loss function.

4. The method for detecting roadbed defects based on YOLOv11 and reverse time migration imaging algorithm according to claim 1, characterized in that: In step 6, the types of hidden defects include voids, loose bodies, and underground pipelines.

5. The method for detecting roadbed defects based on YOLOv11 and reverse time migration imaging algorithm according to claim 1, characterized in that: Step 7 of the reverse time-shift imaging algorithm based on GPU parallel symplectic algorithm includes the following steps: Step 7.1: Initialize the model parameters, which include dielectric constant, conductivity, and permeability; Step 7.2: Discretize Maxwell's equations in isotropic lossy media using the symplectic algorithm to construct forward modeling equations; Maxwell's equations for isotropic lossy media are expressed as follows: In the formula, E and H represent the electromagnetic field vector and the vector magnetomotive force, respectively, and ε, σ and μ are the permittivity, conductivity and permeability, respectively. The vector magnetic potential H = ∇ × A, and given E = -U, the two-dimensional TM wave canonical equation is expressed as: In the formula, A z and U z Let A and U be the components of field components A and U along the z-axis, respectively, and ∇² be the Laplace operator; Discretizing the two-dimensional TM wave canonical equations using the symplectic algorithm yields: In the formula, A n i,j and U n i,j The discrete value of the electromagnetic field vector at grid (i,j) at time step n; Step 7.3: Calculate the forward propagation wave field using GPU parallel acceleration technology; Step 7.4: Store the boundary wavefield and the wavefield at the last moment; Step 7.5: Use GPU parallel acceleration technology to calculate the backpropagation wave field, propagating backwards from the last moment; Step 7.6: Generate profile images by applying cross-correlation imaging conditions.

Images