GPR-HITBZ labeling method based on road cavity defect image feature region division

Abstract

The invention discloses a GPR-HITBZ marking method based on road cavity defect image feature region division, belongs to the field of road cavity detection, and aims to solve the problems that an existing method is insufficient in marking precision, non-uniform in standard, difficult to accurately reflect electromagnetic response features of cavity defects and low in road cavity defect recognition accuracy. The method comprises the following steps: firstly, obtaining road cavity defect image data and determining a corresponding data set, carrying out feature region division on cavity defect features in the data set and determining a corresponding training data set, and carrying out iterative training on a target detection model by utilizing the training data set to obtain corresponding detection values; and determining the process of the GPR-HITBZ labeling method as an optimal labeling method through data information in the detection value.

Description

Technical Field

[0001] The present invention specifically relates to a GPR-HITBZ annotation method based on the division of image feature regions of road cavity defects, and belongs to the field of road cavity detection. Background Art

[0002] During the long-term service of roads, affected by factors such as groundwater erosion, vehicle load impact, and construction quality defects, cavity defects are likely to form in the asphalt road base. Such defects are characterized by strong concealment and progressive development, and are difficult to detect by manual inspection in the initial stage. GPR (Ground Penetrating Radar), as a non-destructive detection technology based on the principle of electromagnetic wave reflection, has shown advantages in rapid detection in the field of road hidden defect identification. However, the current GPR data processing process relies on an experience-driven interpretation mode of professionals, and its complexity and interpretation difficulties seriously restrict the real-time diagnosis and quantitative evaluation capabilities of hidden defects in the intelligent operation and maintenance system of road infrastructure.

[0003] The current research focus is on improving the detection accuracy by comparing model architectures (multi-stage and single-stage detectors) and their optimization strategies. Among them, single-stage detectors are more suitable for engineering practice due to their high-efficiency inference characteristics. The YOLO series, as a typical single-stage detector, has made remarkable progress. However, existing research has not clearly revealed which YOLO framework has the best applicability in the detection of cavity defects in asphalt road ground penetrating radar images. On the one hand, existing model training generally uses pre-trained weights based on general datasets to accelerate convergence. Such general datasets lack the unique electromagnetic wave reflection characteristics of ground penetrating radar images, and the applicability of their transfer learning to road cavity detection still needs to be systematically verified. On the other hand, the performance of object detection models highly depends on the annotation quality. Since the cavity defects in asphalt roads show weak differences between the target body and background features in radar images, there are large subjective biases in the current annotation methods.

[0004] Therefore, there is an urgent need for a method for annotating the features of ground penetrating radar images of cavity defects to solve the above problems. Summary of the Invention

[0005] To solve the problems mentioned in the above background art, the purpose of the present invention is to provide a GPR-HITBZ annotation method based on the division of image feature regions of road cavity defects.

[0006] The GPR-HITBZ annotation method based on the division of image feature regions of road cavity defects includes the following steps:

[0007] The GPR-HITBZ annotation method is a process of obtaining road cavity defect image data and determining the corresponding data set, dividing the cavity defect features in the data set into feature regions and determining the corresponding training data set, iteratively training the target detection model with the training data set to obtain the corresponding detection values, and determining the GPR-HITBZ annotation method through the data information in the detection values.

[0008] Preferably, the determination of the GPR-HITBZ annotation method includes the following steps:

[0009] S1. Collect road cavity defect image data based on the three-dimensional ground penetrating radar device GPR, and perform verification processing on the collected cavity defect image data to determine the data set of the road cavity defect image data;

[0010] S2. Divide the cavity defect features in the cavity defect image data set into feature regions, process the cavity defect image data set after the feature region division with multiple annotation methods, and generate multiple groups of corresponding cavity defect image training data sets;

[0011] S3. Iteratively train the target detection model with multiple groups of training data sets to obtain the AP values corresponding to each annotation method;

[0012] S4. Obtain the corresponding harmonic mean F1 score under each annotation method according to the following formula:

[0013]

[0014] In the above formula, P represents precision, and R represents recall;

[0015] S5. Draw the F1 score curve and determine the data information of the highest point of the F1 score curve corresponding to each annotation method;

[0016] S6. Determine the GPR-HITBZ annotation method according to the AP value and the data information of the highest point of the F1 score curve.

[0017] Preferably, the division of the cavity defect features in the cavity defect image data set includes: dividing the cavity defect features into five cavity defect feature regions, namely region 1, region 2, region 3, region 4, and region 5.

[0018] Preferably, region 1 represents the main feature area of the cavity defect, region 2 represents the area above the main feature area, region 3 represents the area to the left of the main feature area, region 4 represents the area to the right of the main feature area and is symmetric with region 3, and region 5 represents the area below the main feature area and is symmetric with region 2.

[0019] Preferably, after obtaining the AP values corresponding to the respective annotation methods, it further includes: determining one or more annotation methods with the highest AP value among the respective annotation methods.

[0020] Preferably, the data information of the highest point of the F1-score curve corresponding to each annotation method obtained by plotting the F1-score curve includes: confidence threshold, F1-score value.

[0021] Preferably, further processing the dataset of void defect images after feature region division using multiple annotation methods further includes: calibrating the feature regions of the void defect images according to the annotation information to generate annotation images, storing the annotation results in the PASCAL VOC standard XML format, and performing coordinate conversion through the following formula:

[0022]

[0023] Converting to a txt format file required for model training.

[0024] Preferably, the annotation information includes: GPR image size, coordinates of the diagonal vertices of the rectangular box, target category.

[0025] Preferably, the verification processing of the collected void defect image data includes:

[0026] A1. Performing data analysis on the collected radar data of void defect images to mark suspected void defect regions;

[0027] A2. Using a push-type two-dimensional ground penetrating radar to perform a recheck on the suspected void defect regions;

[0028] A3. Performing drilling operations on the regions with abnormal images shown by the two-dimensional ground penetrating radar, and using an endoscope for visual verification;

[0029] A4. Confirming that the image is a void defect when the endoscope image shows that the hole wall is intact and the area below is black.

[0030] Preferably, after the verification processing of the collected void defect image data to determine the dataset of road void defect image data, it further includes: dividing the valid data in the dataset into three parts of train, val, and test according to the ratio of 7:2:1.

[0031] Compared with the prior art, the beneficial effects of the present invention are:

[0032] 1. The present invention collects a large amount of data on asphalt road cavity defects through a three-dimensional ground penetrating radar device, conducts on-site secondary surveys and verifications on the cavity defect data, constructs an asphalt road cavity defect dataset, studies the characteristic regions of cavity defects, explores the influence of different cavity defect image annotation methods on model training, ensures the diversity and representativeness of data samples, alleviates the problem of insufficient data in this field, and divides the effective data into three parts: train, val, and test according to the ratio of 7:2:1. This division strategy not only ensures the generalization ability of model training but also provides a basis for optimizing subsequent data annotation methods.

[0033] 2. The present invention proposes an improved annotation scheme GPR-HITBZ. According to the characteristics of cavity defects, it divides them into five characteristic regions. Based on these five regions, this study designs a variety of different annotation schemes and quantitatively analyzes and compares the effects of each scheme in model training. Obtain the data values of the corresponding AP values and corresponding F1 scores under each annotation method, and comprehensively evaluate based on the data values of the AP value and F1 score. The GPR-HITBZ scheme performs the best. This scheme establishes an operable annotation process by defining the spatial mapping relationship between the bounding box and the cavity defect characteristic region, effectively improving the recognition accuracy of road cavity defects. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] For ease of explanation, the present invention is described in detail by the following specific embodiments and accompanying drawings.

[0035] Figure 1 FIG. is a flowchart of the GPR-HITBZ annotation method based on the division of the characteristic region of the road cavity defect image;

[0036] Figure 2 FIG. is a schematic diagram of the ground penetrating radar cavity defect image labeled by LabelImg;

[0037] Figure 3 FIG. is a flowchart of the conversion of the LabelImg annotation format file;

[0038] Figure 4 FIG. is a schematic diagram of the division of the cavity defect characteristic region;

[0039] Figure 5 FIG. is a schematic diagram of the division of the cavity defect characteristic region corresponding to different annotation methods;

[0040] Figure 6 FIG. is a graph of the AP values corresponding to the detection of cavity defects by different annotation methods;

[0041] Figure 7 FIG. is a schematic diagram of the F1 curve of different annotation methods;

[0042] Figure 8 FIG. is a partial cavity defect image and endoscope image. Detailed implementation manners

[0043] To make the objectives, technical solutions, and advantages of the present invention clearer and more understandable, the present invention will be described below through specific embodiments shown in the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, the descriptions of well-known structures and technologies are omitted to avoid unnecessarily confusing the concepts of the present invention.

[0044] Here, it should also be noted that in order to avoid obscuring the present invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the present invention are shown in the accompanying drawings, while other details less related to the present invention are omitted.

[0045] Detailed implementation manner 1: This embodiment only provides a preferred implementation manner. This implementation manner is a GPR-HITBZ annotation method based on the division of the feature region of the road cavity defect image. The GPR-HITBZ annotation method is to obtain the road cavity defect image data and determine the corresponding data set. Among them, this embodiment does not make specific limitations on the method, time, and quantity of obtaining the road cavity defect image data. This embodiment is applicable to a variety of scenarios for obtaining cavity defect images under the condition of ensuring data accuracy; the cavity defect features in the data set are divided into feature regions and the corresponding training data set is determined. Among them, this embodiment does not make specific limitations on the division method and the number of divided feature regions; the target detection model is iteratively trained using the training data set, and then the corresponding detection values are obtained. The process of the GPR-HITBZ annotation method is determined through the data information in the detection values. Among them, the detection values may include, but are not limited to, the harmonic mean F1 score corresponding to each annotation method and the AP value corresponding to each annotation method, etc. The data information corresponding to the detection values may be the F1 score, confidence threshold, etc.

[0046] Detailed implementation manner 2: This embodiment only provides a preferred implementation manner. Determining the GPR-HITBZ annotation method includes the following steps:

[0047] S1. Collect road cavity defect image data based on the three-dimensional ground penetrating radar device GPR, and perform verification processing on the collected cavity defect image data to determine the data set of the road cavity defect image data;

[0048] Among them, the three-dimensional ground penetrating radar realizes target detection based on the principle of the interaction between electromagnetic waves and underground media, and the characteristics of its reflected signals directly reflect the dielectric constant difference. The present invention provides a preferred three-dimensional radar system, which may but is not limited to adopting a multi-channel array three-dimensional radar system, equipped with a 16-element dipole antenna. Among them, the 16-element dipole antenna can be 8 transmit and 8 receive, and a 15-channel synchronous acquisition architecture is constructed. The specific technical parameters may but are not limited to the following parameters: the antenna configuration has a center frequency of 200 MHz, and the detection width is selected as 178 cm; the working parameters are that the time window range is selected as 1 - 80 ns, the channel spacing is 10 cm, and the maximum detection depth is selected as 4 m; the auxiliary system selects an integrated adjustable hydraulic lifting mechanism, and the vehicle-mounted stable platform ensures a dynamic acquisition rate of 80 km / h.

[0049] S2. Divide the cavity defect features in the cavity defect image dataset into feature regions, process the cavity defect image dataset after feature region division using multiple annotation methods, and generate multiple groups of corresponding cavity defect image training datasets;

[0050] The annotation method for the cavity defect image dataset may but is not limited to using annotation software such as LabelImg to accurately annotate the disease features, and use dark green rectangular frames to calibrate the outer contour of the target area with typical hyperbolic reflection wave features according to the annotation information to generate a standardized annotation image. As Figure 2 shown, the annotation results are stored in the PASCAL VOC standard XML format and coordinate conversion is performed to convert it into a txt format file required for model training, as Figure 3 shown, where the Class_id / 0 category specifically refers to the cavity defect;

[0051] Since the target body contour and background features generally show weak differences in the radar image, the recognition rate of the defect contour edge is relatively low; the cavity defect features in the cavity defect image dataset can be divided into feature regions. Among them, the division of the defect feature regions is not specifically limited, and the types of annotation methods for processing the cavity defect image dataset after feature region division are not limited. The present invention only provides a preferred region division method and a preferred number of annotation methods.

[0052] S3. Iteratively train the target detection model through multiple groups of training datasets to obtain the AP values corresponding to each annotation method;

[0053] Specifically, this application does not make specific limitations on the target detection model. As a classic single-stage detector, the performance evaluation of the YOLO (You Only Look Once) series models on the COCO dataset shows that their running efficiency is significantly better than traditional algorithms such as R-CNN. For road internal crack detection, comparative studies based on the YOLO series models show that the performance of YOLOv4 and YOLOv5 on small datasets has been significantly improved compared to YOLOv3, and the highest mAP of YOLOv5 reaches 94.39%. In terms of the detection of tie bars on cement concrete pavements, the comprehensive detection accuracy of the YOLOv4 model is better than variants such as YOLOv3, YOLOv3-tiny, and YOLOv4-tiny. Experimental results show that the detection accuracy of YOLOv5 on the enhanced dataset is better than models such as YOLOv3, R-CNN, and CenterNet. To improve the performance metrics of the target detection model, researchers have carried out systematic research from aspects such as loss function optimization and module improvement, greatly improving the inference ability of the model. By integrating the gradient boosting algorithm and the Bayesian hyperparameter optimization strategy, the improvement of the GPR detection accuracy for road water damage is achieved. Experimental results show that the recognition accuracy of the optimized model has increased by 1.5%. In terms of road cavity recognition, researchers introduced the SwinTransformer architecture into the improvement of models such as YOLOv3 and YOLOv5. Comparative studies show that the improved YOLOv5 model has the best performance, effectively reducing the false positive rate of other models in recognizing cavities. Therefore, in this embodiment, it is preferably to perform iterative training on the YOLOv5 model, but this application is not limited to using the YOLOv5 model and is also applicable to other models or better models.

[0054] S4. Obtain the corresponding harmonic mean F1 score under each annotation method according to the following formula:

[0055]

[0056] In the above formula, P represents precision, and R represents recall;

[0057] S5. Draw the F1 score curve and determine the data information of the highest point of the F1 score curve corresponding to each annotation method;

[0058] S6. Determine the GPR-HITBZ annotation method according to the AP value and the data information of the highest point of the F1 score curve.

[0059] Specific Embodiment 3: This embodiment only provides a preferred implementation manner. The feature region division of the cavity defect features in the cavity defect image dataset includes: dividing the cavity defect features into five cavity defect feature regions, namely Region 1, Region 2, Region 3, Region 4, and Region 5;

[0060] In this application, through systematic analysis of a large number of ground-penetrating radar images of asphalt road cavity defects, the cavity defect features are divided into five feature regions. As Figure 4 shown, where region 1 represents the main feature region of the cavity defect, with the most significant and clear features. Region 2 represents the region above the main feature region, corresponding to the top feature with relatively weak multiple waves. Region 3 represents the region to the left of the main feature region. Region 4 represents the region to the right of the main feature region, which is symmetrically distributed with region 3. Region 5 represents the region below the main feature region, which is symmetrically distributed with region 2. In this embodiment, a systematic experiment can be set up based on the method of controlling variables. By quantitatively analyzing various differential annotation methods, a standardized annotation paradigm for ground-penetrating radar images of asphalt road cavity defects is constructed;

[0061] The region division method provided in this embodiment is only a preferred division method, not limited to only dividing into 5 regions, and the naming method of the divided regions is not limited to only including the region naming methods in the above embodiments.

[0062] Specific Embodiment 4: This embodiment only provides a preferred implementation method. This embodiment provides a setting of a specific annotation method. For the 5 cavity defect feature regions divided above, eight different annotation methods are differentially annotated as follows. As Figure 5 shown, Figure 5 in (a), the annotation box contains the entire region of the cavity defect feature. In (b), the annotation box does not contain region 5 of the cavity defect feature. In (c), the annotation box does not contain regions 3 and 4 of the cavity defect feature. In (d), the annotation box does not contain regions 3, 4, and 5 of the cavity defect feature. In (e), the annotation box does not contain region 2 of the cavity defect feature. In (f), the annotation box does not contain regions 2 and 5 of the cavity defect feature. In (g), the annotation box does not contain regions 2, 3, and 4 of the cavity defect feature. In (h), the annotation box does not contain regions 2, 3, 4, and 5 of the cavity defect feature;

[0063] Based on the cavity defect data set obtained through on-site secondary verification, this study processes it using the aforementioned eight annotation methods, and finally generates eight groups of training data sets. As shown in Table 1. The annotation methods in Table 1 are in one-to-one correspondence with Figure 5 the annotation methods in;

[0064] Number of cavity defect images annotated by each annotation method in Table 1

[0065]

[0066] This embodiment provides a preferred method of selecting YOLOv5 as the object detection framework. The model is iteratively trained using the cavity defect data sets generated by the above eight annotation methods, and the number of iterations is 300 times for each. Finally, the AP values corresponding to each annotation method are obtained. As Figure 6 shown.

[0067] Specific Embodiment 5: This embodiment only provides a preferred implementation manner. After obtaining the AP values corresponding to each annotation method, it further includes: determining one or more annotation methods with the highest AP value among each annotation method.

[0068] This embodiment provides a preferred specific implementation manner. For the above eight different annotation methods, as Figure 6 shown, among the eight different annotation methods, the AP value of method c is the lowest, and the detection effect on hole defects is the worst. The AP values of the remaining seven methods are the same, indicating that their training effects are consistent. Therefore, the optimal annotation method cannot be distinguished only by the AP value. The present application further introduces the F1 score as a supplementary evaluation index. The F1 score is the harmonic mean of the precision and recall rates, which can comprehensively measure the balance between the two. The higher the F1 score, the stronger the trade-off ability of the model between the precision and recall rates.

[0069] Specific Embodiment 6: This embodiment only provides a preferred implementation manner. The data information of the highest point of the F1 score curve corresponding to each annotation method obtained by plotting the F1 score curve includes: confidence threshold, F1 score value.

[0070] Specifically, for the above eight annotation methods, the F1 score curves corresponding to each annotation method are as Figure 7As shown, in the F1-score curve, "allclasses1.00at0.852" of Method b indicates that when the confidence threshold is 0.852, the average F1-score of the model on the cavity defect category reaches 1.00, indicating that the classification performance is optimal at this time, and P and R reach a high balance. Generally, a higher confidence threshold will screen out more accurate prediction results, but may reduce the recall rate. However, Method b can still maintain an F1-score of 1.00 when the threshold is 0.852, indicating that the model trained by this annotation method can still avoid missed detections and false detections under strict screening conditions. In contrast, for other annotation methods, either the confidence threshold is lower than that of Method b, or the F1-score is inferior to that of Method b. Among the eight annotation methods, the detection effects of Methods g and f are relatively the worst. The annotation methods that include the top feature region of the cavity defect have better performance than those that do not include this region. Notably, the scheme that only annotates the main feature region of the cavity defect performs the worst, which confirms that including the peripheral feature region of the cavity defect can effectively improve the model's detection performance (the F1-score is increased by approximately 30% at most). Based on the comprehensive evaluation of AP and F1-score, Annotation Scheme b (the annotation box does not include the cavity defect feature region 5) is proven to be the optimal choice. In this embodiment, this annotation method can be named GPR-HITBZ. For other different annotation methods, the optimal annotation method can be determined through the comprehensive evaluation of AP and F1-score and named GPR-HITBZ. In practical applications, the GPR-HITBZ annotation method can be referred to to further improve the detection accuracy of road hidden defects from the aspect of data annotation quality.

[0071] Specific Embodiment Seven: This embodiment only provides a preferred implementation manner. The processing of the cavity defect image dataset after feature region division using multiple annotation methods further includes: calibrating the feature regions of the cavity defect images according to the annotation information to generate annotation images. The annotation results are stored in the PASCAL VOC standard XML format and coordinate conversion is performed through the following formula:

[0072]

[0073]

[0074] Converted into the txt format file required for model training.

[0075] Specific Embodiment Eight: This embodiment only provides a preferred implementation manner. Specifically, annotation software such as LabelImg is used to accurately annotate the disease characteristics. The outer contour of the target region with typical hyperbolic reflection wave characteristics is calibrated using dark green rectangular boxes to generate standardized annotation images, where the annotation information includes but is not limited to metadata such as the GPR image size, the diagonal vertex coordinates of the rectangular box, and the target category.

[0076] Specific implementation method 9: This embodiment only provides a preferred implementation method. Specifically, the verification process of the collected cavity defect image data includes:

[0077] A1. Data screening: Professional technicians analyze the collected cavity defect image radar data and mark suspected cavity defect areas;

[0078] A2. On-site retest: According to the guidance of the radar positioning system, the suspected area is blocked and a hand-pushed 2D ground penetrating radar is used to recheck the suspected cavity defect area;

[0079] A3. Drilling verification: Drill holes in areas where abnormalities are detected in 2D GPR images, and use an endoscope for visual verification.

[0080] A4. Confirm that the endoscopic image showing an intact hole wall and a black area underneath is a cavity defect;

[0081] This study went through a rigorous on-site secondary verification process and ultimately obtained a large amount of ground penetrating radar image data of asphalt road cavity defects. Figure 8 The radar images of some cavity defect samples and their corresponding endoscopic verification results are displayed, intuitively presenting the correspondence between the electromagnetic wave reflection characteristics and the physical morphology of road cavity defects.

[0082] Specific Embodiment 10: This example provides only a preferred implementation. Specifically, based on the collected data, this application successfully constructs a dataset of asphalt road cavity defects. To meet the needs of deep learning research, the valid data is divided into three parts: train, val, and test in a ratio of 7:2:1. This partitioning strategy not only ensures the generalization ability of model training but also provides a foundation for optimizing subsequent data annotation methods. The cavity defect image dataset is shown in Table 2.

[0083] Table 2 Asphalt road cavity defect ground penetrating radar dataset

[0084]

[0085] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A GPR-HITBZ annotation method based on the division of the characteristic region of the road cavity defect image, characterized in that: The GPR-HITBZ annotation method is a process of obtaining road cavity defect image data and determining the corresponding data set, dividing the cavity defect features in the data set into feature regions and determining the corresponding training data set, iteratively training the target detection model with the training data set to obtain the corresponding detection values, and determining the GPR-HITBZ annotation method through the data information in the detection values.

2. The GPR-HITBZ annotation method based on the division of the characteristic region of the road cavity defect image according to claim 1, wherein: The determination of the GPR-HITBZ annotation method includes the following steps: S1. Collect road cavity defect image data based on the three-dimensional ground penetrating radar device GPR, and perform verification processing on the collected cavity defect image data to determine the data set of the road cavity defect image data; S2. Divide the cavity defect features in the cavity defect image data set into feature regions, process the cavity defect image data set after the feature region division with multiple annotation methods, and generate multiple groups of corresponding cavity defect image training data sets; S3. Iteratively train the target detection model with multiple groups of training data sets to obtain the AP values corresponding to each annotation method; S4. Obtain the harmonic mean F1 score corresponding to each annotation method according to the following formula: In the above formula, P represents precision, and R represents recall; S5. Draw the F1 score curve and determine the data information of the highest point of the F1 score curve corresponding to each annotation method; S6. Determine the GPR-HITBZ annotation method according to the AP value and the data information of the highest point of the F1 score curve.

3. The GPR-HITBZ annotation method based on the division of the image feature region of road cavity defects according to claim 2, wherein, It is characterized in that: The division of the cavity defect features in the cavity defect image data set into feature regions includes: dividing the cavity defect features into five cavity defect feature regions, namely region 1, region 2, region 3, region 4, and region 5.

4. The GPR-HITBZ annotation method based on the division of the characteristic region of the road cavity defect image according to claim 3, wherein: Region 1 represents the main feature area of the cavity defect, region 2 represents the area above the main feature area, region 3 represents the area on the left side of the main feature area, region 4 represents the area on the right side of the main feature area, which is symmetric with region 3, and region 5 represents the area below the main feature area, which is symmetric with region 2.

5. The GPR-HITBZ annotation method based on the division of the characteristic region of the road cavity defect image according to claim 4, characterized in that: After obtaining the AP values corresponding to each annotation method, it further includes: determining one or more annotation methods with the highest AP value among each annotation method.

6. The GPR-HITBZ annotation method based on the division of the characteristic region of the road cavity defect image according to claim 5, wherein: The drawing of the F1 score curve and determining the data information of the highest point of the F1 score curve corresponding to each annotation method include: confidence threshold, F1 score value.

7. The GPR-HITBZ annotation method based on the division of the characteristic region of the road cavity defect image according to any one of claims 2 to 6, characterized in that: The processing of the cavity defect image data set after the feature region division with multiple annotation methods further includes: calibrating the cavity defect image feature region according to the annotation information to generate an annotation image, storing the annotation result in the PASCAL VOC standard XML format, and performing coordinate conversion through the following formula: Convert it to the txt format file required for model training.

8. The GPR-HITBZ annotation method based on the division of the feature regions of road cavity defect images according to claim 7, characterized in that: The annotation information includes: GPR image size, rectangular box diagonal vertex coordinates, target category.

9. The GPR-HITBZ annotation method based on the division of the characteristic region of the road cavity defect image according to claim 2, wherein: The verification processing of the collected cavity defect image data includes: A1. Perform data analysis on the collected cavity defect image radar data and mark the suspected cavity defect regions; A2. Use a hand-pushed two-dimensional ground-penetrating radar to recheck and detect the suspected cavity defect area; A3. Conduct drilling operations on the areas where the images collected by the two-dimensional ground-penetrating radar show abnormalities, and use an endoscope for visual verification; A4. Confirm that the image is a cavity defect when the endoscope image shows that the hole wall is intact and there is a black area below.

10. The GPR-HITBZ annotation method based on the division of the characteristic region of the road cavity defect image according to claim 9, wherein: After the above-mentioned verification processing of the collected cavity defect image data and determining the data set of the road cavity defect image data, it further includes: dividing the valid data in the data set into three parts of train, val, and test according to the ratio of 7:2:1.

Images