Ground penetrating radar machine learning target detection method

Abstract

The application discloses a ground penetrating radar machine learning target detection method, comprising the following steps: 1) collecting underground scene data by using a ground penetrating radar, and preprocessing the underground scene data to obtain first processing data; 2) constructing a corresponding background and target dielectric property data set; 3) applying a finite difference time domain method to simulate the propagation of electromagnetic waves in the underground medium, decoupling Maxwell equations, and processing reflected waves at the boundary by using a CPML absorbing boundary condition, discretizing the calculation region, and obtaining a forward hyperbolic curve feature; 4) labeling the generated forward hyperbolic curve feature by using Labelimg, and constructing training data; 5) training a DS-YOLOv9 network by using the training data, and obtaining a target detection model; and 6) inputting a related data set into the target detection model to obtain a target hyperbolic curve recognition result. The application can process various types of abnormal bodies in a complex scene, and can well identify sparse or dense rock bodies.

Description

TECHNICAL FIELD

[0001] The application relates to the field of electromagnetic simulation and image recognition, and particularly relates to a ground penetrating radar machine learning target detection method. BACKGROUND

[0002] The ground penetrating radar is a device for non-destructive detection of underground scenes by using electromagnetic wave pulses. Underground target classification and identification is a key task of ground penetrating radar interpretation, and the accuracy directly affects the application effect of the ground penetrating radar in complex underground environments.

[0003] The hyperbolic shape of the target is the most important basis for ground penetrating radar target detection, and therefore curve fitting is a commonly used method for ground penetrating radar target detection. However, there are a large number of background noises in complex environments, such as ionospheric noise, artificial electromagnetic interference, etc. These noises will mask the effective echo signals of the target, making it difficult to distinguish the target from the noise. The shapes and sizes of underground targets are various, such as metal pipes, cavities, rock fissures, etc., and their reflection characteristics change due to medium differences, increasing the difficulty of classification and detection. The traditional curve fitting method needs to rely on artificial design features, such as signal amplitude, hyperbolic characteristics, etc., which may be effective for specific targets, but are difficult to adapt to changes in data distribution in different scenes. SUMMARY

[0004] The purpose of the present application is to provide a ground penetrating radar machine learning target detection method, comprising the following steps:

[0005] 1) Collecting underground scene data by using the ground penetrating radar, and pre-processing the underground scene data to obtain first processing data;

[0006] 2) Processing the first processing data based on geological information and prior knowledge to obtain background abnormal body material information, and constructing a corresponding background and target dielectric property dataset;

[0007] 3) Simulating the propagation of electromagnetic waves in underground media by using the finite difference time domain method, decoupling the Maxwell equation, processing the reflected waves at the boundary by using the CPML absorbing boundary condition, and discretizing the calculation region to obtain the forward hyperbolic curve characteristics;

[0008] 4) Labeling the generated forward hyperbolic curve characteristics by using Labelimg, thereby constructing training data;

[0009] 5) Training the DS-YOLOv9 network by using the training data to obtain a target detection model;

[0010] 6) Inputting the corresponding background and target dielectric property dataset into the target detection model to obtain a target hyperbolic curve identification result.

[0011] Further, in step 1), the step of preprocessing the underground scene data includes direct wave removal processing, time zero point correction, band pass filtering, background noise removal and gain compensation.

[0012] Further, the method of direct wave removal processing includes the mean method.

[0013] Further, the Maxwell equation is as follows:

[0014]

[0015] where H y and H z are the magnetic field components in the y and z directions, E x is the electric field intensity in the x direction, ε is the dielectric constant, σ is the conductivity, and t is time.

[0016] The decoupled Maxwell equation is as follows:

[0017]

[0018] where, represents the value of the magnetic field H y at the grid point (i, j) at the (n+1 / 2)th time step; represents the value of the electric field E x at the (n)th time step.

[0019] Further, in step 4), the step of labeling the forward hyperbolic curve features using Labelimg includes:

[0020] For the target hyperbolic curve region, the position and size of the target box are described based on normalized coordinates, including the target class number class_id, the center point coordinates (x center , y center ) of the target box, the width w and the height h of the target box.

[0021] where the center point coordinates of the target box, the width and the height of the target box are as follows:

[0022]

[0023] where (x min , y min ) and (x max , y max ) are the target box coordinates, W is the image width, and H is the image height.

[0024] Further, the input features of the DS-YOLOv9 network are as follows:

[0025]

[0026] wherein C in is the number of channels. X is an input feature; R is a real number.

[0027] Further, the DS-YOLOv9 network comprises an input layer, a feature extraction layer, a feature fusion layer and an output layer.

[0028] The feature extraction layer fuses a dynamic snake convolution, and the calculation equations of the convolution kernel moving on the x-axis and y-axis are as follows:

[0029]

[0030] wherein K indicates a center point of the convolution kernel, (x i ,y i ) represents a center coordinate, (x i+c ,y i+c ) represents a center coordinate based on c unit offsets, and K i±c , K j±c are center points of the convolution kernel after moving on the x-axis and y-axis; represents a sum of cumulative offsets of the convolution kernel moving on the x-axis.

[0031] Further, in the training process of the DS-YOLOv9 network, the loss function adopted comprises a bounding box loss Loss box and a classification loss Loss cls .

[0032] wherein the bounding box loss Loss box is as follows:

[0033]

[0034] wherein IoU is an intersection over union of the identified bounding box and the real bounding box; FG represents an index set of all positive samples, is a sum of weights of all positive samples, ω i is a weight of the sample i, i.e., its confidence score, b i is a predicted bounding box, is a corresponding real bounding box; Area of Overlap and Area of Union are the identified bounding box and the real bounding box.

[0035] The classification loss Loss cls is as follows:

[0036]

[0037] wherein N is a total number of all predictions, is a real class probability, is a real class label.

[0038] Further, the learning rate used in the DS-YOLOv9 network training process is as follows:

[0039]

[0040] lr start is the initial value; lr end is the decay minimum value. current epoch and total epochs represent the current iteration number and the total iteration number, respectively.

[0041] The technical effects of the present application are self-evident. The target detection algorithm based on the improved network DS-YOLOv9 of YOLOv9 is successfully applied to the abnormal body recognition field of ground penetrating radar. The algorithm is used to train and predict the synthetic data of rock. The algorithm performs more excellently than Mask-RCNN and Swin-Transformer in the detection of radar echo hyperbola, and can effectively avoid missed detection and false detection. In addition, the high efficiency of the method also lies in its strong adaptability. DS-YOLOv9 exceeds other methods in the indicators of mAPIOU=0.5, mAPIoU=0.5:0.95 and recall rate, and can handle multiple types of abnormal bodies in complex scenes, whether sparse or dense rock bodies can be well identified. This has important practical significance for application scenarios that require high-precision positioning and identification of underground targets. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is a flowchart of a ground penetrating radar machine learning target detection method provided by an embodiment of the present application.

[0043] Figure 2 is a schematic diagram of labelming labeling forward hyperbola provided by an embodiment of the present application.

[0044] Figure 3 is a whole structure framework diagram of DS-YOLOv9 provided by an embodiment of the present application.

[0045] Figure 4 is a DSELAN4 module structure diagram provided by an embodiment of the present application.

[0046] Figure 5 is an abnormal body model and its finite difference forward diagram provided by an embodiment of the present application. The dark blue part is the air layer, the light blue part is the soil layer, and the yellow part is the rock block. The red dot is the radar motion trace.

[0047] Figure 6are the identification results of different networks on a single rock anomaly body: (a) a true model setting, one rock anomaly body is placed in the soil; (b) Mask-RCNN identification result; (c) Swin-Transformer identification result; (d) DS-YOLOv9 identification result.

[0048] Figure 7 are the identification results of different networks on sparse rock anomaly bodies: (a) (e) a true model setting, three rock anomaly bodies are placed in the soil; (b) (f)

[0049] Mask-RCNN identification result; (c) (g) Swin-Transformer identification result; (d) (h) DS-YOLOv9 identification result.

[0050] Figure 8 are the identification results of different networks on dense rock anomaly bodies: (a) a true model setting, ten rock anomaly bodies are placed in the soil; (b) Mask-RCNN identification result; (c) Swin-Transformer identification result; (d) DS-YOLOv9 identification result. DETAILED DESCRIPTION

[0051] The application will be further described below in conjunction with examples, but should not be understood as limiting the above-mentioned subject matter of the application to the following examples. Various substitutions and modifications can be made according to ordinary technical knowledge and conventional means in the art without departing from the above-mentioned technical idea of the application, and all should be included in the protection scope of the application.

[0052] Example 1

[0053] Reference Figures 1 to 8 A ground penetrating radar machine learning target detection method, comprising the following steps:

[0054] 1) Collecting underground scene data by using a ground penetrating radar, and pre-processing the underground scene data to obtain first processing data;

[0055] 2) Based on geological information and prior knowledge, processing the first processing data to obtain background anomaly body material information, and constructing a corresponding background and target dielectric property data set;

[0056] 3) Simulating the propagation of electromagnetic waves in underground media by using the finite difference time domain method, decoupling Maxwell's equations, and using CPML absorbing boundary conditions to process reflected waves at the boundary, discretizing the calculation region, and obtaining forward hyperbolic curve characteristics;

[0057] 4) Labeling the generated forward hyperbolic curve characteristics by using Labelimg, thereby constructing training data;

[0058] 5) training the DS-YOLOv9 network with the training data to obtain a target detection model;

[0059] 6) inputting the corresponding background and target dielectric property dataset into the target detection model to obtain a target hyperbolic curve recognition result.

[0060] In step 1), the step of pre-processing the underground scene data includes direct wave removal, time zero point correction, band pass filtering, background noise removal and gain compensation.

[0061] The method of direct wave removal includes mean method.

[0062] The Maxwell equation is as follows:

[0063]

[0064] where H y and H z are the magnetic field components in y and z directions, E x is the electric field intensity in x direction, ε is the dielectric constant, σ is the conductivity, and t is the time;

[0065] The decoupled Maxwell equation is as follows:

[0066]

[0067] where, represents the value of the magnetic field H y at the grid point (i,j) at the (n+1 / 2)th time step; represents the value of the electric field E x at the (n)th time step.

[0068] In step 4), the step of labeling the forward hyperbolic curve features using Labelimg includes:

[0069] For the target hyperbolic curve area, the position and size of the target box are described based on normalized coordinates, including target class number class_id, target box center point coordinates (x center , y center ), target box width w and height h;

[0070] where the target box center point coordinates, target box width and height are as follows:

[0071]

[0072] where (x min , y min ) and (x max , y max) is the target box coordinate, W is the image width, and H is the image height.

[0073] The input features of the DS-YOLOv9 network are as follows:

[0074]

[0075] wherein C in is the number of channels, X is the input feature, and R is a real number.

[0076] The DS-YOLOv9 network comprises an input layer, a feature extraction layer, a feature fusion layer, and an output layer.

[0077] The feature extraction layer fuses a dynamic snake convolution, and the calculation equations of the convolution kernel moving on the x-axis and y-axis are as follows:

[0078]

[0079] wherein K indicates the center point of the convolution kernel, (x i ,y i ) represents the center coordinates, (x i+c ,y i+c ) represents the center coordinates based on c unit offsets, and ΣK i±c , K j±c are the center points of the convolution kernel after moving on the x-axis and y-axis; ΣK box represents the sum of the cumulative offsets of the convolution kernel moving on the x-axis.

[0080] In the training process of the DS-YOLOv9 network, the loss function adopted includes a bounding box loss Loss cls and a classification loss Loss box .

[0081] wherein the bounding box loss Loss i is as follows:

[0082]

[0083] wherein IoU is the intersection over union of the identified bounding box and the real bounding box, FG represents the index set of all positive samples, is the sum of the weights of all positive samples, ω i is the weight of sample i, i.e., its confidence score, b i is the predicted bounding box, is the corresponding real bounding box; Area of Overlap and Area of Union are the identified bounding box and the real bounding box.

[0084] Classification loss Loss cls As shown below:

[0085]

[0086] In the formula, N is the total number of all predictions, is the true class probability, is the true class label.

[0087] During the training process of the DS-YOLOv9 network, the learning rate used is as follows:

[0088]

[0089] In the formula, lr start is the initial value; lr end is the decay minimum value. current epoch and total epochs represent the current iteration number and the total iteration number, respectively.

[0090] Example 2:

[0091] A ground penetrating radar machine learning target detection method, comprising the following steps:

[0092] 1) Collecting underground scene data using a ground penetrating radar, and preprocessing the underground scene data to obtain first processed data;

[0093] 2) Based on geological information and prior knowledge, processing the first processed data to obtain background abnormal body material information, and constructing a corresponding background and target dielectric property dataset;

[0094] 3) Applying the finite difference time domain method to simulate the propagation of electromagnetic waves in underground media, decoupling Maxwell's equations, and using CPML absorbing boundary conditions to process reflected waves at the boundary, discretizing the calculation region, and obtaining forward hyperbolic curve characteristics;

[0095] 4) Labeling the generated forward hyperbolic curve characteristics using Labelimg, thereby constructing training data;

[0096] 5) Training the DS-YOLOv9 network using the training data to obtain a target detection model;

[0097] 6) Inputting the corresponding background and target dielectric property dataset into the target detection model to obtain target hyperbolic curve recognition results.

[0098] Example 3:

[0099] The application discloses a ground penetrating radar machine learning target detection method, and the technical content is the same as that of the embodiment 2, and further, the step 1) comprises the steps of direct wave removal processing, time zero point correction, band pass filtering, background noise removal and gain compensation.

[0100] Embodiment 4:

[0101] The application discloses a ground penetrating radar machine learning target detection method, and the technical content is the same as that of any one of the embodiments 2-3, and further, the method of removing direct waves comprises a mean method.

[0102] Embodiment 5:

[0103] The application discloses a ground penetrating radar machine learning target detection method, and the technical content is the same as that of any one of the embodiments 2-4, and further, the Maxwell equation is as follows:

[0104]

[0105] Wherein, H y and H z are magnetic field components in y and z directions, E x is an electric field intensity in x direction, epsilon is a dielectric constant, sigma is an electric conductivity, and t is time.

[0106] The decoupled Maxwell equation is as follows:

[0107]

[0108] Wherein, represents a value of a magnetic field H y of a grid point (i, j) at an (n+1 / 2)th time step; represents a value of an electric field E x at an nth time step.

[0109] Embodiment 6:

[0110] The application discloses a ground penetrating radar machine learning target detection method, and the technical content is the same as that of any one of the embodiments 2-5, and further, in the step 4), the step of labeling the forward hyperbolic curve characteristics by using Labelimg comprises:

[0111] For a target hyperbolic curve region, the position and size of a target frame are described based on a normalized coordinate, that is:

[0112] class_id x center y center w h(3)

[0113] Wherein, class_id is a target category number, x center , y centeris the center point coordinate of the target frame, w and h are the width and height of the target frame;

[0114] The center coordinates, width, and height of the target frame are as follows:

[0115]

[0116] Where (x min ,y min ) and (x max ,y max ) are the target frame coordinates, W is the image width, and H is the image height.

[0117] Example 7:

[0118] A ground penetrating radar machine learning target detection method, the technical content of which is the same as any one of Examples 2-6, further, the input features of the DS-YOLOv9 network are as follows:

[0119]

[0120] Where C in is the number of channels.

[0121] Example 8:

[0122] A ground penetrating radar machine learning target detection method, the technical content of which is the same as any one of Examples 2-7, further, the DS-YOLOv9 network includes an input layer, a feature extraction layer, a feature fusion layer, and an output layer;

[0123] The feature extraction layer integrates dynamic snake convolution, and the calculation equations for the convolution kernel moving on the x-axis and y-axis are as follows:

[0124]

[0125] Where K refers to the center point of the convolution kernel, (x i ,y i ) represents the center coordinate, (x i+c ,y i+c ) represents the center coordinate based on c unit offset, Represents the sum of the cumulative offsets of the convolution kernel moving on the y-axis.

[0126] Example 9:

[0127] A ground penetrating radar machine learning target detection method, the technical content is the same as any one of Examples 2-8, further, during the DS-YOLOv9 network training process, the loss function used includes the bounding box loss Loss box and classification loss Loss cls ;

[0128] wherein the bounding box loss Loss box As follows:

[0129]

[0130] wherein IoU is the intersection over union of the identified bounding box and the true bounding box; FG represents an index set of all positive samples, is the sum of all positive sample weights, ω i is the weight of sample i, i.e., its confidence score, b i is the predicted bounding box, is the corresponding true bounding box;

[0131] Classification loss Loss cls As follows:

[0132]

[0133] wherein N is the total number of all predictions, is the true class probability, is the true class label.

[0134] Embodiment 10:

[0135] A ground penetrating radar machine learning target detection method, the technical content is the same as any one of embodiments 2-9, further, in the DS-YOLOv9 network training process, the learning rate used is as follows:

[0136]

[0137] wherein lr start is the initial value; lr end is the decay minimum value.

[0138] Embodiment 10:

[0139] A ground penetrating radar machine learning target detection method, the method is as Figure 1 shown, comprising the following steps:

[0140] S1: Collecting data by a ground penetrating radar, performing a direct wave removal processing process and a data preprocessing process on the data collected by the ground penetrating radar to obtain first processed data.

[0141] In the embodiment of the application, the mean method is used for the direct wave removal processing process, the average value of each channel data is calculated and subtracted from the original data, and the main content of the data preprocessing includes time zero point correction, band pass filtering, background noise removal and gain compensation.

[0142] S2: Constructing a related data set for the first processed data based on geological information and prior knowledge.

[0143] In the embodiment of the present application, in the data set established for the input of DS-YOLOv9, it is necessary to pay attention to the fact that the synthetic data set and the real data are too different, and direct training and prediction will lead to false detection and missed detection, so it is necessary to establish a similar physical model of underground medium and target, and use numerical methods to simulate the propagation process of electromagnetic waves; in order to make the synthetic data closer to the real data, the statistical characteristics of the real data are combined to further process the simulated data, and a high-quality data set containing rich target and noise characteristics is constructed.

[0144] S3: The finite difference time domain method is applied to simulate the propagation of electromagnetic waves in the underground medium, the Maxwell equation is decoupled, the CPML absorbing boundary condition is used to process the reflected wave at the boundary, the Yee grid is selected, and the calculation area is efficiently discretized.

[0145] In the embodiment of the present application, the formula for discretizing the Maxwell equation set based on the finite difference method includes:

[0146]

[0147] where H y and H z are the magnetic field components in the y and z directions, E x is the electric field intensity in the x direction, ε is the dielectric constant, σ is the conductivity, and t is the time.

[0148] By spatial difference and time difference, the partial derivative is converted into a difference form:

[0149]

[0150] where represents the value of the magnetic field H y at the grid point (i, j) at the n+1 / 2 time step. represents the value of the electric field E x at the n time step.

[0151] S4: Labelimg is used to accurately label the generated forward hyperbolic curve characteristics to provide training data for the DS-YOLOv9 intelligent recognition network.

[0152] In the embodiment of the present application, as shown in Figure 2 , Labelimg is used to label the forward hyperbolic curve characteristics, including:

[0153] For the target hyperbolic curve region, the position and size of the target frame are described based on normalized coordinates:

[0154] class_id x center ycenter w h (3)

[0155] where class_id is the target class number, x center , y center are the center point coordinates of the target frame, w, h are the width and height of the target frame, normalized to the image width and height range:

[0156]

[0157] where (x min , y min ) and (x max , y max ) are the target frame coordinates, W is the image width, and H is the image height.

[0158] S5: input the generated forward data and the corresponding label file into the DS-YOLOv9 intelligent target detection network. Design the core loss function, select the appropriate learning rate setting, and ensure that the model avoids overfitting or underfitting problems. Use the fully trained model weight to predict the actual data and obtain the final recognition result.

[0159] In the embodiment of the present application, as shown in Figure 3 and Figure 4 , a DS-YOLOv9 intelligent recognition network is built, which includes:

[0160] DSConv convolution block, the channel number of the input feature tensor is C in :

[0161]

[0162] The calculation equation of the convolution kernel of the dynamic snake convolution moving in the x-axis:

[0163]

[0164] where K refers to the center point of the convolution kernel, (x i , y i ) represents the center coordinates, (x i+c , y i+c ) represents the center coordinates based on c unit offsets, and represents the sum of the cumulative offsets of the convolution kernel moving in the y-axis.

[0165] In the embodiment of the present application, the core loss function is designed, the appropriate learning rate setting is selected, and the model is ensured to avoid overfitting or underfitting problems, including:

[0166] The intersection over union of the recognition bounding box and the real bounding box can be represented as:

[0167]

[0168] Boundary box loss of DS-YOLOv9 box which can be expressed as:

[0169]

[0170] where FG represents the index set of all positive samples, ω is the sum of weights of all positive samples, i bi is the weight of sample i, i.e., its confidence score, i bi is the predicted boundary box, bi is the corresponding real boundary box.

[0171] Classification loss Loss cls which can be expressed as formula (10), where N is the total number of all predictions, pi is the real class probability, yi is the real class label:

[0172]

[0173] A linear decay learning rate schedule is adopted, and at each epoch, the learning rate gradually decays from the initial value lr start to lr end :

[0174]

[0175] As shown in Figure 5 , (a) is an anomaly body model. The dark blue part is the air layer, the light blue part is the soil layer, and the yellow part is the rock block. The red dot is the radar motion trace. Figure 5 (b) is the Bscan synthetic data. Figure 5 (a) is an anomaly body model, the horizontal distance is 8m, and the vertical distance is 3.3m. The model is composed of three parts: an air layer with a relative permittivity of 1 and a conductivity of 0mS / m; a soil layer with a relative permittivity of 2.1 to 3 along the depth, and a conductivity of 1mS / m, with the ground surface as the horizontal interface; a rock block with a relative permittivity of 5 and a conductivity of 10mS / m. The model area is divided into 800×330 square grid units with a longitudinal and lateral grid spacing of 0.01m×0.01m. To eliminate the interference of artificial boundaries, the CPML absorption layer is set to 20, and the thickness is 0.2m. Figure 5 The simulation adopts a center frequency of 800MHz as the excitation source, records for 40ns, and the sampling interval is 0.0167ns.

[0176] Figure 5 ​(a) The running acquisition track of GPR. The transmitting antenna is set in the air layer 0.3 m away from the ground, and 1 transmitting antenna is placed every 0.01 m in the horizontal direction, a total of 800 transmitting antennas. The forward result is shown in Figure 5 (b).

[0177] Figure 6 are the identification results of different networks on a single rock anomaly body: (a) real model setting, 10 rock anomaly bodies are placed in the soil; (b) Mask-RCNN identification result; (c) Swin-Transformer identification result; (d) DS-YOLOv9 identification result. This study designed three different anomaly body target detection methods: (1) Mask-RCNN relying on high-precision data enhancement and segmentation mask annotation; (2) Swin-Transformer capable of global modeling, optimizing multi-scale features and global annotation; (3) DS-YOLOv9 improved to have efficient background and target separation capability. The training set uses 300 forward B-SCAN images, which are divided into training set: test set: validation set according to the ratio of 8:1:1.

[0178] The target detection results of the above three different networks are shown in Figure 6 . The real anomaly body model setting is shown in Figure 6 (a). The Mask-RCNN identification result is shown in Figure 6 (b). Although it can identify certain target areas on the boundary of the rock anomaly body, at the vertex of the hyperbolic curve, the recognition box does not completely fit the outline of the real anomaly body, showing some loss of details and segmentation accuracy problems. The identification effect of Swin-Transformer is shown in Figure 6 (c). The identification effect of DS-YOLOv9 is shown in Figure 6 (d). The detection box fits the boundary of the anomaly body very well, proving that the network has a significant advantage in the ability to separate background and target.

[0179] Figure 7 are the identification results of different networks on sparse rock anomaly bodies: (a) (e) real model setting, three rock anomaly bodies are placed in the soil; (b) (f) Mask-RCNN identification result; (c) (g) Swin-Transformer identification result; (d) (h) DS-YOLOv9 identification result.

[0180] The target detection results of the above three different networks are shown in Figure 7 . The real anomaly body model setting is shown in Figure 7 (a). The Mask-RCNN identification result is shown in Figure 7(b) shown, the upper and lower left rock blocks are well identified, but the anomaly body in the lower right is missed; the recognition effect of Swin-Transformer is as shown in Figure 7 (c) shown, the recognition effect of DS-YOLOv9 is as shown in Figure 7 (d) shown, it can be seen that the three rock blocks are all circled with high confidence. The real anomaly body model setting is shown in Figure 7 (e). The Mask-RCNN recognition result is as shown in Figure 7 (f) shown, the DS-YOLOv9 recognition result is as shown in Figure 7 (h) shown, the three rock blocks are not missed; the recognition effect of Swin-Transformer is as shown in Figure 7 (g) shown, the result shows that four rock echo hyperbolas are circled, and a false detection phenomenon occurs. Compared with the first two target detection networks, the DS-YOLOv9 built by the present application has a dynamic snake-shaped convolution block sensitive to irregular anomaly bodies such as hyperbolas, which can better avoid false detection and missed detection.

[0181] Figure 8 are the recognition results of different networks on dense rock anomaly bodies: (a) real model setting, 10 rock anomaly bodies are placed in the soil; (b) Mask-RCNN recognition result; (c) Swin-Transformer recognition result; (d) DS-YOLOv9 recognition result. The target detection results of the above three different networks are as shown in Figure 8 respectively. The real anomaly body model setting is shown in Figure 8 (a). The Mask-RCNN recognition result is as shown in Figure 8 (b) shown, the recognition effect of Swin-Transformer is as shown in Figure 8 (c) shown, both networks have missed detection and false detection; the recognition effect of DS-YOLOv9 is as shown in Figure 8 (d) shown, although one rock is missed, no false detection result occurs.

[0182] To further quantify the detection effect of different networks on the hyperbolic curve of ground penetrating radar echo, the mAPIOU=0.5, mAPIoU=0.5:0.95 and Recall performance indicators are compared, and the results are shown in Table 1. From the table, it can be seen that DS-YOLOv9 performs well in all indicators, among which mAPIOU=0.5 reaches 0.861, which is significantly better than 0.772 of Swin-Transformer and 0.750 of Mask-RCNN. At the same time, DS-YOLOv9 also achieves high scores of 0.477 and 0.731 in mAPIoU=0.5:0.95 and recall rate respectively, fully embodying its advantages in detection accuracy and target recognition ability. In contrast, Swin-Transformer is in the middle in mAPIoU=0.5:0.95 and recall rate, which are 0.392 and 0.563 respectively, while the comprehensive performance of Mask-RCNN is relatively weak. The above results show that the improved model DS-YOLOv9 based on YOLO architecture has significant performance advantages in target detection tasks and is more adaptable in complex scenes.

[0183] Table 1

[0184]

[0185]

Claims

1. A ground penetrating radar machine learning target detection method, characterized in that: The following steps are involved: 1) using a ground penetrating radar to collect underground scene data, and preprocessing the underground scene data to obtain first processed data; 2) Based on geological information and prior knowledge, the first processed data is processed to obtain background anomaly material information and construct corresponding background and target dielectric property data sets; 3) The finite-difference time-domain method is used to simulate the propagation of electromagnetic waves in underground media, decouple the Maxwell equations, and use the CPML absorbing boundary conditions to process the reflected waves at the boundaries. The calculation area is discretized to obtain the forward hyperbolic characteristics; 4) Use Labelimg to label the generated forward hyperbola features to construct training data; 5) Use the training data to train the DS-YOLOv9 network to obtain the target detection model; 6) Inputting the corresponding background and target dielectric property data sets into the target detection model to obtain the target hyperbola recognition result; The DS-YOLOv9 network includes an input layer, a feature extraction layer, a feature fusion layer and an output layer; The feature extraction layer integrates dynamic snake convolution, and the calculation equations for the convolution kernel moving on the x-axis and y-axis are as follows: Among them, K refers to the center point of the convolution kernel, (x i ,y i ) represents the center coordinate, (x i+c ,y i+c ) represents the center coordinate based on c unit offset, Represents the sum of the cumulative offsets of the convolution kernel moving on the y-axis; K i±c , K j±c is the center point of the convolution kernel after moving on the x-axis and y-axis; Represents the sum of the cumulative offsets of the convolution kernel moving on the x-axis; During the DS-YOLOv9 network training process, the loss functions used include bounding box loss Loss box and classification loss Loss cls ; Among them, the bounding box loss Loss box As shown below: Where IoU is the intersection-over-union ratio between the identified bounding box and the true bounding box; FG represents the index set of all positive samples, is the sum of all positive sample weights, ω i is the weight of sample i, that is, its confidence score, b i is the predicted bounding box, is the corresponding true bounding box; Area of ​​Overlap and Area of ​​Union are the recognized bounding box and the true bounding box; Classification loss cls As shown below: Where N is the total number of all predictions, is the true class probability, is the true category label; During DS-YOLOv9 network training, the learning rate lr used is as follows: Where, lr start is the initial value; lr end is the minimum decay value; current epoch and total epochs represent the current number of iterations and the total number of iterations respectively.

2. The method for ground penetrating radar machine learning target detection according to claim 1, characterized in that: In step 1), the steps of pre-processing the underground scene data include direct wave removal, time zero point correction, bandpass filtering, background noise removal and gain compensation.

3. The method for ground penetrating radar machine learning target detection according to claim 2, characterized in that: Methods for removing direct waves include the averaging method.

4. The method for ground penetrating radar machine learning target detection according to claim 1, characterized in that: Maxwell's equations are shown below: Among them, H y and H z are the magnetic field components in the y and z directions, E x is the electric field intensity in the x direction, ε is the dielectric constant, σ is the conductivity, and t is the time; The decoupled Maxwell equations are shown below: in, Indicates the magnetic field H at the grid point (i, j) at the n+1 / 2th time step y The value of Indicates that at the nth time step, the electric field E x The value of .

5. The method for ground penetrating radar machine learning target detection according to claim 1, characterized in that: In step 4), the steps of labeling the forward hyperbola features using Labelimg include: For the target hyperbolic area, the position and size of the target frame are described based on the normalized coordinate method, including the target category number class_id, the center point coordinates of the target frame (x center ,y center ), the width w and height h of the target box; The center coordinates, width, and height of the target frame are as follows: Where (x min ,y min ) and (x max ,y max ) are the target frame coordinates, W is the image width, and H is the image height.

6. The method for ground penetrating radar machine learning target detection according to claim 1, characterized in that: The input features of the DS-YOLOv9 network are as follows: Where C in is the number of channels; X is the input feature; R is a real number.

Images