Automatic labeling method for deep learning YOLO format sample of ground penetrating radar

Abstract

The invention provides a ground penetrating radar deep learning YOLO format sample automatic labeling method, comprising the following steps: S1, saving or saving road hidden disease ground penetrating radar artificial interpretation results in a unified data format, each ground penetrating radar data corresponding to an artificial interpretation result file; s2, reading ground penetrating radar data and corresponding artificial interpretation results, sequentially retrieving starting point coordinates and terminal point coordinates of each disease in the artificial interpretation results in the ground penetrating radar data, and delineating a ground penetrating radar data space corresponding to the road hidden disease; s3, taking a ground penetrating radar data space where the disease is located as a center, properly reserving surrounding ground penetrating radar data as a sample background, and intercepting disease data and surrounding background data as disease sample data; s4, self-adaptive gain adjustment is carried out; and S5, carrying out normalization [0 < 1 >] on the whole sample data space, calculating central point coordinates (x, y) of the disease and width and height of the central point coordinates, and automatically generating a ground penetrating radar deep learning sample conforming to a YOLO (Yttrium-Learning-Oriented Yttrium-Learning-Oriented Yttrium-Learning-Oriented Yttrium-Learning-Oriented Yttrium-Learning-Oriented) format.

Description

Technical Field

[0001] This invention belongs to the technical field of ground penetrating radar, specifically an automatic annotation method for YOLO format samples from deep learning ground penetrating radar. Background Technology

[0002] Ground penetrating radar is a method that utilizes 10 6 -10 9 Ground-penetrating radar (GPR) is a geophysical exploration method that uses high-frequency electromagnetic waves (Hz) to visualize the internal distribution patterns of underground media. It boasts advantages such as high precision, high efficiency, high speed, continuous non-destructive imaging, real-time imaging, and intuitive results. It is currently the primary technology for long-distance, engineered, and highly efficient non-destructive detection of hidden road defects. However, the interpretation of GPR data and the identification of hidden road defects still largely rely on manual, experience-based interpretation, which is time-consuming, labor-intensive, and highly subjective. In long-distance detection of hidden road defects, GPR primarily employs the common-offset profiling method to scan and image the road structure. This eliminates the need for complex data processing, generating radar profiles corresponding to the distribution patterns of underground geological structures. Therefore, compared to other geophysical methods, intelligent identification and classification of GPR profiles based on deep learning has significant advantages.

[0003] In the creation of deep learning samples for ground-penetrating radar (GPR) of hidden road defects, manual annotation using open-source annotation software such as LabelImg, Labelme, and VOTT is commonly employed. This method is time-consuming, labor-intensive, slow, and inefficient, making it difficult to meet the large sample size requirements of GPR deep learning. Therefore, efficiently utilizing existing manual interpretation results of hidden road defects is one of the effective means to overcome the current shortage of GPR deep learning samples and improve the accuracy of deep learning. Summary of the Invention

[0004] To address the shortcomings in the aforementioned background technology, this invention proposes an automatic annotation method for YOLO format samples of ground-penetrating radar deep learning, which solves the problem of low efficiency caused by the need for manual annotation of learning samples in the production of ground-penetrating radar deep learning samples for hidden road defects.

[0005] The technical solution of this application is as follows:

[0006] An automatic annotation method for YOLO format samples from ground-penetrating radar deep learning includes the following steps:

[0007] S1: Save or save the manual interpretation results of ground-penetrating radar for hidden road defects in a unified data format. Each ground-penetrating radar data corresponds to a manual interpretation result file. In order to facilitate retrieval and automatic matching, the manual interpretation results should have the same file name as the ground-penetrating radar data, but with a different file extension. Each ground-penetrating radar data corresponds to a manual interpretation result txt file.

[0008] S2: Read the ground-penetrating radar data and its corresponding manual interpretation results. In the ground-penetrating radar data, sequentially search for the starting coordinates and ending coordinates of each defect in the manual interpretation results to delineate the ground-penetrating radar data space corresponding to the hidden road defects.

[0009] S3: Taking the ground-penetrating radar data space where the disease is located as the center, appropriately retain the surrounding ground-penetrating radar data as the sample background, and extract the disease data and the surrounding background data as the disease sample data.

[0010] S4: Adaptive gain adjustment;

[0011] S5: Normalize the entire sample data space [0 1], and calculate the center point coordinates (x, y) of the disease and its horizontal and vertical dimension ratios (width, height), and automatically generate ground-penetrating radar deep learning samples that conform to the YOLO format.

[0012] This invention designs a simple and easy-to-understand data format to store manually interpreted results, establishes a mapping relationship between hidden road defects and their ground-penetrating radar (GPR) data, and enables autonomous retrieval, location, cutting, and automatic annotation of GPR data. Employing an automatic annotation method and adaptive gain adjustment for deep learning samples of hidden road defects using GPR, it can quickly generate deep learning samples, significantly saving sample annotation time and enabling the generation of large batches of GPR deep learning samples in a short time. In long-distance detection of hidden road defects, GPR mainly uses the common offset profiling method to scan and image the road structure. Without complex data processing, it can obtain radar profile maps corresponding to the distribution of underground geological structures. The intelligent identification and classification of GPR profiles based on deep learning has significant advantages.

[0013] Furthermore, the manually interpreted result file in S1 has the same filename as the ground-penetrating radar data, but with a different file extension.

[0014] Furthermore, the data format of the manually interpreted result file is as follows:

[0015] Line 1: The total number of defects n contained in the ground-penetrating radar data;

[0016] Lines 2 to 6: These are the five parameters corresponding to the first disease. The first four parameters are the number of the starting point of the disease and the number of sampling points, i.e., the starting point coordinates; the number of the ending point of the disease and the number of sampling points, i.e., the ending point coordinates; and the fifth parameter is the disease type or the disease type number.

[0017] Rows (n-1)×5+2 to 6: five parameters corresponding to the nth disease.

[0018] Furthermore, the specific steps of S4 are as follows:

[0019] S4.1: Extract the mean amplitude A1 of the phase axis of the first half-cycle of the early signal from the ground penetrating radar and the corresponding number of sampling points m;

[0020] S4.2: Extract and calculate the mean amplitude within the CD time window as the mean amplitude A2 of the deep reflected wave;

[0021] S4.3: Calculate the ratio of A1 to A2. When the ratio is less than the threshold, it is determined that the radar data already contains gain information and no gain compensation is needed. When the ratio is greater than the threshold, it is determined that the radar data does not contain gain information and amplitude compensation needs to be performed using an exponential gain function.

[0022] Furthermore, the expression for the exponential gain function is:

[0023]

[0024] A i This represents the original amplitude value corresponding to the i-th sampling point in the raw ground-penetrating radar data. For A i The amplitude value after gain amplification, where v is the average propagation speed of the electromagnetic wave in the underground medium. t i Let be the arrival time of the reflected wave corresponding to the i-th sampling point.

[0025] Furthermore, the expression for the exponential gain factor α is:

[0026]

[0027] κ is a constant.

[0028] Furthermore, the arrival time t of the reflected wave corresponding to the i-th sampling point i The expression is:

[0029] t i =(im)·△t

[0030] △t is the time sampling interval, and m is the number of sampling points. When i≤m, the signal corresponding to sampling point i is earlier than or equal to the arrival time of the early signal of the ground penetrating radar. In this case, no amplitude compensation is required, so the gain function value is set to be equal to 1, and the radar wave amplitude value remains unchanged. When i>m, the signal corresponding to sampling point i is later than the arrival time of the early signal of the ground penetrating radar. In this case, the exponential gain function is used to perform amplitude compensation on the radar wave.

[0031] Furthermore, the specific steps of S5 are as follows:

[0032] S5.1: Batch import N ground-penetrating radar files;

[0033] S5.2: Sequentially query whether the current radar file n (n = 1, 2, 3, ... N) has a corresponding manual interpretation result file. If the radar file does not have a corresponding manual interpretation result, then sequentially query whether the next radar file (n = n + 1) has a corresponding manual interpretation result, and so on. If the radar file has a corresponding manual interpretation result, then proceed to step S6.3; when n > N, execute step S5.1;

[0034] S5.3: Read the total number M of hidden road defects from the manual interpretation results corresponding to the current radar file;

[0035] S5.4: Read the starting point, ending point coordinates, and disease type or number of the m-th disease (m = 1, 2, 3, ... M) in the manual interpretation results in sequence. Find the radar channel number and sampling point number of the starting point and the radar channel number and sampling point number of the ending point of the disease in the radar file. Expand the total number of channels of the sample data to the left and right with the disease as the center, and expand it to the corresponding integer multiple in the upward and downward directions to include all sampling points. Use the expanded radar data as the sample data for deep learning. That is, the sample data includes both the radar data corresponding to the disease and the background data around the disease. When m > M, execute step S5.2.

[0036] S5.5: Perform adaptive gain adjustment and normalization on the sample data, automatically generate radar images and tag pairs, and set m = m + 1, then execute step S5.4.

[0037] Furthermore, the deep learning samples include one-to-one corresponding radar images and label pairs, with the radar image filename being the same as the label filename, and placed in the file paths / image / *.jpg and / label / *.txt respectively.

[0038] Furthermore, when there are n defective targets in a radar image, the tag file is labeled with n lines.

[0039] The specific beneficial effects of this invention include:

[0040] 1. This invention designs a simple and easy-to-understand data format to save the results of manual interpretation, establish a mapping relationship between hidden road defects and their ground-penetrating radar data, and realize autonomous retrieval, positioning, cutting and automatic annotation of ground-penetrating radar data;

[0041] 2. This invention adopts an automatic annotation method and adaptive gain adjustment for deep learning samples of ground-penetrating radar for hidden road defects, which can quickly generate deep learning samples, save sample annotation time on a large scale, and generate ground-penetrating radar deep learning samples in a short time and in large batches.

[0042] 3. In the detection of hidden defects on long-distance roads, the ground-penetrating radar mainly uses the common offset profile method to scan and image the road structure. Without the need for complex data processing, it can obtain radar profile maps corresponding to the distribution of underground geological structures. The intelligent identification and classification of ground-penetrating radar profiles based on deep learning has significant advantages. Attached Figure Description

[0043] To more clearly illustrate the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0044] Figure 1 This is a flowchart illustrating the automatic generation of deep learning samples for ground-penetrating radar in this invention.

[0045] Figure 2 This is the data format of the manually interpreted result file in this invention;

[0046] Figure 3 This is a flowchart of the automatic annotation process for deep learning samples (YOLO format) of ground-penetrating radar in this invention;

[0047] Figure 4 This is a record of raw ground-penetrating radar data without any gain, acquired using a ground-coupled shielded antenna in this invention.

[0048] Figure 5 This is the curve of the exponential gain function in this invention;

[0049] Figure 6 This is a comparison diagram of single-channel waveforms before and after the exponential gain function in this invention;

[0050] Figure 7 This is the post-gain ground-penetrating radar profile in this invention. Detailed Implementation

[0051] The technical solutions of 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 scope of protection of the present invention.

[0052] An automatic annotation method for YOLO format samples from ground-penetrating radar deep learning includes the following steps:

[0053] S1: Save or save the manual interpretation results of ground-penetrating radar for hidden road defects in a unified data format. Each ground-penetrating radar data corresponds to a manual interpretation result file. In order to facilitate retrieval and automatic matching, the manual interpretation results should have the same file name as the ground-penetrating radar data, but with a different file extension. Each ground-penetrating radar data corresponds to a manual interpretation result txt file.

[0054] S2: Read the ground-penetrating radar data and its corresponding manual interpretation results. In the ground-penetrating radar data, sequentially search for the starting coordinates and ending coordinates of each defect in the manual interpretation results to delineate the ground-penetrating radar data space corresponding to the hidden road defects.

[0055] S3: Taking the ground-penetrating radar data space where the disease is located as the center, appropriately retain the surrounding ground-penetrating radar data as the sample background, and extract the disease data and the surrounding background data as the disease sample data.

[0056] S4: Adaptive gain adjustment;

[0057] S5: Normalize the entire sample data space [0 1], and calculate the center point coordinates (x, y) of the disease and its horizontal and vertical dimension ratios (width, height), and automatically generate ground-penetrating radar deep learning samples that conform to the YOLO format.

[0058] This invention designs a simple and easy-to-understand data format to store manually interpreted results, establishes a mapping relationship between hidden road defects and their ground-penetrating radar (GPR) data, and enables autonomous retrieval, location, cutting, and automatic annotation of GPR data. Employing an automatic annotation method and adaptive gain adjustment for deep learning samples of hidden road defects using GPR, it can quickly generate deep learning samples, significantly saving sample annotation time and enabling the generation of large batches of GPR deep learning samples in a short time. In long-distance detection of hidden road defects, GPR mainly uses the common offset profiling method to scan and image the road structure. Without complex data processing, it can obtain radar profile maps corresponding to the distribution of underground geological structures. The intelligent identification and classification of GPR profiles based on deep learning has significant advantages.

[0059] Based on the above implementation method, as a preferred implementation method, the manually interpreted result file and the ground penetrating radar data have the same file name but different file extensions.

[0060] Based on the above embodiments, as a preferred embodiment, such as... Figure 2 As shown, the data format of the manually interpreted result file is as follows:

[0061] Line 1: The total number of defects n contained in the ground-penetrating radar data;

[0062] Lines 2 to 6: These are the five parameters corresponding to the first disease. The first four parameters are the number of the starting point of the disease and the number of sampling points, i.e., the starting point coordinates; the number of the ending point of the disease and the number of sampling points, i.e., the ending point coordinates; and the fifth parameter is the disease type or the disease type number.

[0063] Rows (n-1)×5+2 to 6: five parameters corresponding to the nth disease.

[0064] Based on the above implementation method, as a preferred implementation method, the specific steps of S4 are as follows:

[0065] S4.1: Extract the mean amplitude A1 of the phase axis of the first half-cycle of the early signal from the ground penetrating radar and the corresponding number of sampling points m;

[0066] S4.2: Extract and calculate the average amplitude within the CD time window range as the average amplitude A2 of the deep reflected wave; specifically, the CD time window range is 0.5-0.625;

[0067] S4.3: Calculate the ratio of A1 to A2. When the ratio is less than the threshold, it is determined that the radar data already contains gain information and no gain compensation is needed. When the ratio is greater than the threshold, it is determined that the radar data does not contain gain information and amplitude compensation needs to be performed using an exponential gain function.

[0068] Specifically, electromagnetic waves propagating in underground media are affected by factors such as geometric diffusion, absorption attenuation, scattering loss, and transmission loss. This results in stronger amplitudes for early radar signals from near the surface and weaker amplitudes for later radar signals from deeper layers in the raw ground-penetrating radar (GPR) records. Therefore, during GPR data acquisition, to highlight the radar wave response characteristics of deep targets, constant gain, automatic gain control, or manual gain control are typically used to compensate for the weak radar reflection signals from deeper layers. Constant gain and automatic gain control do not conform to the actual physical laws of GPR amplitude attenuation, while manual gain control can perform similar exponential gain control, but it is not precise or accurate enough. Furthermore, in some commercial GPR systems, gain adjustment is only used for real-time on-site display, and the raw GPR records do not contain any gain information. Other commercial GPR systems save the gain function and the raw GPR records separately, so that in later GPR data processing, dedicated software can be used to extract the gain adjustment function set during acquisition and perform amplitude compensation on the radar records. Still other commercial GPR systems save the gain-adjusted radar signal as the raw radar record.

[0069] In the process of automatically generating deep learning samples from ground-penetrating radar (GPR), it is necessary to deal with tens of thousands of raw GPR data from different manufacturers, with different data formats, and different application scenarios. It is impossible to verify the rationality of the attenuation coefficient settings for each one. Only by using an adaptive gain adjustment method to compensate for amplitude based on the amplitude attenuation characteristics of the raw GPR data, and enhancing the amplitude intensity of deep reflected waves to highlight the radar wave response characteristics of hidden road defects, can high-quality GPR deep learning samples be automatically and in batches generated. Figure 4 It is a raw data record of ground-penetrating radar without any gain, collected using a ground-coupled shielded antenna.

[0070] Figure 4 In the early stages of ground-penetrating radar (GPR) signals and reflected waves from shallow near-surface layers exhibit strong amplitudes, while reflected waves from deeper layers have very weak amplitudes. Particularly noteworthy is that the phase axis of the first half-cycle of the early GPR signal is flat, the waveform is stable, and the amplitude is uniform, making it an ideal reference for assessing the attenuation of reflected wave amplitude from deep targets. The mean amplitude A1 of the phase axis of the first half-cycle of the early GPR signal and its corresponding sampling point number m are extracted. The mean amplitude A1 is used to assess the attenuation of deep reflected wave amplitude and estimate the exponential gain factor. The sampling point number m is the starting point for applying the exponential gain function for amplitude compensation. The mean amplitude within a time window range of 0.5–0.625 is extracted and calculated as the mean amplitude A2 of the deep reflected wave. First, the ratio of A1 to A2 is calculated. If the ratio is less than a certain threshold, the radar data is considered to contain gain information and no gain compensation is needed. If the ratio is greater than a certain threshold, the radar data is considered to not contain gain information and amplitude compensation using the exponential gain function is required.

[0071] Based on the above implementation method, as a preferred implementation method, considering factors such as geometric diffusion and absorption attenuation, the radar wave amplitude is compensated, and the expression of the exponential gain function is:

[0072]

[0073] A i This represents the original amplitude value corresponding to the i-th sampling point in the raw ground-penetrating radar data. For A i The amplitude value after gain amplification, where v is the average propagation speed of the electromagnetic wave in the underground medium. t i Let be the arrival time of the reflected wave corresponding to the i-th sampling point.

[0074] Based on the above implementation method, as a preferred implementation method, the expression for the exponential gain factor α is:

[0075]

[0076] κ is a constant.

[0077] Specifically, with κ = 0.002, it can be seen that the gain function value is inversely proportional to A2. The larger the mean amplitude A2 of the deep reflected wave, the smaller the exponential gain factor α, and the slower the growth rate of the gain function curve; conversely, the smaller the mean amplitude A2 of the deep reflected wave, the larger the exponential gain factor α, and the faster the growth rate of the gain function curve.

[0078] Based on the above implementation method, as a preferred implementation method, the arrival time t of the reflected wave corresponding to the i-th sampling point is... i The expression is: t i =(im)·△t

[0079] △t is the time sampling interval, and m is the number of sampling points. When i≤m, the signal corresponding to sampling point i is earlier than or equal to the arrival time of the early signal of the ground penetrating radar. In this case, no amplitude compensation is required, so the gain function value is set to be equal to 1, and the radar wave amplitude value remains unchanged. When i>m, the signal corresponding to sampling point i is later than the arrival time of the early signal of the ground penetrating radar. In this case, the exponential gain function is used to perform amplitude compensation on the radar wave.

[0080] Based on the expression for the exponential gain function, the expression for the exponential gain factor α, and the arrival time t of the reflected wave corresponding to the i-th sampling point. i The expression for this yields its exponential gain curve as shown below. Figure 5 As shown, Figure 6 for Figure 4 The comparison chart shows the single-channel waveforms before and after amplitude compensation using the exponential gain function for the intermediate channel data. Figure 7 for Figure 4 Ground-penetrating radar profile after amplitude compensation using the exponential gain function.

[0081] Based on the above implementation method, as a preferred implementation method, the specific steps of S5 are as follows:

[0082] S5.1: Batch import N ground-penetrating radar files; theoretically, all radar files in the same folder can be imported at once.

[0083] S5.2: Sequentially query whether the current radar file n (n = 1, 2, 3, ... N) has a corresponding manual interpretation result file. If the radar file does not have a corresponding manual interpretation result, then sequentially query whether the next radar file (n = n + 1) has a corresponding manual interpretation result, and so on. If the radar file has a corresponding manual interpretation result, then proceed to step S6.3; when n > N, execute step S5.1;

[0084] S5.3: Read the total number M of hidden road defects from the manual interpretation results corresponding to the current radar file;

[0085] S5.4: Read the starting point, ending point coordinates, and disease type or number of the m-th disease (m = 1, 2, 3, ... M) in the manual interpretation results in sequence. Find the radar channel number and sampling point number of the starting point and the radar channel number and sampling point number of the ending point of the disease in the radar file. Expand the total number of channels of the sample data to the left and right with the disease as the center, and expand it to the corresponding integer multiple in the upward and downward directions to include all sampling points. Use the expanded radar data as the sample data for deep learning. That is, the sample data includes both the radar data corresponding to the disease and the background data around the disease. When m > M, execute step S5.2.

[0086] S5.5: Perform adaptive gain adjustment and normalization on the sample data, automatically generate radar images and tag pairs, and set m = m + 1, then execute step S5.4.

[0087] Based on the above process, a computer program can be developed to import hundreds of radar files at once, automatically label and generate tens of thousands of sample (image and label pairs) data, realize the generation of ground-penetrating radar deep learning samples in a short time, and transform manual interpretation results and expert experience into sample data for deep learning training.

[0088] Based on the above implementation, as a preferred implementation, the deep learning samples include one-to-one radar image and label pairs, specifically image and label pairs, wherein the filenames of the radar images and the labels are the same, and are respectively placed in the file paths / image / *.jpg and / label / *.txt.

[0089] Based on the above implementation method, as a preferred implementation method, when there are n defect targets in a radar image, the tag file is identified by n lines. Specifically, each line adopts the format (class xy ​​width height), where class represents the defect type identifier, consistent with the configuration file (*.yaml) in YOLO training.

[0090] This invention designs a simple and easy-to-understand data format to store manually interpreted results, establishes a mapping relationship between hidden road defects and their ground-penetrating radar (GPR) data, and enables autonomous retrieval, location, cutting, and automatic annotation of GPR data. Employing an automatic annotation method and adaptive gain adjustment for deep learning samples of hidden road defects using GPR, it can quickly generate deep learning samples, significantly saving sample annotation time and enabling the generation of large batches of GPR deep learning samples in a short time. In long-distance detection of hidden road defects, GPR mainly uses the common offset profiling method to scan and image the road structure. Without complex data processing, it can obtain radar profile maps corresponding to the distribution of underground geological structures. The intelligent identification and classification of GPR profiles based on deep learning has significant advantages.

[0091] All aspects not detailed herein are conventional technical means known to those skilled in the art. The above content shows and describes the basic principles, main features, and beneficial effects of the present invention. The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automatic annotation method for YOLO format samples from ground-penetrating radar deep learning, characterized in that: Includes the following steps: S1: Save or save the manual interpretation results of ground-penetrating radar for hidden road defects in a unified data format, with each ground-penetrating radar data corresponding to a manual interpretation result file; S2: Read the ground-penetrating radar data and its corresponding manual interpretation results. In the ground-penetrating radar data, sequentially search for the starting coordinates and ending coordinates of each defect in the manual interpretation results to delineate the ground-penetrating radar data space corresponding to the hidden road defects. S3: Centered on the ground-penetrating radar data space where the disease is located, retain the surrounding ground-penetrating radar data as the sample background, and extract the disease data and the surrounding sample background data as the disease sample data. S4: Adaptive gain adjustment; S5: Normalize the entire sample data space [0 1], and calculate the center point coordinates (x, y) of the disease and its horizontal and vertical dimension ratios (width, height), and automatically generate ground-penetrating radar deep learning samples that conform to the YOLO format.

2. The automatic annotation method for YOLO format samples of ground-penetrating radar deep learning according to claim 1, characterized in that: The manually interpreted result file in S1 has the same filename as the ground-penetrating radar data, but with a different file extension.

3. The automatic annotation method for YOLO format samples of ground-penetrating radar deep learning according to claim 1, characterized in that: The data format of the manually interpreted result file is as follows: Line 1: Total number of defects n contained in the ground penetrating radar data; Lines 2 to 6: These are the five parameters corresponding to the first disease. The first four parameters are the number of the starting point of the disease and the number of sampling points, i.e., the starting point coordinates; the number of the ending point of the disease and the number of sampling points, i.e., the ending point coordinates; and the fifth parameter is the disease type or the disease type number. Rows (n-1)×5+2 to 6: five parameters corresponding to the nth disease.

4. The automatic annotation method for YOLO format samples of ground-penetrating radar deep learning according to claim 1, characterized in that: The specific steps for S4 are as follows: S4.1: Extract the mean amplitude A1 of the phase axis of the first half-cycle of the early signal from the ground penetrating radar and the corresponding number of sampling points m; S4.2: Extract and calculate the mean amplitude within the CD time window as the mean amplitude A2 of the deep reflected wave; S4.3: Calculate the ratio of A1 to A2. When the ratio is less than the threshold, it is determined that the radar data already contains gain information and no gain compensation is needed. When the ratio is greater than the threshold, it is determined that the radar data does not contain gain information and amplitude compensation needs to be performed using an exponential gain function.

5. The automatic annotation method for YOLO format samples of ground-penetrating radar deep learning according to claim 6, characterized in that: The expression for the exponential gain function is: A i This represents the original amplitude value corresponding to the i-th sampling point in the raw ground-penetrating radar data. For A i The amplitude value after gain amplification, where v is the average propagation speed of electromagnetic waves in the underground medium, and t i Let be the arrival time of the reflected wave corresponding to the i-th sampling point.

6. The automatic annotation method for YOLO format samples of ground-penetrating radar deep learning according to claim 7, characterized in that: The expression for the exponential gain factor α is: κ is a constant.

7. The automatic annotation method for YOLO format samples of ground-penetrating radar deep learning according to claim 7, characterized in that: The arrival time t of the reflected wave corresponding to the i-th sampling point i The expression is: t i =(im)·△t △t represents the time sampling interval, and m represents the number of sampling points.

8. The automatic annotation method for YOLO format samples of ground-penetrating radar deep learning according to claim 1, characterized in that: The specific steps for S5 are as follows: S5.1: Batch import N ground-penetrating radar files; S5.2: Sequentially query whether the current radar file n (n = 1, 2, 3, ... N) has a corresponding manual interpretation result file. If the radar file does not have a corresponding manual interpretation result, then sequentially query whether the next radar file (n = n + 1) has a corresponding manual interpretation result, and so on. If the radar file has a corresponding manual interpretation result, then proceed to step S6.3; when n > N, execute step S5.1; S5.3: Read the total number M of hidden road defects from the manual interpretation results corresponding to the current radar file; S5.4: Read the starting point, ending point coordinates, and disease type or number of the m-th disease (m = 1, 2, 3, ... M) in the manual interpretation results in sequence. Find the radar channel number and sampling point number of the starting point and the radar channel number and sampling point number of the ending point of the disease in the radar file. Expand the total number of channels of the sample data to the left and right with the disease as the center, and expand it to the corresponding integer multiple in the upward and downward directions to include all sampling points. Use the expanded radar data as the sample data for deep learning. That is, the sample data includes both the radar data corresponding to the disease and the background data around the disease. When m > M, execute step S5.

2. S5.5: Perform adaptive gain adjustment and normalization on the sample data, automatically generate radar images and tag pairs, and set m = m + 1, then execute step S5.

4.

9. The automatic annotation method for YOLO format samples of ground-penetrating radar deep learning according to claim 8, characterized in that: The deep learning samples in S5 include one-to-one radar image and label pairs. The filenames of the radar images and the labels are the same, and are placed in the file paths / image / *.jpg and / label / *.txt, respectively.

10. The automatic annotation method for YOLO format samples of ground-penetrating radar deep learning according to claim 3, characterized in that: When there are n defective targets in a radar image, the tag file uses n lines to identify them.

Images