Ground penetrating radar urban road hidden disease image detection method based on YOLO11n-GPR

Abstract

The invention relates to a ground penetrating radar urban road hidden disease image detection method based on YOLO11n-GPR, and the method comprises the steps: obtaining an urban road underground crack and cavity image data set collected by a ground penetrating radar, and carrying out the preprocessing; a target detection model YOLO11n-a based on the improved YOLO11n is constructed, and YOLO11n-b is further obtained according to a structured pruning strategy; the YOLO11n-a serves as a teacher model, the YOLO11n-b serves as a student model, knowledge distillation training is conducted through the training set, and a hidden disease detection model YOLO11n-GPR is obtained; and obtaining a to-be-detected urban road underground ground penetrating radar scanning image, inputting the image into the hidden disease detection model YOLO11n-GPR, and outputting an urban road hidden disease detection result. Compared with the prior art, the method has the advantages of solving the problems of complexity, accuracy and timeliness of existing detection and the like.

Description

Technical Field

[0001] The present invention relates to the interdisciplinary technical field of deep learning and geophysical exploration, and in particular to a method for detecting hidden defects in urban roads using ground-penetrating radar images based on YOLO11n-GPR. Background Art

[0002] As an integral part of urban transportation, urban roads play a vital role in ensuring the safety of drivers and pedestrians. Affected by traffic and environmental factors, urban roads are prone to hidden defects such as cracks and voids. Furthermore, issues such as insufficient roadbed compaction, excessive groundwater extraction, and damaged municipal underground pipelines can lead to waterlogging, loosening, and voids in the roadbed. These defects, collectively known as hidden dangers of urban roads, are characterized by their concealment, insidiousness, and sudden onset. If these hidden dangers are neglected, they will gradually expand and eventually spread to the road surface, causing uneven settlement and collapse, leading to significant economic losses.

[0003] Traditional methods for detecting road hazards can be divided into two categories. Core sampling, a destructive method, can provide a comprehensive understanding of road structure, detect the location of defects, and classify them. However, core sampling damages the road surface, requires a lot of manpower, and is inefficient. In contrast, methods such as drop weight rebound testers, ultrasonic testing, high-density electrical testing, and geological radar (GPR) are non-destructive testing methods. These methods do not damage the road surface, are fast, and have low labor costs. Among non-destructive testing methods, GPR stands out due to its high accuracy and efficiency.

[0004] However, manual interpretation of GPR images requires a high level of expertise, and due to its slow recognition speed, it is difficult to meet the needs of daily maintenance. In recent years, deep learning technology has rapidly developed. By deepening the training network and gradually extracting high-dimensional features from the data, it avoids the strong subjectivity and complex image processing steps of traditional methods, greatly improving recognition accuracy. YOLO11 is the latest object detection model in the YOLO series, capable of high-precision and high-speed detection in a variety of natural scenes. However, due to the complexity and unique characteristics of GPR images, YOLO11's detection accuracy still needs to be further improved. Summary of the Invention

[0005] The purpose of the present invention is to provide a ground penetrating radar urban road hidden disease image detection method based on YOLO11n-GPR, which can improve both detection accuracy and detection efficiency.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] A method for detecting hidden damage images of urban roads using ground penetrating radar (GPR) based on YOLO11n-GPR includes the following steps:

[0008] Obtain a dataset of underground cracks and cavities in urban roads collected by ground-penetrating radar, perform preprocessing, and divide it into training, validation, and test sets.

[0009] Build the target detection model YOLO11n-a based on the improved YOLO11n, and further obtain YOLO11n-b according to the structured pruning strategy;

[0010] The YOLO11n-a is used as the teacher model, the YOLO11n-b is used as the student model, and the training set is used to perform knowledge distillation training to obtain the latent disease detection model YOLO11n-GPR;

[0011] Obtain the underground ground penetrating radar scanning image of the urban road to be detected, input it into the hidden disease detection model YOLO11n-GPR, and output the detection results of the hidden disease of the urban road.

[0012] Furthermore, the pre-processing step includes:

[0013] The urban road underground crack and cavity image dataset was subjected to operations including zero point correction, bandpass filtering, background removal, and time-varying gain, and the hidden defects were annotated using the Makesense website.

[0014] Furthermore, the target detection model YOLO11n-a based on the improved YOLO11n is obtained by improving the backbone network, neck network and loss function of the YOLO11n model, wherein:

[0015] The improved backbone network is obtained by replacing the Conv module in the backbone network before improvement with the GhostConv module.

[0016] The improved neck network is obtained by adding the BiFormer attention mechanism module to the neck network before improvement.

[0017] The improved loss function is obtained by replacing the loss function before improvement with the MDPIoU loss function.

[0018] Furthermore, the GhostConv module includes a Conv layer, a Cheap operation layer, and a Concat layer connected in sequence, wherein the Conv layer is further connected to the Concat layer. The execution steps of the GhostConv module include:

[0019] In the Conv layer, a convolution kernel of half size is used to generate half of the feature map, and a 5x5 convolution kernel is used to perform a linear operation in the Cheap operation layer to obtain the other half of the feature map. Finally, in the Concat layer, the feature map of one half and the feature map of the other half are spliced ​​to obtain a complete feature map.

[0020] Furthermore, the BiFormer attention mechanism module includes a DWConv layer, a first normalization layer, a two-layer routing attention layer, a second normalization layer and an MLP layer connected in sequence.

[0021] Furthermore, the MDPIoU loss function is expressed as:

[0022]

[0023] Where, is the MDPIoU loss, IoU is the intersection-over-union loss, w and h represent the width and height of the input image in the network, and Represents the coordinates of the upper left and lower right corners of the prediction box, and Represents the coordinates of the upper left corner and lower right corner of the ground truth box.

[0024] Furthermore, the step of obtaining YOLO11n-b includes:

[0025] Use the training set to perform sparse training on YOLO11n-a, and obtain the weight last.pt after the training;

[0026] According to the closeness of the BN layer γ coefficient in the weight last.pt to zero, channel pruning is performed to obtain the weight prune.pt after pruning;

[0027] Fine-tune the weight prune.pt to obtain YOLO11n-b.

[0028] Furthermore, during the sparsification training process, the loss function used is:

[0029]

[0030] Where L is the sparsification training loss, the first term is the training loss of the network, x and y are the training input and target, W is the trainable parameter in the grid, the second term is the L1 regularization constraint term of the γ coefficient of the BN layer, that is, g(γ) = |γ|, and λ is the penalty factor used to balance the first and second terms.

[0031] Furthermore, the step of obtaining the hidden disease detection model YOLO11n-GPR includes:

[0032] According to the isomorphic distillation theory, a middle layer is selected from the teacher model as a guiding layer, and a corresponding middle layer is selected from the student model as a guided layer;

[0033] Based on the training set, the intermediate feature expression in the guiding layer is used to guide the guided layer to learn, and the mimic method is used to calculate the intermediate feature knowledge distillation loss;

[0034] After the training is completed, the hidden disease detection model YOLO11n-GPR is obtained.

[0035] Furthermore, the calculation expression of the intermediate feature knowledge distillation loss is:

[0036]

[0037] Where, L kd is the intermediate feature knowledge distillation loss, N i is the middle layer feature dimension, is the middle layer feature of the teacher model, is the intermediate layer feature of the student model, r is the feature dimension adapter, and n is the number of intermediate feature knowledge distillation layers.

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

[0039] (1) The present invention addresses the complexity and particularity of ground penetrating radar images and the detection accuracy issues of the existing YOLO11n. By improving YOLO11n and introducing the ideas of model pruning and model distillation, the detection and recognition speed is significantly improved while maintaining high-precision detection.

[0040] (2) The present invention establishes a real radar image dataset, including underground cracks and cavities, and preprocesses the dataset to improve image quality, reduce the impact of noise on disease identification, and help improve the stability and robustness of the target detection algorithm.

[0041] (3) The present invention improves the YOLO11n model using the GhostConv module, the BiFormer attention mechanism module, and the MDPIoU loss function. The GhostConv module reduces the computational and parameter overhead while maintaining the feature map size. The BiFormer attention mechanism module can effectively reduce the number of model parameters and computational complexity, accelerate training, and enhance the generalization ability of the model. It can also improve the model's ability to focus on different features by processing and adjusting attention weights, and better detect small target defects. The MDPIoU loss function combines the minimum point distance and redefines the loss function through a metric, thereby reducing the overall degrees of freedom. By improving the YOLO11n model, the present invention improves recognition efficiency while improving recognition accuracy.

[0042] (4) The MDPIoU loss function of the present invention solves the limitations of the CIoU loss function before improvement, making it more suitable for GPR latent disease detection. According to the formula of the MPDIoU loss function, it encourages the predicted box to approach the true box under conditions such as non-overlapping center points. In addition, when the predicted box and the true box have overlapping center points and consistent aspect ratios, but the predicted sizes are different, the penalty term in MPDIoU is still not zero, preventing degeneration into IoU loss. Therefore, compared with the CIoU loss function, the use of MPDIoU not only simplifies the calculation, but also stabilizes the convergence of the model, and improves the accuracy of detecting subtle diseases in complex environments.

[0043] (5) In the knowledge distillation process, the present invention adopts the concept of isomorphic distillation, using the feature expression of the intermediate layer of the teacher model to guide the learning of the corresponding layer of the student model. By imitating the feature expression of the intermediate layer of the teacher model, the student model can learn more general and representative feature patterns, thereby improving the generalization ability of the student model. In addition, when the training data is limited, isomorphic distillation enables the student model to fully utilize the knowledge of the teacher model. Even on a small data set, it can learn better feature representations, avoid problems such as overfitting due to insufficient data, and improve the performance of the model in small data set scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 Schematic diagram of the method flow of the present invention;

[0045] Figure 2 Schematic diagram of the GhostConv module of the present invention;

[0046] Figure 3 Schematic diagram of the Biformer attention mechanism module of the present invention;

[0047] Figure 4 Schematic diagram of the YOLO11n-a model structure of the present invention;

[0048] Figure 5 Schematic diagram of the intermediate feature knowledge distillation of the present invention. DETAILED DESCRIPTION

[0049] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0050] This embodiment provides a method for detecting hidden defects in urban road images using ground penetrating radar based on YOLO11n-GPR. Figure 1 As shown, the method includes the following steps:

[0051] S1. Data construction: Construct a dataset of underground cracks and voids in urban roads and perform data preprocessing, dividing the training set, validation set, and test set into a certain proportion.

[0052] Ground-penetrating radar equipment is used to obtain ground-penetrating radar B-scan profile images of underground cracks and voids in urban roads. Professional technicians screen the collected images to retain images of underground cracks and voids with obvious features. The image processing software Photoshop is used to crop out useless information such as blank areas around them.

[0053] The original data set is preprocessed to generate label files of underground cracks and cavities ground penetrating radar images, and the training set, validation set, and test set are divided according to a certain ratio. The specific process is as follows:

[0054] 1. Zero point correction: Take the starting point of ground reflection as the zero point to eliminate the time error caused by the gap between the antenna and the ground;

[0055] 2. Bandpass filtering: Use low-pass filtering to filter out high-frequency components in the signal, and use high-pass filtering to filter out low-frequency components in the signal, retaining the antenna's main frequency, thereby improving the signal-to-noise ratio of the radar wave;

[0056] 3. Background elimination: Use the amplitude value of the original signal to subtract the amplitude mean of all single-channel signal waves to eliminate the background noise of the radar wave;

[0057] 4. Time-varying gain: By multiplying each single-channel signal by an amplification coefficient, the energy attenuation of the radar wave when propagating into the underground medium is compensated, thereby improving the reflection characteristics of the target;

[0058] 5. Use the MakeSense website to create label files for the dataset images. Combine the calibrated PNG images with the generated txt annotation information to build a database of underground fractures and cavities. Split the database into training, validation, and test sets in a ratio of 0.7:0.15:0.15.

[0059] S2. Improved YOLO11n model: Build the target detection model YOLO11n-a based on the improved YOLO11n. The specific operations include improving the backbone network, neck network and loss function.

[0060] This step mainly improves YOLO11n to form the target detection model YOLO11n-a based on the improved YOLO11n. The improvements include the backbone network, the neck network, and the loss function:

[0061] The GhostConv module is used to replace the Conv module in the backbone network, which reduces the computational and parameter overhead while maintaining the feature map size. The BiFormer attention mechanism module is added to the neck structure to better detect small target defects. The loss function is replaced with MDPIoU, which improves the bounding box regression by minimizing the distance between the predicted box and the ground-truth box corner points.

[0062] Specifically, if Figure 2 As shown in the figure, the GhostConv module consists of a sequentially connected Conv layer, a Cheap operation layer, and a Concat layer. The Conv layer is also connected to the Concat layer. "Cheap operation" refers to a low-cost linear operation. In this GhostConv module, the Conv layer first uses a convolution kernel with half the size of the original convolution to generate half of the feature map. Next, a Cheap operation is performed in the Cheap operation layer using a 5x5 convolution kernel with a stride of 1 to obtain the other half of the feature map. Finally, the Concat layer uses the Concat operation to splice these two sets of feature maps together to form the complete feature map.

[0063] Specifically, if Figure 3As shown in the figure, the BiFormer attention mechanism module is designed based on a two-layer routing attention, including a DWConv layer (depthwise separable convolution), a first normalization layer (LN), a two-level routing attention layer (Bi-level Routing Attention), a second normalization layer (LN), and an MLP layer connected in sequence. In this module, DWConv refers to the depthwise separable convolution, which effectively reduces the number of parameters and computational complexity of the model. LN represents layer normalization, which accelerates training and enhances the generalization ability of the model. MLP refers to the multi-layer perceptron, which further processes and adjusts the attention weights, thereby improving the model's ability to focus on different features. Figure 3 In , the plus sign represents the concatenation of two eigenvectors.

[0064] Specifically, the MDPIoU loss function combines the minimum point distance and redefines the loss function through a metric, thereby reducing the overall degrees of freedom. The specific formula is as follows:

[0065]

[0066] Where w and h represent the width and height of the input image to the network, and Represents the coordinates of the upper left and lower right corners of the prediction box, and Represents the coordinates of the upper left corner and lower right corner of the ground truth box.

[0067] The MPDIoU loss function addresses the limitations of the CIoU loss function, making it more suitable for GPR latent disease detection. According to the MPDIoU loss function formula, it encourages the predicted box to approach the ground-truth box under conditions such as non-overlapping center points. Furthermore, when the predicted and ground-truth boxes have overlapping center points and consistent aspect ratios, but different predicted sizes, the penalty term in MPDIoU remains non-zero, preventing degradation to IoU loss. Therefore, compared to the CIoU loss function, using MPDIoU not only simplifies computation but also stabilizes model convergence, improving the accuracy of detecting subtle diseases in complex environments.

[0068] According to the above improvements, the improved YOLO11n-a model diagram is as follows Figure 4 shown.

[0069] S3. Model pruning: Design a pruning strategy to obtain YOLO11n-b. The specific operations include sparse training, structured pruning, and model fine-tuning of YOLO11n-a.

[0070] This step uses YOLO11n-a to perform sparse training on the ground penetrating radar dataset. During the sparse training phase, an L1 regularization penalty term is added to the original loss function. The formula is as follows:

[0071]

[0072] In the above formula, the first term is the training loss of the network, x and y are the training input and target, W is the trainable parameter in the grid, the second term is the L1 regularization constraint term of the γ coefficient of the BN layer, that is, g(γ) = |γ|, and λ is the penalty factor used to balance the first and second terms.

[0073] In this embodiment, the penalty factors are selected as 0.004, 0.008, and 0.0012 for sparse training. After the three parameter sparse training, most of the β coefficients of the BN layer are small enough. When the penalty factor is 0.004, the accuracy of the model decreases the least compared with the original model, as shown in Table 1. Therefore, the penalty factor is selected as 0.004, and the weight last.pt is obtained after the training.

[0074] Table 1. Sparse training results

[0075]

[0076] Prune the channels in last.pt where γ is close to zero to compress the model structure and improve the recognition speed. After pruning, we get the weight prune.pt. The pruning result is:

[0077] Global pruning threshold: 1.2744

[0078] Reserved channel ratio: 16 / 16 (100.00%)

[0079] Retained channel ratio: 13 / 32 (40.62%)

[0080] Retained channel ratio: 22 / 32 (68.75%)

[0081] Retained channel ratio: 52 / 64 (81.25%)

[0082] Retained channel ratio: 56 / 64 (87.50%)

[0083] Reserved channel ratio: 108 / 128 (84.38%)

[0084] Retained channel ratio: 44 / 64 (68.75%)

[0085] Retained channel ratio: 18 / 32 (56.25%)

[0086] Retained channel ratio: 49 / 64 (76.56%)

[0087] Retained channel ratio: 107 / 128 (83.59%)

[0088] Fine-tune the weights obtained after pruning to recover some of the accuracy lost due to pruning. When the accuracy no longer improves, the YOLO11n-b model is obtained.

[0089] S4. Model distillation: Perform knowledge distillation on the pruned model, use YOLO11n-b as the student model, and YOLO11n-a as the teacher model. After knowledge distillation training, the YOLO11n-GPR model is obtained.

[0090] Since pruning does not change the overall architecture of the model, the teacher model and the student model have the same architecture, and there is a one-to-one correspondence between layers, so the present invention adopts Figure 5 The isomorphic distillation shown directly uses the feature expressions of the intermediate layers of the teacher model to guide the learning of the corresponding layers of the student model.

[0091] The present invention selects the 17th, 20th and 23rd layers of the teacher model and the student model as the guiding layer and the guided layer during the intermediate feature knowledge distillation, and uses the mimic method to calculate the intermediate feature knowledge distillation loss. The specific formula is:

[0092]

[0093] In the above formula, L kd is the intermediate feature knowledge distillation loss, N i is the middle layer feature dimension, is the middle layer feature of the teacher model, is the intermediate layer feature of the student model, r is the feature dimension adapter, and n is the number of intermediate feature knowledge distillation layers.

[0094] At the end of training, the YOLO11n-GPR model with significantly reduced parameter and computational complexity and no loss of accuracy was obtained.

[0095] S5. Model testing: Introduce the test set to test YOLO11n-GPR, output various detection numerical indicators, and determine whether they reach the expected value.

[0096] This embodiment also sets up an ablation experiment to measure the effect of the improved model through evaluation indicators, including: precision P, recall R, average precision mAP and detection speed FPS. The specific formula is as follows:

[0097] Precision P—is defined as the accuracy of all detected targets, which can be expressed as

[0098]

[0099] Recall rate R—defined as the detection accuracy rate among all positive samples, which can be expressed as

[0100]

[0101] The area enclosed by the AP-PR curve and the coordinate axis can be expressed as

[0102]

[0103] mAP—the mean of all class APs, which can be expressed as

[0104]

[0105] FPS—defined as the number of detection images per second for the YOLOv8-CBAM-AFPN model

[0106] Among them, TP is the number of positive samples predicted as positive; FP is the number of negative samples predicted as positive; FN is the number of positive samples predicted as negative; TN is the number of negative samples predicted as negative, n is the category, which is 2 in the present invention.

[0107] Table 2 Ablation experiment

[0108]

[0109] Analysis of Table 2 shows that when the network replaces the C2f module with the C2f-CBAM module, or introduces the AFPN feature fusion network, the precision and recall rates are slightly improved, the mAP value also increases, but the FPS decreases slightly. The reason for this may be that the introduction of the attention mechanism and the adaptive feature fusion module increases the complexity of the model, and the corresponding performance also improves, but the inference efficiency decreases. After the introduction of transfer learning, all indicators are improved, indicating that transfer learning helps to improve the model training speed and model performance.

[0110] Replacing the convolution module in the backbone network with the Ghostconv module results in a slight decrease in recall rate, but a significant improvement in accuracy and average precision, and the detection efficiency is also improved; the introduction of the Biformer attention mechanism improves the focus on key information in the feature map, which increases the mAP50 by 3.9%; the use of the MPDIoU loss function improves the detection accuracy of the model, effectively overcoming the limitations of the CIoU loss function in detecting small anomalies in actual scenarios.

[0111] Table 3 Comparative experiments after model pruning and model distillation

[0112]

[0113] From the results in Table 3, we can see that the model after model pruning and model distillation does not lose too much accuracy, but the detection efficiency is significantly improved.

[0114] In summary, a ground-penetrating radar urban road hidden disease image detection method based on YOLO11n-GPR can help improve the accuracy and efficiency of disease identification, which is beneficial to road maintenance work.

[0115] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0116] Those skilled in the art will appreciate that the embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention may be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0117] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0118] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0119] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0120] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0121] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A method for detecting hidden damage images of urban roads using ground penetrating radar based on YOLO11n-GPR, characterized in that: The following steps are involved: Obtain a dataset of underground cracks and cavities in urban roads collected by ground-penetrating radar, perform preprocessing, and divide it into training, validation, and test sets. Build the target detection model YOLO11n-a based on the improved YOLO11n, and further obtain YOLO11n-b according to the structured pruning strategy; The YOLO11n-a is used as the teacher model, the YOLO11n-b is used as the student model, and the training set is used to perform knowledge distillation training to obtain the latent disease detection model YOLO11n-GPR; Obtain the underground ground penetrating radar scanning image of the urban road to be detected, input it into the hidden disease detection model YOLO11n-GPR, and output the detection results of the hidden disease of the urban road.

2. The method for detecting hidden defects in urban road images using ground penetrating radar based on YOLO11n-GPR according to claim 1, characterized in that: The pre-processing steps include: The urban road underground crack and cavity image dataset was subjected to operations including zero point correction, bandpass filtering, background removal, and time-varying gain, and the hidden defects were annotated using the Makesense website.

3. The method for detecting hidden defects in urban road images using ground penetrating radar based on YOLO11n-GPR according to claim 1, characterized in that: The target detection model YOLO11n-a based on the improved YOLO11n is obtained by improving the backbone network, neck network and loss function of the YOLO11n model, wherein: The improved backbone network is obtained by replacing the Conv module in the backbone network before improvement with the GhostConv module. The improved neck network is obtained by adding the BiFormer attention mechanism module to the neck network before improvement. The improved loss function is obtained by replacing the loss function before improvement with the MDPIoU loss function.

4. The method for detecting hidden defects in urban road images using ground penetrating radar based on YOLO11n-GPR according to claim 3, characterized in that: The GhostConv module includes a Conv layer, a Cheap operation layer, and a Concat layer connected in sequence, wherein the Conv layer is further connected to the Concat layer. The execution steps of the GhostConv module include: In the Conv layer, a convolution kernel of half size is used to generate half of the feature map, and a 5x5 convolution kernel is used to perform a linear operation in the Cheap operation layer to obtain the other half of the feature map. Finally, in the Concat layer, the feature map of one half and the feature map of the other half are spliced ​​to obtain a complete feature map.

5. The method for detecting hidden defects in urban road images using ground penetrating radar based on YOLO11n-GPR according to claim 3, characterized in that: The BiFormer attention mechanism module includes a DWConv layer, a first normalization layer, a two-layer routing attention layer, a second normalization layer and an MLP layer connected in sequence.

6. The method for detecting hidden defects in urban road images using ground penetrating radar based on YOLO11n-GPR according to claim 3, characterized in that: The MDPIoU loss function is expressed as: Where, is the MDPIoU loss, IoU is the intersection-over-union loss, w and h represent the width and height of the input image in the network, and Represents the coordinates of the upper left and lower right corners of the prediction box, and Represents the coordinates of the upper left corner and lower right corner of the ground truth box.

7. The method for detecting hidden defects in urban road images using ground penetrating radar based on YOLO11n-GPR according to claim 1, characterized in that: The steps of obtaining YOLO11n-b include: Use the training set to perform sparse training on YOLO11n-a, and obtain the weight last.pt after the training; According to the closeness of the BN layer γ coefficient in the weight last.pt to zero, channel pruning is performed to obtain the weight prune.pt after pruning; Fine-tune the weight prune.pt to obtain YOLO11n-b.

8. The method for detecting hidden defects in urban road images using ground penetrating radar based on YOLO11n-GPR according to claim 7, characterized in that: During the sparsification training process, the loss function used is: Where L is the sparsification training loss, the first term is the training loss of the network, x and y are the training input and target, W is the trainable parameter in the grid, the second term is the L1 regularization constraint term of the γ coefficient of the BN layer, that is, g(γ) = |γ|, and λ is the penalty factor used to balance the first and second terms.

9. The method for detecting hidden defects in urban road images using ground penetrating radar based on YOLO11n-GPR according to claim 1, characterized in that: The steps of obtaining the hidden disease detection model YOLO11n-GPR include: According to the isomorphic distillation theory, a middle layer is selected from the teacher model as a guiding layer, and a corresponding middle layer is selected from the student model as a guided layer; Based on the training set, the intermediate feature expression in the guiding layer is used to guide the guided layer to learn, and the mimic method is used to calculate the intermediate feature knowledge distillation loss; After the training is completed, the hidden disease detection model YOLO11n-GPR is obtained.

10. The method for detecting hidden defects in urban road images using ground penetrating radar based on YOLO11n-GPR according to claim 9, characterized in that: The calculation expression of the intermediate feature knowledge distillation loss is: Where, L kd is the intermediate feature knowledge distillation loss, N i is the middle layer feature dimension, is the middle layer feature of the teacher model, is the intermediate layer feature of the student model, r is the feature dimension adapter, and n is the number of intermediate feature knowledge distillation layers.

Images