Array ground penetrating radar target automatic reconstruction method based on image segmentation

Abstract

The invention discloses an array ground penetrating radar target automatic reconstruction method based on image segmentation. The method comprises the following steps of 1, performing data preprocessing on data acquired by a multi-frequency array ground penetrating radar and accumulated data; step 2, extracting a profile map of data of each multi-frequency array ground penetrating radar; step 3, obtaining a final data set and a tag; step 4, carrying out image segmentation training on the marked data set through Deeplabv3 +; step 5, obtaining a binarized mask image; and step 6, based on a three-dimensional array formed by splicing mayavi visual mask images, obtaining reconstruction of the cavity. According to the method disclosed by the invention, an actual measurement sample database is utilized, a construction method of an intelligent identification deep learning network model is researched, an existing model is optimized and improved, and a large amount of actual measurement data is subjected to data preprocessing and data enhancement, so that the number and diversity of samples capable of training ground penetrating radar data are increased.

Description

Technical Field

[0001] The present invention belongs to the field of ground penetrating radar, and particularly relates to an automatic reconstruction method for array ground penetrating radar targets based on image segmentation in this field. Background Art

[0002] Safe and unobstructed roads and underground projects are the foundation of urban development. Urban roads and various underground municipal facilities are carriers of each other, interacting with each other to jointly ensure the safe operation of the city. However, with the rapid development of urban construction, urban road collapse accidents occur frequently, seriously affecting the safety of urban operation and people's normal life. Three-dimensional reconstruction of underground cavities can accurately restore the three-dimensional structure of the cavities, thereby providing more accurate spatial positioning and reducing the cost and time required by traditional exploration methods. However, the reflection and diffraction of radar signals are interfered by various factors, and the interpretation process is complex and vulnerable to influence. Radar images may contain noise or artifacts, especially during the data collection process due to interference from equipment, environment, or other underground structures, resulting in poor accuracy of three-dimensional reconstruction of underground cavities. Summary of the Invention

[0003] In order to solve the technical problem of poor ability of traditional three-dimensional reconstruction of underground cavities, the present invention provides an automatic reconstruction method for array ground penetrating radar targets based on image segmentation. By using deep learning semantic segmentation on the longitudinal profile diagrams collected by the array ground penetrating radar, it can adapt to the changes in the target shape and improve the accuracy of target detection and reconstruction.

[0004] The present invention adopts the following technical solutions:

[0005] An automatic reconstruction method for array ground penetrating radar targets based on image segmentation, which is improved in that it includes the following steps:

[0006] Step 1, perform data preprocessing on the data collected by the multi-frequency array ground penetrating radar and the accumulated data;

[0007] Step 2, extract the profile diagrams of each multi-frequency array ground penetrating radar data;

[0008] Step 3, label the targets in each profile diagram, and at the same time enhance all sample data to obtain the final data set and labels;

[0009] Step 4, perform image segmentation training on the labeled data set through deeplabv3+ to obtain model weights;

[0010] Step 5, use the trained model weights to detect the array three-dimensional data to be reconstructed to obtain a binary mask image;

[0011] Step 6, based on mayavi, visualize the three-dimensional array composed of spliced mask images to obtain the reconstruction of the cavity.

[0012] Further, in step 1, the data collected by the multi-frequency array ground penetrating radar is 10-channel data; the accumulated data includes the data of urban arterial roads, branch roads and sidewalks detected in the past three years.

[0013] Further, in step 1, the means of data preprocessing is automatic gain, and the automatic gain is achieved by multiplying the radar data by an automatic gain weight function that changes with time:

[0014] y~(t)=P(t)y(t)

[0015] In the above formula, represents the radar data after automatic gain, P(t) represents the automatic gain weight function, and y(t) represents the radar data before automatic gain;

[0016] The number of time windows is determined by the number of parameter control points. When calculating the average amplitude, there should be an overlap of half a time window between every two adjacent time windows. The following parameters are calculated according to the number of control points: the length of the time window, the start time of the time window, and the end time of the time window;

[0017] The average amplitude A within the s-th time window s The calculation formula is:

[0018]

[0019] In the above formula, q(t) represents the single-channel data collected by the radar to be processed; T s-1 represents the start time of the s-th time window; T s+1 represents the end time of the s-th time window; B represents the number of sample points in the s-th time window;

[0020] Finally, the average amplitude of each time window is stored as the gain parameter of the control point corresponding to the center of its respective time window;

[0021] The weighting function is:

[0022]

[0023] In the above formula, p s represents the weighting factor corresponding to the center of the s-th time window, and F represents the balance coefficient used to adjust the magnitude of the effective amplitude after processing; for the weighting factors of points other than the center of the time window, they are obtained by linearly interpolating the weighting factors of the centers of two adjacent time windows.

[0024] Further, in step 2, the images with void targets are intercepted from each three-dimensional data profile; the number of intercepted images is 6000, the data size is 10*512, and the size of the image generated by linear interpolation is 130*512.

[0025] Further, step 3 specifically includes the following steps:

[0026] Step 31: Use the open-source image annotation tool labelme to annotate the underground cavity targets in the radar images;

[0027] Step 32: Combine the annotated radar images and annotation information to establish a radar image database of underground cavity targets;

[0028] Step 33: Adopt data augmentation methods to increase the data volume, and the data augmentation methods include horizontal flipping of images and scaling of images;

[0029] Step 34: Randomly allocate the data according to the ratio of 7:2:1 to obtain the final training set, validation set, and test set.

[0030] Further, during the training process of the deeplabv3+ model in step 4, the number of epochs is 300 times, and the Batch_Size is 16.

[0031] Further, in step 6, after splicing the masks into a three-dimensional array, visualize this three-dimensional array through visualization software.

[0032] The beneficial effects of the present invention are as follows:

[0033] The method disclosed by the present invention uses the measured sample database to study the construction method of the intelligent recognition deep learning network model, optimizes and improves the existing model, and performs data preprocessing and data augmentation on a large amount of measured data to increase the sample quantity and diversity of the trainable ground-penetrating radar data. The method disclosed by the present invention has good adaptability and generalization ability for extracting target features with various shapes, strong target recognition ability, and accurate reconstruction. Description of the Drawings

[0034] Figure 1 is the flow schematic diagram of the method disclosed by the present invention;

[0035] Figure 2 is the network structure schematic diagram of deeplabv3+;

[0036] Figure 3 is the image segmentation result diagram;

[0037] Figure 4 is the binary mask image;

[0038] Figure 5 is the result diagram I of the cavity three-dimensional reconstruction;

[0039] Figure 6 is the result diagram II of the cavity three-dimensional reconstruction. Detailed Embodiments

[0040] To make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0041] Embodiment 1 discloses an automatic reconstruction method for array ground penetrating radar targets based on image segmentation. Measured data collected by a multi-frequency array ground penetrating radar is used to make a dataset, and a model is trained to complete the segmentation of void targets. The deep learning network based on Deeplabv3+ enhances the reconstruction accuracy of voids with different shapes. As Figure 1 shown, it includes the following steps:

[0042] Step 1: Perform data preprocessing on the data collected by the multi-frequency array ground penetrating radar and the accumulated data to obtain data with obvious target features;

[0043] The multi-frequency array ground penetrating radar used is a three-dimensional ground penetrating radar detection vehicle + 400 MHz array, and the data collected is 10-channel data; the accumulated data includes data on urban arterial roads, branch roads, and sidewalks detected in the past three years.

[0044] The means for performing data preprocessing on the original data is automatic gain, which is achieved by multiplying the radar data by an automatic gain weight function that changes with time:

[0045]

[0046] In the above formula, represents the radar data after automatic gain, P(t) represents the automatic gain weight function, and y(t) represents the radar data before automatic gain;

[0047] The number of time windows is determined by the parameter control points. When calculating the average amplitude, there should be an overlap of half a time window between every two adjacent time windows. The following parameters are calculated according to the control points: the time window length, the start time of the time window, and the end time of the time window;

[0048] The average amplitude A in the s-th time window s The calculation formula is:

[0049]

[0050] In the above formula, q(t) represents the single-channel data collected by the radar to be processed; T s-1 represents the start time of the s-th time window; T s+1 represents the end time of the s-th time window; B represents the number of samples in the s-th time window;

[0051] Finally, store the average amplitude of each time window corresponding to the center of its respective time window as the gain parameter of the control point.

[0052] The weighting function is:

[0053]

[0054] In the above formula, p s represents the weighting factor corresponding to the center of the s-th time window, and F represents the balance coefficient used to adjust the size of the effective amplitude after processing; for the weighting factors of points other than the time window center, they are obtained by linearly interpolating the weighting factors of two adjacent time window centers.

[0055] Step 2: Extract the cross-sectional view of each multi-frequency array ground penetrating radar data.

[0056] Intercept the images with void targets in each three-dimensional data cross-sectional view; the number of intercepted images is 6000, the data size is 10 * 512, and the image size generated by linear interpolation is 130 * 512.

[0057] Step 3: Label the targets in each cross-sectional view, and at the same time enhance all sample data to obtain the final dataset and labels.

[0058] Step 31: Use the open-source image annotation tool labelme to annotate the underground void targets in the radar images.

[0059] Step 32: Combine the annotated radar images and annotation information to establish a radar image database of underground void targets.

[0060] Step 33: Adopt data augmentation methods to expand the data volume and increase sample diversity. The data augmentation methods include horizontal flipping of images and scaling of images.

[0061] Step 34: Randomly allocate the data in the ratio of 7:2:1 to obtain the final training set, validation set, and test set.

[0062] Step 4: Perform image segmentation training on the annotated dataset through deeplabv3+, Figure 3 is the image segmentation result diagram, obtain the model weights. During the training process of the deeplabv3+ model, the number of epochs is 300 times, and the Batch_Size is 16; the network structure of deeplabv3+ is as Figure 2 shown;

[0063] Step 5: Use the model weights obtained from training to detect the array three-dimensional data to be reconstructed, and obtain a binary mask image as Figure 4 shown;

[0064] Step 6, after splicing the masks into a three-dimensional array, visualize this three-dimensional array through visualization software, that is, visualize the three-dimensional array formed by splicing the mask images based on mayavi to obtain the reconstruction of the cavity. The result of the three-dimensional reconstruction of the cavity is shown in Figure 5 , as shown in 6.

Claims

1. A method for automatic reconstruction of array ground penetrating radar targets based on image segmentation, characterized in that: The steps include: Step 1, preprocessing the data collected and accumulated by the multi-frequency array ground penetrating radar; Step 2, extracting a profile of each multi-frequency array ground penetrating radar data; Step 3: Label the targets in each profile image and enhance all sample data to obtain the final data set and labels; Step 4: Use deeplabv3+ to perform image segmentation training on the labeled dataset to obtain the model weights; Step 5, using the trained model weights to detect the array three-dimensional data to be reconstructed, to obtain a binary mask image; Step 6: Based on Mayavi visualization, the three-dimensional array formed by splicing the mask images is used to obtain the reconstruction of the holes.

2. The automatic reconstruction method of array ground penetrating radar target based on image segmentation according to claim 1 is characterized by: In step 1, the data collected by the multi-frequency array ground penetrating radar is 10 channels of data; the accumulated data includes the data of urban trunk roads, branch roads and sidewalks detected in the past three years.

3. The automatic reconstruction method of array ground penetrating radar target based on image segmentation according to claim 1 is characterized by: In step 1, the data preprocessing method is automatic gain, which is achieved by multiplying the radar data by the automatic gain weight function that varies with time: In the above formula, represents the radar data after automatic gain, P(t) represents the automatic gain weight function, and y(t) represents the radar data before automatic gain; The number of time windows is determined by the number of parameter control points. When calculating the average amplitude, half of the time window should be overlapped between every two adjacent time windows. The following parameters are calculated based on the number of control points: time window length, start time of the time window, and end time of the time window. The average amplitude A in the sth time window s The calculation formula is: In the above formula, q(t) represents the single-channel data collected by the radar to be processed; T s-1 Indicates the start time of the sth time window; T s+1 represents the end time of the sth time window; B represents the number of samples in the sth time window; Finally, the average amplitude of each time window corresponding to the center of the respective time window is stored as the gain parameter of the control point; The weighting function is: In the above formula, p s represents the weighting factor corresponding to the center of the sth time window, and F represents the balance coefficient used to adjust the effective amplitude after processing. The weighting factors of points other than the center of the time window are obtained by linearly interpolating the weighting factors of the centers of two adjacent time windows.

4. The automatic reconstruction method of array ground penetrating radar target based on image segmentation according to claim 1 is characterized by: In step 2, images with hollow targets in each 3D data profile are captured; the number of captured images is 6000, the data size is 10*512, and the image size generated by linear interpolation is 130*512.

5. The automatic reconstruction method of array ground penetrating radar target based on image segmentation according to claim 1, characterized in that: The step 3 specifically includes the following steps: Step 31, using the open source image annotation tool labelme to annotate underground cavity targets in the radar image; Step 32, combining the annotated radar image and the annotation information to establish an underground cavity target radar image database; Step 33, using a data enhancement method to expand the data volume, the data enhancement method includes horizontally flipping the image and scaling the image; Step 34, randomly distribute the data in a ratio of 7:2:1 to obtain the final training set, validation set, and test set.

6. The method for automatic reconstruction of array ground penetrating radar targets based on image segmentation according to claim 1, characterized in that: Step 4: During the deeplabv3+ model training process, the epoch is 300 and the Batch_Size is 16.

7. The automatic reconstruction method of array ground penetrating radar target based on image segmentation according to claim 1, characterized in that: In step 6, after the masks are spliced ​​into a three-dimensional array, the three-dimensional array is visualized through visualization software.

Images