A method and electronic equipment for detecting underground targets on airport pavements

By constructing a ground-penetrating radar dataset and graph relation network, target features are extracted and enhanced, and prediction models are optimized, thus solving the limitations in underground target detection on airport pavements and achieving higher accuracy and reliability in detection.

CN118070212BActive Publication Date: 2026-11-17CIVIL AVIATION UNIV OF CHINA
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202410043853.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-01-11
Publication Date
2026-11-17
Estimated Expiration
2044-01-11

AI Technical Summary

Technical Problem

Existing technologies for detecting underground targets on airport pavements have drawbacks such as localized monitoring, limited monitoring depth, complex installation and operation, difficulty in real-time monitoring, and reliance on surface reference points, resulting in poor detection performance.

Method used

By constructing a ground-penetrating radar dataset, establishing co-occurrence relation matrices and adjacency matrices, extracting target features, constructing a graph relation network for feature enhancement, and combining it with a prediction model for training and optimization, the detection results of underground targets are obtained.

Benefits of technology

It improves the accuracy and reliability of underground target detection, reduces missed detections and false detections, and enhances detection effectiveness.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118070212B_ABST
    Figure CN118070212B_ABST
Patent Text Reader

Abstract

This invention relates to the field of underground target detection technology and discloses a method for detecting underground targets on airport pavements. The method includes establishing a ground-penetrating radar dataset, establishing a co-occurrence matrix and adjacency matrix between underground targets, extracting target features from underground targets, calculating the weights of relationships between underground targets and constructing a graph relationship network, performing feature enhancement on the target features to obtain enhanced features, fusing the enhanced features corresponding to the training set with the target features and inputting them into a prediction model to obtain the corresponding prediction results, training and optimizing the prediction model, inputting the test set into the optimized prediction model for evaluation to obtain the target prediction model, and inputting newly acquired detection data into the target prediction model to obtain the underground target detection results. This invention, by constructing a co-occurrence matrix and a graph relationship network, and by performing feature enhancement and feature fusion on the target features, improves the accuracy and reliability of the target prediction model and achieves better detection results.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of underground target detection technology, and in particular to a method and electronic equipment for detecting underground targets on airport pavements. Background Technology

[0002] Airport runways, as a crucial infrastructure for the safe operation of aircraft, inevitably suffer varying degrees of damage to their underground structures due to prolonged high-intensity loads and complex environmental factors. This results in defects such as runway delamination, settlement, loosening, and cracking. The concealed location and potential hazards of these defects pose a significant risk to the safe and stable operation of airport runways. However, detecting these underground defects presents a major challenge. Since the underground structure of airport runways mainly consists of soil foundations, subbases, base courses, and cement concrete slabs, and the material structures of different geological layers vary, direct detection of underground defects is extremely difficult. Detection typically requires underground detection equipment, such as deflectometers, ground-penetrating radar, and ground-penetrating radar.

[0003] The disadvantages of using deflectometers for underground target detection include five aspects: 1. Localized monitoring: Deflectometers typically only monitor underground structures at specific locations, failing to provide an overall picture of underground targets, which may lead to some targets being overlooked. 2. Limited monitoring depth: The monitoring depth of deflectometers is usually limited by the length of the detection equipment and the installation method. For deeper underground targets, it may be impossible to fully detect them or accurately assess their extent and impact. 3. Complex installation and operation: The installation and operation of deflectometers usually require specialized technicians and involve complex procedures and equipment. This increases the cost and difficulty of using deflectometers for underground target monitoring. 4. Difficulty in real-time monitoring: Deflectometers typically require periodic manual measurement and recording, and cannot provide real-time monitoring and data transmission. This may result in changes and evolution of underground targets not being captured and responded to in a timely manner. 5. Reliance on surface reference points: Deflectometer measurements usually need to be compared with surface reference points to determine the displacement and deformation of underground structures. However, the stability and accuracy of surface reference points may be affected by the surface environment and other factors, potentially introducing errors. In summary, deflectometers have drawbacks in underground target detection, including limited monitoring depth, complex installation and operation, difficulty in real-time monitoring, and reliance on surface reference points. Therefore, in practical applications, these factors must be comprehensively considered, making it impossible to detect underground targets quickly and with high accuracy.

[0004] Ground penetrating radar (GPR) is a non-destructive testing device widely used for underground target detection. It boasts advantages such as real-time performance, multi-dimensionality, high precision, high efficiency, wide coverage, and fast detection speed, making it one of the mainstream methods for underground target detection and widely applied in various underground defect detection fields. Currently, GPR-based underground target detection is mainly divided into three types: 1. GPR target detection methods based on traditional methods primarily target high-frequency underground targets, detecting only a single target and requiring the design of independent feature extractors for each defect, thus exhibiting strong limitations. In GPR B-scan images, pipeline targets typically exhibit hyperbolic features, and methods such as Hough Transform and Least Squares (LS) are used to detect underground pipeline targets and locate their positions. 2. Methods based on traditional machine learning algorithms, such as Support Vector Machine (SVM) and Random Forest (RF), are also widely used for underground defect target detection. Third, using YOLO series models on GPR-based 3D C-scan data, we applied them to road inspection engineering, achieving the detection of hidden cracks, defects, and cavities beneath roads. However, traditional methods and machine learning methods require substantial data and feature engineering support, limiting their performance in complex geological environments and restricting their accuracy and reliability for underground targets. Deep learning models can overcome the shortcomings of traditional and machine learning methods, but they lack domain-specific knowledge and have insufficient understanding of the characteristics and mechanisms of underground targets. This lack of domain knowledge may prevent the model from fully utilizing the experience and observations of domain experts, and deep learning models perform poorly in detecting multiple underground targets.

[0005] Therefore, based on the above problems, how to improve the effectiveness of underground target detection on the road surface is an urgent issue to be addressed. Summary of the Invention

[0006] To address the technical problems in the prior art, the present invention provides a method and electronic device for detecting underground targets on airport pavements.

[0007] This invention includes a method for detecting underground targets on airport pavements, comprising the following steps:

[0008] Acquire detection data from several ground-penetrating radars, establish a ground-penetrating radar dataset based on the detection data, and divide the ground-penetrating radar dataset into a training set and a test set;

[0009] A co-occurrence matrix among underground targets is established based on the ground-penetrating radar dataset.

[0010] Construct an adjacency matrix between the underground targets based on the co-occurrence relationship matrix;

[0011] According to a preset feature extraction strategy, target features of different types of underground targets are extracted and stored from the ground-penetrating radar dataset;

[0012] The relationship weights between different types of underground targets are calculated based on the preset feature similarity calculation strategy and the target features, and a graph relationship network is constructed based on the relationship weights and the adjacency matrix.

[0013] Based on the graph relation network, the target features are enhanced to obtain enhanced features;

[0014] The enhanced features corresponding to the training set are fused with the target features and input into a preset prediction model to obtain the corresponding prediction results. The prediction model is then trained and optimized based on the prediction results, the true results, and the set loss function.

[0015] The test set is input into the optimized prediction model for evaluation, and the target prediction model is obtained after the preset evaluation conditions are met.

[0016] The newly acquired detection data from any ground-penetrating radar is input into the target prediction model to obtain the underground target detection results.

[0017] Furthermore, the acquisition of detection data from several ground-penetrating radars, the establishment of a ground-penetrating radar dataset based on the detection data, and the division of the ground-penetrating radar dataset into a training set and a test set include:

[0018] Acquire B-scan image data from several ground-penetrating radars;

[0019] The B-scan image data is cropped into a single-channel grayscale image of a set size;

[0020] Obtain manual annotations for the single-channel grayscale image; the manual annotations are used to mark underground targets in the single-channel grayscale image;

[0021] The ground-penetrating radar dataset is established based on the single-channel grayscale image and the manual annotation.

[0022] The ground-penetrating radar dataset is divided into a training set and a test set according to a random classification algorithm.

[0023] Furthermore, establishing the co-occurrence matrix among underground targets based on the ground-penetrating radar dataset includes:

[0024] Acquire information about underground targets in a single-channel grayscale image;

[0025] An empty N*N matrix is ​​created to record the co-occurrence relationships among the underground targets; N is the number of types of underground targets.

[0026] Increment the co-occurrence value of every two underground targets in the single-channel grayscale image by one;

[0027] Process all single-channel grayscale images in the ground-penetrating radar dataset according to the above steps to obtain the co-occurrence matrix of all underground targets.

[0028] Furthermore, constructing the adjacency matrix between the underground targets based on the co-occurrence relation matrix includes:

[0029] The co-occurrence relation matrix is ​​used to obtain the association relation probability matrix through row normalization.

[0030] Based on a preset association probability threshold, an adjacency matrix is ​​constructed between the underground targets.

[0031] Furthermore, the step of extracting and storing target features of different types of underground targets from the ground-penetrating radar dataset according to a preset feature extraction strategy includes:

[0032] The training set is input into a preset residual network to obtain disease features {C1, C2, C3, C4, C5} at different scales; the residual network contains several residual blocks, and each residual block consists of multiple convolutional layers and skip connections;

[0033] The disease features are input into a preset pyramid network, which captures target features at different scales and semantic levels. The target features at different scales are fused to obtain target features {P2, P3, P4, P5}, and the last layer of deep features {C5} of the disease features is retained.

[0034] The depth feature {C5} is convolved to obtain the target feature {P6} with a larger receptive field;

[0035] The target features are input into the RPN network, which generates disease prior boxes through AnchorsGenerator. At the same time, the RPN network filters the input target features based on the disease prior boxes and extracts the target features that are most likely to be underground targets.

[0036] Furthermore, the step of calculating the relationship weights between different types of underground targets based on a preset feature similarity calculation strategy and the target features, and constructing a graph relationship network based on the relationship weights and the adjacency matrix, includes:

[0037] The dimensions of the target features are transformed to be consistent, and then used as nodes;

[0038] according to Calculate the weight e of the relationship between adjacent nodes in the graph generated by the association of different nodes. ij σ is a nonlinear activation function, e ij Let α represent the importance of node j to node i, and let α represent the concatenation mapping of the target features of node i and the target features of node j. This represents the target feature corresponding to node i. This represents the target feature corresponding to node j;

[0039] Using the adjacency matrix as the adjacency matrix of the graph, and setting the relation weights e... ij The graph's edge weights are used to construct relationships, resulting in the graph relationship network.

[0040] Furthermore, the step of enhancing the target features based on the graph relation network to obtain enhanced features includes:

[0041] according to Calculate the weights of adjacent nodes and input them as propagation weights into the graph relationship network; α ij N represents the degree of association between node j and node i, k represents the k-th node that is associated with node i, and N represents the degree of association between node j and node i. i This represents the number of N nodes that are associated with node i;

[0042] pass Calculate the enhanced features; This represents the enhanced feature corresponding to node i. This represents the target feature corresponding to node j.

[0043] Furthermore, the step of training and optimizing the prediction model based on the predicted results, the actual results, and the set loss function includes:

[0044] The loss function is: in, The classification loss is calculated as follows: The regression loss is calculated as follows: p i and t i The prediction result, and p represents the actual result, t represents the classification result, λ represents the location result, and λ represents the set coefficient.

[0045] Furthermore, the step of inputting the test set into the optimized prediction model for evaluation includes:

[0046] Evaluation indicators include:

[0047] Accuracy: TP represents the number of correct predictions, and FP represents the number of incorrect predictions.

[0048] Recall rate: FN represents the target number of missed detections;

[0049] mAP value: AP represents the prediction accuracy for each type of target;

[0050] F1 value:

[0051] The present invention also includes an electronic device comprising a memory and a processor, wherein the memory stores a calculator program and the processor is configured to implement the above method via a computer program.

[0052] The present invention relates to an airport pavement underground target detection method and electronic device, which acquires detection data from several ground-penetrating radars (GPRs) and establishes a GPR dataset based on the detection data. A co-occurrence relation matrix between underground targets is then established based on the GPR dataset. An adjacency matrix between underground targets is constructed based on the co-occurrence relation matrix. Target features of different types of underground targets are extracted and stored from the GPR dataset according to a preset feature extraction strategy. Next, the relationship weights between different types of underground targets are calculated based on a preset feature similarity calculation strategy and the target features. A graph relation network is constructed based on the relation weights and the adjacency matrix. Feature enhancement is then performed on the target features based on the graph relation network to obtain enhanced features. The enhanced features corresponding to the training set are then fused with the target features and input into a preset prediction model to obtain the corresponding prediction results. The prediction model is trained and optimized based on the prediction results, the actual results, and a set loss function. A test set is then input into the optimized prediction model for evaluation. After meeting preset evaluation conditions, a target prediction model is obtained. After obtaining the target prediction model, newly acquired detection data from any newly acquired GPR is input into the target prediction model to obtain the underground target detection results. This invention constructs a co-occurrence relation matrix and a graph relation network, and enhances and fuses the target features. It considers not only the characteristics of the pavement underground disease itself, but also the environmental characteristics of the disease and the feature distortion caused by the feature complexity and similarity between different underground target samples. Thus, domain knowledge is integrated into the prediction model as guiding prior knowledge, making the accuracy and reliability of the trained target prediction model higher and the detection effect better. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is a flowchart (I) of a method for detecting underground targets on an airport pavement according to an embodiment of the present invention;

[0055] Figure 2 This is a flowchart (II) of a method for detecting underground targets on an airport pavement according to an embodiment of the present invention;

[0056] Figure 3 This is a flowchart (III) of a method for detecting underground targets on an airport pavement according to an embodiment of the present invention;

[0057] Figure 4 This is a schematic diagram of the co-occurrence relationship matrix of an airport pavement underground target detection method according to an embodiment of the present invention;

[0058] Figure 5 This is a schematic diagram of the correlation probability matrix of an airport pavement underground target detection method according to an embodiment of the present invention;

[0059] Figure 6 This is a schematic diagram (a) of the residual block of an airport pavement underground target detection method according to an embodiment of the present invention;

[0060] Figure 7 This is a schematic diagram (II) of the residual block of an airport pavement underground target detection method according to an embodiment of the present invention;

[0061] Figure 8 This is a schematic diagram of the RPN network structure of an airport pavement underground target detection method according to an embodiment of the present invention;

[0062] Figure 9 This is a schematic diagram of the node association relationship in an airport pavement underground target detection method according to an embodiment of the present invention;

[0063] Figure 10 This is an experimental comparison diagram of an airport pavement underground target detection method according to an embodiment of the present invention;

[0064] Figure 11 This is a structural diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0065] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0066] An embodiment of the present invention provides a method for detecting underground targets on airport pavements, such as... Figure 1 As shown, the steps include:

[0067] Step S10: Obtain detection data from several ground-penetrating radars, establish a ground-penetrating radar dataset based on the detection data, and divide the ground-penetrating radar dataset into a training set and a test set.

[0068] Specifically, such as Figure 2 As shown, step S10 includes:

[0069] Step S101: Acquire B-scan image data from several ground-penetrating radars.

[0070] Ground penetrating radar can detect reflected signals from underground objects by emitting and receiving electromagnetic waves, and then generate B-scan images through signal processing software.

[0071] This step uses ground-penetrating radar to scan the airport runway and obtain B-scan images of the underground structure.

[0072] Step S102: Crop the B-scan image data into a single-channel grayscale image of a set size.

[0073] As an example, the acquired B-scan image is cropped into a 448x448 pixel single-channel grayscale image.

[0074] Step S102 is equivalent to the preprocessing of the detection data, so that all B-scan images reach a uniform specification, which facilitates the subsequent extraction and processing of information.

[0075] Step S103: Obtain manual annotations for the single-channel grayscale image; manual annotations are used to mark underground targets in the single-channel grayscale image.

[0076] Based on the experience and knowledge of experts, the cropped single-channel grayscale images are manually annotated to mark underground targets of the airport runway that appear in the single-channel grayscale images, such as cracks, voids, settlement, loosening, pipes, ground lights and steel bars.

[0077] Step S104: Establish a ground-penetrating radar dataset based on single-channel grayscale images and manual annotations.

[0078] Based on the above manual annotations, the data can be formatted as PASCAL VOC dataset to form a ground-penetrating radar dataset for underground targets.

[0079] Step S105: Divide the ground-penetrating radar dataset into a training set and a test set according to the random classification algorithm.

[0080] The number of training and test sets can be set based on experience, and this embodiment does not impose a specific limit.

[0081] After establishing the ground-penetrating radar dataset, proceed to step S20.

[0082] Step S20: Establish a co-occurrence relationship matrix among underground targets based on the ground penetrating radar dataset.

[0083] Specifically, such as Figure 3 As shown, step S20 includes:

[0084] Step S201: Obtain underground target information from a single-channel grayscale image.

[0085] Step S202: Create an empty N*N matrix to record the co-occurrence relationships between underground targets.

[0086] In this step, N represents the number of different types of underground targets.

[0087] Step S203: Increment the co-occurrence value of every two underground targets in the single-channel grayscale image by one.

[0088] Step S204: Process all single-channel grayscale images in the ground-penetrating radar dataset according to the above steps to obtain the co-occurrence matrix of all underground targets.

[0089] like Figure 4 The diagram shown is a schematic of a co-occurrence matrix, where the numbers represent the number of times two corresponding targets co-occur.

[0090] After proposing the co-occurrence relation matrix, proceed to step S30: construct the adjacency matrix between underground targets based on the co-occurrence relation matrix.

[0091] Specifically, step S30 includes:

[0092] Step S301: Obtain the association probability matrix from the co-occurrence relation matrix using the row normalization method.

[0093] Normalizing the co-occurrence relation matrix yields the following: Figure 5 The association probability matrix shown indicates that darker colors represent higher association probability values, while lighter colors represent lower association probability values. Figure 5 Each colored square in the diagram corresponds to a probability value for association, and the value is less than or equal to 1.

[0094] Step S302: Construct an adjacency matrix between underground targets based on a preset association probability threshold.

[0095] The association probability threshold is set manually by technical personnel, and the specific value can be determined according to the circumstances.

[0096] Step S40: Extract and store target features of different types of underground targets from the ground-penetrating radar dataset according to the preset feature extraction strategy.

[0097] Specifically, step S40 in this embodiment includes:

[0098] Step S401: Input the training set into the preset residual network to obtain disease features {C1, C2, C3, C4, C5} at different scales.

[0099] The residual network in this embodiment, also known as the ResNet50 network, contains several residual blocks, such as... Figure 6 and Figure 7 As shown, each residual block consists of multiple convolutional layers and skip connections, resulting in disease features {C1, C2, C3, C4, C5} at different scales.

[0100] Step S402: Input the disease features into the preset pyramid network. The pyramid network captures target features at different scales and semantic levels. The target features at different scales are fused to obtain target features {P2, P3, P4, P5}. The last layer of deep features {C5} of the disease features is retained.

[0101] The pyramid network FPN fuses {C2, C3, C4, C5} to obtain the target features {P2, P3, P4, P5}, and then retains the last layer deep features {C5} of ResNet50 for the next step of processing.

[0102] Step S403: Perform convolution operation on the depth feature {C5} to obtain the target feature {P6} with a larger receptive field.

[0103] Step S404: Input the target features into the RPN network. The RPN network generates disease prior boxes through AnchorsGenerator. At the same time, the RPN network filters the input target features based on the disease prior boxes and extracts the target features that are most likely to be underground targets.

[0104] The generated multi-scale target features are fed into the RPN network for processing, such as... Figure 8 The diagram shown is a schematic of the RPN network structure.

[0105] Step S50: Calculate the relationship weights between different types of underground targets according to the preset feature similarity calculation strategy and target features, and construct a graph relationship network based on the relationship weights and adjacency matrix.

[0106] Specifically, this step includes:

[0107] Step S501: Convert the dimensions of the target features to be consistent and use them as nodes.

[0108] Step S502: According to Calculate the weight e of the relationship between adjacent nodes in the graph generated by the association of different nodes. ij σ is a nonlinear activation function, e ij Let α represent the importance of node j to node i, and let α represent the concatenation mapping of the target features of node i and the target features of node j. This represents the target feature corresponding to node i. This represents the target feature corresponding to node j;

[0109] Step S503: Use the adjacency matrix as the adjacency matrix of the graph, and set the relation weights e. ij As edge weights of the graph, association relationships are constructed to obtain a graph relationship network.

[0110] Step S60: Enhance the target features based on the graph relation network to obtain enhanced features.

[0111] Specifically, this step includes:

[0112] according to Calculate the weights of adjacent nodes and input them into the graph relation network as propagation weights; α ij N represents the degree of association between node i and node j, k represents the k-th node that is associated with node i, and N represents the degree of association between node i and node j. i This represents the number of N nodes that are associated with node i;

[0113] pass Calculate enhanced features; This represents the enhanced feature corresponding to node i. This represents the target feature corresponding to node j.

[0114] The implementation process of the above steps is as follows: Figure 9 As shown, this illustrates the relationships between nodes.

[0115] Step S70: Fuse the enhanced features and target features corresponding to the training set and input them into the preset prediction model to obtain the corresponding prediction results. Train and optimize the prediction model based on the prediction results, the true results and the set loss function.

[0116] In this step, the loss function can be set to in, The classification loss is calculated as follows: The regression loss is calculated as follows: p i and t i For the predicted results, and p represents the actual result, t represents the classification result, λ represents the location result, and λ represents the set coefficient.

[0117] The prediction results in this step may include the confidence level, category, and location information of possible diseases.

[0118] In this embodiment, the loss function includes classification loss and regression loss. Using the backpropagation algorithm, starting with the loss, the gradient of each model parameter with respect to the loss is calculated using the chain rule. Backpropagation propagates the gradient layer by layer from the output layer to the input layer, updating the value of each parameter based on the gradient, iteratively obtaining more accurate model parameters for the prediction model.

[0119] Step S80: Input the test set into the optimized prediction model for evaluation, and obtain the target prediction model after the preset evaluation conditions are met.

[0120] In this embodiment, when the test set is input into the optimized prediction model for evaluation, the evaluation metrics include:

[0121] Accuracy: TP represents the number of correct predictions, and FP represents the number of incorrect predictions.

[0122] Recall rate: FN represents the target number of missed detections;

[0123] mAP value: AP represents the prediction accuracy for each type of target;

[0124] F1 value:

[0125] After obtaining the optimal target prediction model, the detection result can be predicted based on the newly detected detection data, that is:

[0126] Step S90: Input the newly acquired detection data from any ground-penetrating radar into the target prediction model to obtain the underground target detection results.

[0127] Before the detection data in this step is input into the target prediction model, it should be preprocessed, i.e., step S102 of the above embodiment, to ensure that it is of the same specification as the sample data, so as to obtain more accurate detection results. The prediction results may also include information such as the location, type and confidence level of underground diseases.

[0128] Comparative experiments were conducted using embodiments of this invention and five mainstream existing target detection algorithms. The dataset consisted of 80% as the training set and 20% as the test set. The comparison and evaluation of the five algorithms were performed under conditions of IoU of 0.5 and a confidence threshold of 0.5. Table 1 shows the results of each algorithm in defect detection. The algorithm proposed in this application outperforms other algorithms in F1 score for cracks, voids, settlement, ground lights, rebar bevel angle, and rebar parallelism. Its accuracy and recall are also partially superior to other algorithms. Figure 10 The visualization results comparing the embodiments of the present invention with other algorithms are listed. Figure 10 In experiment a), Faster R-CNN and YOLOX falsely detected ground lights and void defects, while RetinaNet, SSD, and RetinaNet falsely detected loose defects and missed rebar targets. In experiment b, RetinaNet and SSD missed loose defects, while Faster R-CNN and YOLOX missed small void defects. In experiment c, Faster R-CNN missed void defects, while RetinaNet, SSD, and YOLOX missed rebar and loose defect targets. Based on the experimental data, it can be confirmed that the present invention can effectively reduce the number of missed and false detections of defects, effectively verifying the effectiveness of the correlation reasoning. Table 2 lists the effect of the method of the present invention on improving the mAP value. The comparison results of experiments 1, 2, 3, and 4 show that the deep features and correlations of defects are effectively extracted, and the correlations between defects are effectively utilized, resulting in a 12.85% improvement in the mAP value.

[0129] Table 1 Comparison of the performance of detection methods

[0130]

[0131] Table 2 mAP performance improvement

[0132]

[0133] In some embodiments, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the airport pavement underground target detection method of the foregoing embodiments.

[0134] like Figure 11As shown, in some embodiments, this application also provides an electronic device, which may be a mobile phone, computer or tablet computer, etc., including a memory 201 and a processor 202. The memory 201 stores a calculator program, and the processor 202 is configured to implement the airport pavement underground target detection method of the aforementioned embodiments through a computer program.

[0135] The processor 202 executes a computer program to implement the airport pavement underground target detection method described in the above embodiment. The memory 201 is used to store various types of data, which may include, for example, instructions for any application or method in an electronic device, as well as application-related data.

[0136] The processor 202 may be implemented as an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field-programmable gate array (FPGA), a controller, a microcontroller, a microprocessor, or other electronic components, and is used to execute the airport pavement underground target detection method in the above embodiments.

[0137] The memory 201 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0138] like Figure 11 As shown, it can be understood that the electronic device 200 may also include an input / output (I / O) interface 203 and a communication component 204.

[0139] I / O interface 203 provides an interface between processor 202 and other interface modules, such as keyboards, mice, and buttons. These buttons can be virtual or physical. Communication component 204 is used for wired or wireless communication between the electronic device 200 and other devices. Wireless communication includes Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof. Therefore, the corresponding communication component 204 may include a Wi-Fi module, a Bluetooth module, and an NFC module.

[0140] The airport pavement underground target detection method and electronic device of this invention acquires detection data from several ground-penetrating radars (GPRs), establishes a GPR dataset based on the detection data, establishes a co-occurrence relation matrix between underground targets based on the GPR dataset, constructs an adjacency matrix between underground targets based on the co-occurrence relation matrix, extracts and stores target features of different types of underground targets from the GPR dataset according to a preset feature extraction strategy, calculates the relationship weights between different types of underground targets based on a preset feature similarity calculation strategy and target features, constructs a graph relation network based on the relation weights and adjacency matrix, enhances the target features based on the graph relation network to obtain enhanced features, fuses the enhanced features corresponding to the training set with the target features and inputs them into a preset prediction model to obtain the corresponding prediction results, trains and optimizes the prediction model based on the prediction results, the true results and the set loss function, and then inputs the test set into the optimized prediction model for evaluation. After reaching the preset evaluation conditions, the target prediction model is obtained. After obtaining the target prediction model, any newly acquired detection data from any GPR is input into the target prediction model to obtain the underground target detection result. This invention constructs a co-occurrence relation matrix and a graph relation network, and enhances and fuses the target features. It considers not only the characteristics of the pavement underground disease itself, but also the environmental characteristics of the disease and the feature distortion caused by the feature complexity and similarity between different underground target samples. Thus, domain knowledge is integrated into the prediction model as guiding prior knowledge, making the accuracy and reliability of the trained target prediction model higher and the detection effect better.

[0141] The present invention has been further described above with reference to specific embodiments. However, it should be understood that the specific description herein should not be construed as limiting the nature and scope of the present invention. Various modifications made to the above embodiments by those skilled in the art after reading this specification are all within the scope of protection of the present invention.

Claims

1. An airport pavement subsurface target detection method, characterized by, Including the following steps: Acquire detection data from several ground-penetrating radars, establish a ground-penetrating radar dataset based on the detection data, and divide the ground-penetrating radar dataset into a training set and a test set; A co-occurrence matrix among underground targets is established based on the ground-penetrating radar dataset. Construct an adjacency matrix between the underground targets based on the co-occurrence relationship matrix; According to a preset feature extraction strategy, target features of different types of underground targets are extracted and stored from the ground-penetrating radar dataset; The relationship weights between different types of underground targets are calculated based on the preset feature similarity calculation strategy and the target features, and a graph relationship network is constructed based on the relationship weights and the adjacency matrix. Based on the graph relation network, the target features are enhanced to obtain enhanced features; The enhanced features corresponding to the training set are fused with the target features and input into a preset prediction model to obtain the corresponding prediction results. The prediction model is then trained and optimized based on the prediction results, the true results, and the set loss function. The test set is input into the optimized prediction model for evaluation, and the target prediction model is obtained after the preset evaluation conditions are met. The newly acquired detection data from any ground-penetrating radar is input into the target prediction model to obtain the underground target detection results; The step of establishing a co-occurrence matrix among underground targets based on the ground-penetrating radar dataset includes: Acquire information about underground targets in a single-channel grayscale image; An empty N*N matrix is ​​created to record the co-occurrence relationships among the underground targets; N is the number of types of underground targets. Increment the co-occurrence value of every two underground targets in the single-channel grayscale image by one; Process all single-channel grayscale images in the ground-penetrating radar dataset according to the above steps to obtain the co-occurrence matrix of all underground targets; Constructing the adjacency matrix between the underground targets based on the co-occurrence relation matrix includes: The co-occurrence relation matrix is ​​used to obtain the association relation probability matrix through row normalization. Based on a preset association probability threshold, an adjacency matrix is ​​constructed between the underground targets; The step of extracting and storing target features of different types of underground targets from the ground-penetrating radar dataset according to a preset feature extraction strategy includes: The training set is input into a preset residual network to obtain disease features {C1, C2, C3, C4, C5} at different scales; the residual network contains several residual blocks, and each residual block consists of multiple convolutional layers and skip connections; The disease features are input into a preset pyramid network, which captures target features at different scales and semantic levels. The target features at different scales are fused to obtain target features {P2, P3, P4, P5}, and the last layer of deep features {C5} of the disease features is retained. The depth feature {C5} is convolved to obtain the target feature {P6} with a larger receptive field; The target features are input into the RPN network, which generates disease prior boxes through AnchorsGenerator. At the same time, the RPN network filters the input target features based on the disease prior boxes and extracts the target features that are most likely to be underground targets. The step of calculating the relationship weights between different types of underground targets based on a preset feature similarity calculation strategy and the target features, and constructing a graph relationship network based on the relationship weights and the adjacency matrix, includes: The dimensions of the target features are transformed to be consistent, and then used as nodes; According to The relationship weight between the adjacent graph nodes is generated according to the calculation of the different node associations ; is a nonlinear activation function, represents the importance of node j to node i, represents splicing and mapping the target feature of node i and the target feature of node j, represents the target feature corresponding to node i, represents the target feature corresponding to node j; Using the adjacency matrix as the adjacency matrix of the graph, the relation weights are... The edge weights of the graph are used to construct the association relationships, thus obtaining the graph relationship network. The step of enhancing the target features based on the graph relation network to obtain enhanced features includes: according to Calculate the weight values ​​of adjacent nodes and input them into the graph relationship network as propagation weights; This indicates the degree of association between node j and node i. Indicates the first There are nodes that are associated with node i. This represents the number of N nodes that are associated with node i; pass Calculate the enhanced features; This represents the enhanced feature corresponding to node i. This represents the target feature corresponding to node j.

2. The method for detecting underground targets on airport pavements as described in claim 1, characterized in that, The process of acquiring detection data from several ground-penetrating radars, establishing a ground-penetrating radar dataset based on the detection data, and dividing the ground-penetrating radar dataset into a training set and a test set includes: Acquire B-scan image data from several ground-penetrating radars; The B-scan image data is cropped into a single-channel grayscale image of a set size; Obtain manual annotations for the single-channel grayscale image; the manual annotations are used to mark underground targets in the single-channel grayscale image; The ground-penetrating radar dataset is established based on the single-channel grayscale image and the manual annotation. The ground-penetrating radar dataset is divided into a training set and a test set according to a random classification algorithm.

3. The method for detecting underground targets on airport pavements as described in claim 2, characterized in that, The step of training and optimizing the prediction model based on the predicted results, the actual results, and the set loss function includes: The loss function is: ,in, N cls For the number of categories, N reg The number of detection boxes; The classification loss is calculated as follows: , The regression loss is calculated as follows: ,in, , For input variables, and The prediction result, and For the true result, To set the coefficients.

4. The method for detecting underground targets on an airport pavement as described in claim 3, characterized in that, The step of inputting the test set into the optimized prediction model for evaluation includes: Evaluation indicators include: Accuracy: ; This indicates the number of times the prediction was correct. Indicates the number of incorrect predictions; Recall rate: ; Indicates the target number of missed detections; mAP value: ; This represents the prediction accuracy for each type of target. F1 value: .

5. An electronic device, comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to implement the method of any one of claims 1 to 4 through the computer program.

Citation Information

Patent Citations

  • Radar target detection method based on multi-scale feature fusion

    CN115019141A

  • Aerial photography target detection method based on composite backbone network and multiple prediction heads

    CN115035429A