Ground penetrating radar C-scan inversion method and system

Abstract

The invention discloses a ground penetrating radar C-scan inversion method and system. The method comprises the steps that XZ slice data, YZ slice data and XY slice data of a to-be-inverted target are acquired and preprocessed to construct a training data set; constructing a ground penetrating radar slice data inversion initial model based on a convolutional layer, a linear layer, a Patch Merging layer, a Patch Expanding layer, a projection layer, residual connection and jump connection, and training to obtain a ground penetrating radar slice data inversion model; inputting detection data obtained in the actual three-dimensional detection scene into the ground penetrating radar slice data inversion model for processing to obtain inversion slice data; and performing splicing and three-dimensional reconstruction on the obtained inversion slice data to complete C-scan inversion of the ground penetrating radar. According to the method, the detection data of the XZ slice, the YZ slice and the XY slice are processed by adopting the constructed inversion model, and splicing and three-dimensional reconstruction are combined, so that the C-scan inversion of the ground penetrating radar is realized, the reliability is higher, the accuracy is better, and the efficiency is higher.

Description

Technical Field

[0001] The present invention belongs to the field of geological exploration, and particularly relates to a ground penetrating radar C-scan inversion method and system. Background Art

[0002] A ground penetrating radar is a commonly used non-destructive testing device. It detects internal defects of a structure by transmitting and receiving electromagnetic waves, thereby inferring the internal condition of the structure. Ground penetrating radar inversion can intuitively and clearly reflect the shape, size and position distribution of the detected object, and transform the interpretation of the radar profile into the analysis of the dielectric parameter image, which is beneficial to analyzing the shape information of the target. Therefore, the inversion scheme of the ground penetrating radar is of great significance to the ground penetrating radar.

[0003] At the present stage, the traditional two-dimensional inversion methods of ground penetrating radar mainly include the RTM imaging algorithm and two-dimensional full waveform inversion (FWI). RTM was first applied to seismic wave inversion. Due to the similar kinematic and dynamic characteristics of the electromagnetic waves of the ground penetrating radar, it has gradually been introduced into the field of ground penetrating radar. Two-dimensional FWI is a quantitative inversion technique, which was initially applied to seismic wave inversion. In recent years, with the improvement of computing power, it has been introduced into the inversion of ground penetrating radar data. In addition, through deep learning technology, the analysis and processing ability of ground penetrating radar data has been significantly improved. For example, the method based on a deep neural network takes the B-scan data of the ground penetrating radar as input and realizes the inversion of the dielectric constant of the detection area.

[0004] However, the above ground penetrating radar inversion schemes are mainly for two-dimensional inversion; for more complex three-dimensional scenarios, the existing schemes usually jointly analyze multiple B-scan images, form a three-dimensional C-scan from multiple two-dimensional B-scans, and directly map the C-scan to a three-dimensional spatial structure through a three-dimensional inversion network. However, the number of network parameters of such schemes is extremely large and the system efficiency is low; moreover, due to the different input dimensions, the two-dimensional inversion network with fewer parameters cannot be applied to three-dimensional inversion. Summary of the Invention

[0005] One object of the present invention is to provide a ground penetrating radar C-scan inversion method with high reliability, good accuracy and high efficiency.

[0006] Another object of the present invention is to provide a system for implementing the ground penetrating radar C-scan inversion method.

[0007] The ground penetrating radar C-scan inversion method provided by the present invention includes the following steps: S1. In a three-dimensional detection scenario, obtain the XZ slice, YZ slice and XY slice data of the target to be inverted from the ground penetrating radar C-scan data; S2. Preprocess the three-view detection data obtained in step S1 to construct a training dataset; S3. Based on convolutional layers, linear layers, Patch Merging layers, Patch Expanding layers, projection layers, residual connections, and skip connections, construct an initial model for ground penetrating radar slice data inversion; The constructed initial model for ground penetrating radar slice data inversion includes an embedding layer, an encoder, a decoder, and a projection layer; the embedding layer adjusts the channels of the input data; the encoder is used to perform feature extraction and downsampling operations on the input data with adjusted channels; the decoder is used to perform feature decoding and upsampling operations on the features output by the encoder; the projection layer is used to restore the features output by the decoder; S4. Use the training dataset obtained in step S2 to train the initial model for ground penetrating radar slice data inversion constructed in step S3 to obtain a ground penetrating radar slice data inversion model; S5. Input the detection data obtained in the actual three-dimensional detection scenario into the ground penetrating radar slice data inversion model obtained in step S4 for processing to obtain inverted slice data; S6. Stitch and three-dimensionally reconstruct the inverted slice data obtained in step S5 to complete the C-scan inversion of the ground penetrating radar.

[0008] The specific steps of step S1 are as follows: In a three-dimensional detection scenario, set a number of parallel detection lines on the OXZ plane where the ground surface is located; Based on the set parallel detection lines, collect several groups of B-scan data to form C-scan data; In the B-scan data collected at the first parallel detection line, select the scattering curve region of the potential target to be inverted and use it as the target to be tracked; Use a tracking algorithm to track the target to be tracked in the B-scan collected at the subsequent parallel detection lines to obtain the target box of the target scattering curve region in the B-scan; the length of the target box is The width of the target box is The coordinates of the upper left corner point of the target box are The target box is represented as ; The B-scan data corresponds to the XY slice data; Slice the views of different dimensions to obtain XZ slice data and YZ slice data: In the Top-scan data, the length of the target box is The width of the target box is N, the coordinates of the upper left corner point of the target box are The coordinates of the lower right corner point of the target box are The target box is represented as , where the Top-scan data corresponds to the XZ slice data; In the Crossline-scan data, the length of the target box is N, and the width of the target box is , and the coordinates of the upper left corner point of the target box are , and the coordinates of the lower right corner point of the target box are , and the target box is represented as , where the Crossline-scan data corresponds to the YZ slice data.

[0009] The step S2 specifically includes the following steps: The XZ slice data, YZ slice data, and XY slice data obtained in step S1 are adjusted to a set size through image size operations , where A is the side length of the square with the set size; the image size operations include padding, cropping, and scaling; First, calculate the expansion parameter p as ; then calculate the side length L of the square area to be cropped as ; next, calculate the scaling factor s as ; During cropping, judge whether it exceeds the boundary of the data according to the center position and side length information of the target: If it exceeds the boundary of the data, the exceeded part is filled with the average pixel value of the picture; If it does not exceed the boundary of the data, no filling is performed; Adjust the size of the square area to the set size according to the scaling factor s ; Slice the three-dimensional spatial structure model of the target based on the positions of each survey line; to ensure that the size of the target model remains unchanged, set a matrix with a set size, align the center of the slice data with the label matrix, assign the slice data to the label matrix correspondingly, and at the same time set the value of the target area in the label matrix to 1 and the value of the background area in the label matrix to 0; finally, use the label matrix as the inversion label.

[0010] The step S3 includes the following steps: Use the Patch Embedding layer as the embedding layer to adjust the channels of the input data; Based on the linear layer, Patch Merging layer, and residual connection, construct an encoder to perform feature extraction and downsampling operations on the input data; Based on the Patch Expanding layer and linear layer, construct a decoder to perform feature decoding and upsampling operations on the features output by the encoder; a skip connection is used between the encoder and the decoder; The projection layer is used to restore the features output by the decoder.

[0011] The initial model for the inversion of ground penetrating radar slice data includes the following: The inputs of the model include the preprocessed XZ slice data, YZ slice data, and XY slice data; The preprocessed XZ slice data, YZ slice data, and XY slice data are input to the encoder after being processed by their respective Patch Embedding layers; The encoder includes an XZ slice data sub-encoder, a YZ slice data sub-encoder, and an XY slice data sub-encoder; the structures of the XZ slice data sub-encoder, YZ slice data sub-encoder, and XY slice data sub-encoder are the same; the input of the XZ slice data sub-encoder is the XZ slice data processed by the Patch Embedding layer, the input of the YZ slice data sub-encoder is the YZ slice data processed by the Patch Embedding layer, and the input of the XY slice data sub-encoder is the XY slice data processed by the Patch Embedding layer; The processing process of the sub-encoder includes: the input slice data is processed by the first feature extraction module, the first Patch Merging layer, and the first linear layer to obtain the first sub-feature; the first sub-feature is processed by the second feature extraction module, the second Patch Merging layer, and the second linear layer to obtain the second sub-feature; the second sub-feature is processed by the third feature extraction module, the third Patch Merging layer, and the third linear layer to obtain the third sub-feature; the third sub-feature is processed by the fourth feature extraction module to obtain the fourth sub-feature; among them, the Patch Merging layer and the linear layer are both used for downsampling operations; The fourth sub-features of the XZ slice data, YZ slice data, and XY slice data output by the encoder are added together and used as the input to the decoder; The processing process of the decoder includes: after the input of the decoder is processed by the fifth feature extraction module, it is added to the third sub-feature of the XZ slice data, the third sub-feature of the YZ slice data, and the third sub-feature of the XY slice data output by the encoder to obtain the fifth sub-feature; after the fifth sub-feature is processed by the sixth Patch Expanding layer, the sixth linear layer, and the sixth feature extraction module, it is added to the second sub-feature of the XZ slice data, the second sub-feature of the YZ slice data, and the second sub-feature of the XY slice data output by the encoder to obtain the sixth sub-feature; after the sixth sub-feature is processed by the seventh Patch Expanding layer, the seventh linear layer, and the seventh feature extraction module, it is added to the first sub-feature of the XZ slice data, the first sub-feature of the YZ slice data, and the first sub-feature of the XY slice data output by the encoder to obtain the seventh sub-feature; the seventh sub-feature is processed by the eighth Patch Expanding layer, the eighth linear layer, and the eighth feature extraction module to obtain the output of the encoder; wherein, both the Patch Expanding layer and the linear layer are used for upsampling operations; The output of the encoder is processed by the projection layer to restore the feature channels, and finally the inversion slice data is obtained.

[0012] The structures of the first feature extraction module to the eighth feature extraction module are the same; the processing process of the feature extraction module includes the following steps: The input feature data is processed by the normalization layer and then divided into G channels; Each channel of data is subjected to feature extraction through the feature extraction sub-module; The obtained G-channel feature extraction data is concatenated, and then processed by the normalization layer and the linear layer to obtain the feature extraction module.

[0013] The processing process of the feature extraction sub-module includes the following steps: The input feature data is processed by the first normalization layer and then divided into two channels: The first channel of data is processed by the first linear layer and the first activation function layer to obtain the first branch data; The second channel of data is successively processed by the second linear layer, the depthwise separable convolution layer, the two-dimensional selective scanning module SS2D, and the second normalization layer to obtain the second branch data; After the first branch data and the second branch data are multiplied element by element, they are then processed by the third linear layer to obtain the third branch data; The third branch data and the feature data input to the feature extraction sub-module are added as the output of the feature extraction sub-module.

[0014] The training described in step S4 specifically includes the following steps: The following formula is used as the loss function for the training process: Where L is the value of the loss function; n is the number of samples; is the actual label of the i-th sample; is the predicted label of the i-th sample.

[0015] The 3D reconstruction described in step S6 specifically includes the following steps: For the data obtained after splicing, the Marching Cubes algorithm is used for 3D reconstruction.

[0016] The present invention also provides a system for implementing the ground penetrating radar C-scan inversion method, including a data acquisition module, a data processing module, a model construction module, a model training module, a model inversion module, and a 3D reconstruction module; the data acquisition module, the data processing module, the model construction module, the model training module, the model inversion module, and the 3D reconstruction module are connected in series in sequence; the data acquisition module is used to obtain the XZ slice, YZ slice, and XY slice data of the target to be inverted from the ground penetrating radar C-scan data in a 3D detection scenario, and upload the data information to the data processing module; the data processing module is used to perform data preprocessing on the obtained three-view detection data according to the received data information to construct a training data set, and upload the data information to the model construction module; the model construction module is used to construct an initial model for inverting the ground penetrating radar slice data based on a convolutional layer, a linear layer, a Patch Merging layer, a Patch Expanding layer, a projection layer, a residual connection, and a skip connection according to the received data information, and upload the data information to the model training module; the constructed initial model for inverting the ground penetrating radar slice data includes an embedding layer, an encoder, a decoder, and a projection layer; the embedding layer adjusts the channels of the input data; the encoder is used to perform feature extraction and downsampling operations on the input data with adjusted channels; the decoder is used to perform feature decoding and upsampling operations on the features output by the encoder; the projection layer is used to restore the features output by the decoder; the model training module is used to train the constructed initial model for inverting the ground penetrating radar slice data with the obtained training data set according to the received data information to obtain a model for inverting the ground penetrating radar slice data, and upload the data information to the model inversion module; the model inversion module is used to input the detection data obtained in the actual 3D detection scenario into the obtained model for inverting the ground penetrating radar slice data for processing according to the received data information to obtain inverted slice data, and upload the data information to the 3D reconstruction module; the 3D reconstruction module is used to splice and perform 3D reconstruction on the obtained inverted slice data according to the received data information to complete the C-scan inversion of the ground penetrating radar.

[0017] The ground penetrating radar C-scan inversion method and system provided by the present invention processes the detection data of XZ slices, YZ slices and XY slices by using the constructed inversion model, and combines stitching and three-dimensional reconstruction, which not only realizes the inversion of ground penetrating radar C-scan, but also greatly reduces the number of parameters of the inversion network and has higher efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic flowchart of the method of the present invention.

[0019] Figure 2 It is a schematic diagram of three-view data for extracting potential targets of the method of the present invention.

[0020] Figure 3 It is a schematic diagram of the detection scene of the embodiment of the method of the present invention.

[0021] Figure 4 It is a schematic diagram of each frame of C-scan of the embodiment of the method of the present invention. Figure 4 (1)~ Figure 4 (20) are schematic diagrams of the data of 20 groups of B-scan that make up the C-scan.

[0022] Figure 5 It is a schematic diagram of the tracking result of B-scan of the embodiment of the method of the present invention. Figure 5 (1)~ Figure 5 (20) are schematic diagrams of the data of 20 groups of B-scan.

[0023] Figure 6 It is a schematic diagram of the tracking result of Top-scan of the embodiment of the method of the present invention. Figure 6 (1)~ Figure 6 (20) are schematic diagrams of the data of 20 groups of Top-scan.

[0024] Figure 7 It is a schematic diagram of the tracking result of Crossline-scan of the embodiment of the method of the present invention. Figure 7 (1)~ Figure 7 (20) are schematic diagrams of the data of 20 groups of Crossline-scan.

[0025] Figure 8 It is a schematic diagram of the three-dimensional reconstruction result of the embodiment of the method of the present invention, where Figure 8 (1) is a schematic diagram of the model label of the high-resolution STL file, Figure 8 (2) is a schematic diagram of the low-resolution label (the label corresponding to the inversion network) composed of the target slices stitched and reconstructed at each survey line, Figure 8 (3) is a schematic diagram of the inversion result of the three-dimensional reconstruction of the present invention, Figure 8(4) Schematic diagram of the comparison effect between the three-dimensional inversion result after reconstruction of the present invention and the high-resolution label stack.

[0026] Figure 9 Schematic diagram of the functional modules of the system of the present invention. Specific implementation manners

[0027] As Figure 1 Shown is the schematic diagram of the method flow of the method of the present invention: The ground penetrating radar C-scan inversion method disclosed by the present invention includes the following steps: S1. In a three-dimensional detection scenario, obtain the XZ slice, YZ slice, and XY slice data of the target to be inverted from the ground penetrating radar C-scan data; specifically including the following steps: The data from three perspectives is as Figure 2 shown, namely the XZ slice (Top-scan), YZ slice (Crossline-scan), and XY slice (B-scan); where Top-scan is sliced in the depth dimension, Crossline-scan is the slice survey line perpendicular to the radar scanning direction, and B-scan is the slice along the radar detection survey line direction; In a three-dimensional detection scenario, set a number of parallel detection lines in the OXZ plane where the ground surface is located; Based on the set parallel detection lines, collect several groups of B-scan data to form C-scan data; In the B-scan data collected at the first parallel detection line, select the scattering curve region of the potential target to be inverted and use it as the target to be tracked; Adopt a tracking algorithm (such as the KCF algorithm) to track the target to be tracked in the B-scan collected at the subsequent parallel detection lines, and obtain the target box of the target scattering curve region in the B-scan; the length of the target box is and the width of the target box is and the coordinates of the upper left corner point of the target box are and the target box is represented as ; the B-scan data corresponds to the XY slice data; Slice the perspectives of different dimensions to obtain the XZ slice data and YZ slice data: Since when collecting and obtaining C-scan, it is usually sparse in the Z-axis direction, and all are included in the target box range under this perspective in this axis direction. Therefore, in the Top-scan data, the length of the target box is and the width of the target box is N, and the coordinates of the upper left corner point of the target box are and the coordinates of the lower right corner point of the target box are and the target box is represented as , where the Top-scan data corresponds to the XZ slice data; In the Crossline-scan data, the length of the target box is N, and the width of the target box is , and the coordinates of the upper left corner point of the target box are , and the coordinates of the lower right corner point of the target box are , and the target box is represented as , where the Crossline-scan data corresponds to the YZ slice data.

[0028] S2. Perform data preprocessing on the three-view detection data obtained in step S1 to construct a training data set; specifically, it includes the following steps: Adjust the XZ slice data, YZ slice data, and XY slice data obtained in step S1 to a set size through image size operations , where A is the side length of a square with a set size (preferably 255); the image size operations include padding, cropping, and scaling; First, calculate the expansion parameter p as ; then calculate the side length L of the square area to be cropped as ; next, calculate the scaling factor s as ; Judge whether it exceeds the boundary of the data according to the center position and side length information of the target: If it exceeds the boundary of the data, the exceeded part is filled with the average pixel value of the picture; If it does not exceed the boundary of the data, no filling is performed; Adjust the size of the square area to the set size according to the scaling factor s ; Slice the three-dimensional spatial structure model of the target based on the positions of each survey line; to ensure that the size of the target model remains unchanged, set a matrix with a set size (preferably ), align the center of the slice data with the label matrix, assign the slice data to the label matrix correspondingly, and at the same time set the value of the target area in the label matrix to 1 and the value of the background area in the label matrix to 0; finally, use the label matrix as the label of the data.

[0029] S3. Based on the convolutional layer, linear layer, Patch Merging layer, Patch Expanding layer, projection layer, residual connection, and skip connection, construct an initial model for the inversion of ground penetrating radar slice data; The initially constructed inversion model for ground penetrating radar slice data includes an embedding layer, an encoder, a decoder, and a projection layer; the embedding layer adjusts the channels of the input data; the encoder is used to perform feature extraction and downsampling operations on the input data with adjusted channels; the decoder is used to perform feature decoding and upsampling operations on the features output by the encoder; the projection layer is used to recover the features output by the decoder.

[0030] When specifically implemented, it includes the following steps: Use the Patch Embedding layer as the embedding layer to adjust the channels of the input data; Based on the linear layer, the Patch Merging layer, and the residual connection, construct an encoder to perform feature extraction and downsampling operations on the input data; Based on the Patch Expanding layer and the linear layer, construct a decoder to perform feature decoding and upsampling operations on the features output by the encoder; a skip connection is used between the encoder and the decoder; The projection layer is used to recover the features output by the decoder.

[0031] Among them, the initially constructed inversion model for ground penetrating radar slice data includes the following: The input of the model includes preprocessed XZ slice data, YZ slice data, and XY slice data; The preprocessed XZ slice data, YZ slice data, and XY slice data are input to the encoder after being processed by their respective Patch Embedding layers; The encoder includes a sub-encoder for XZ slice data, a sub-encoder for YZ slice data, and a sub-encoder for XY slice data; the structures of the sub-encoder for XZ slice data, the sub-encoder for YZ slice data, and the sub-encoder for XY slice data are the same; the input of the sub-encoder for XZ slice data is the XZ slice data processed by the Patch Embedding layer, the input of the sub-encoder for YZ slice data is the YZ slice data processed by the Patch Embedding layer, and the input of the sub-encoder for XY slice data is the XY slice data processed by the Patch Embedding layer; The processing process of the sub-encoder includes: after the input slice data is processed by the first feature extraction module, the first Patch Merging layer, and the first linear layer, the first sub-feature is obtained; after the first sub-feature is processed by the second feature extraction module, the second Patch Merging layer, and the second linear layer, the second sub-feature is obtained; after the second sub-feature is processed by the third feature extraction module, the third Patch Merging layer, and the third linear layer, the third sub-feature is obtained; after the third sub-feature is processed by the fourth feature extraction module, the fourth sub-feature is obtained; among them, both the Patch Merging layer and the linear layer are used for downsampling operations; After adding the fourth sub-feature of the XZ slice data, the fourth sub-feature of the YZ slice data, and the fourth sub-feature of the XY slice data output by the encoder, it is used as the input of the decoder; The processing process of the decoder includes: after the input of the decoder is processed by the fifth feature extraction module, it is added to the third sub-feature of the XZ slice data, the third sub-feature of the YZ slice data, and the third sub-feature of the XY slice data output by the encoder to obtain the fifth sub-feature; after the fifth sub-feature is processed by the sixth Patch Expanding layer, the sixth linear layer, and the sixth feature extraction module, it is added to the second sub-feature of the XZ slice data, the second sub-feature of the YZ slice data, and the second sub-feature of the XY slice data output by the encoder to obtain the sixth sub-feature; after the sixth sub-feature is processed by the seventh Patch Expanding layer, the seventh linear layer, and the seventh feature extraction module, it is added to the first sub-feature of the XZ slice data, the first sub-feature of the YZ slice data, and the first sub-feature of the XY slice data output by the encoder to obtain the seventh sub-feature; after the seventh sub-feature is processed by the eighth Patch Expanding layer, the eighth linear layer, and the eighth feature extraction module, the output of the encoder is obtained; among them, both the Patch Expanding layer and the linear layer are used for upsampling operations; The output of the encoder is processed by the projection layer to restore the feature channels, and finally the inversion slice data is obtained.

[0032] The structures of the first feature extraction module to the eighth feature extraction module are the same; the processing process of the feature extraction module includes the following steps: The input feature data is processed by the normalization layer and then divided into G channels; Each channel of data is subjected to feature extraction through the feature extraction sub-module; The obtained G-channel feature extraction data is concatenated, and then processed by the normalization layer and the linear layer to obtain the feature extraction module; In the traditional solution, the number of channels, the size of the SSM state dimension, the size of the internal 1D convolution kernel, the projection expansion multiplier, and the rank of the stride all affect the parameters, and the influence of the number of channels is explosive. Therefore, in this feature extraction module proposed by the present invention, based on the idea of grouping (corresponding to dividing into G paths of data for feature extraction), the feature extraction is realized; through the idea of grouping, the problem of the sharp increase in the number of parameters of the inversion network caused by the increase in the number of channels can be better avoided, and the number of parameters of the inversion model is greatly reduced.

[0033] The processing process of the feature extraction sub-module includes the following steps: After the input feature data is processed by the first normalization layer, it is divided into two paths: The first path of data is processed by the first linear layer and the first activation function layer to obtain the first branch data; The second path of data is sequentially processed by the second linear layer, the depthwise separable convolution layer, the two-dimensional selective scanning module SS2D, and the second normalization layer to obtain the second branch data; After the first branch data and the second branch data are element-wise multiplied, they are then processed by the third linear layer to obtain the third branch data; The third branch data and the feature data input to the feature extraction sub-module are added as the output of the feature extraction sub-module; This feature extraction sub-module provided by the present invention is more efficient than Transformer and has a linear time complexity; compared with the CNN network, this feature extraction sub-module provided by the present invention can capture long-term dependencies.

[0034] S4. Use the training data set obtained in step S2 to train the initial model for ground penetrating radar slice data inversion constructed in step S3 to obtain a ground penetrating radar slice data inversion model; During specific training, it includes the following steps: Use the following formula as the loss function during the training process: In the formula, L is the value of the loss function; n is the number of samples; is the actual label of the i-th sample; is the predicted label of the i-th sample.

[0035] S5. Input the detection data obtained in the actual three-dimensional detection scene into the ground penetrating radar slice data inversion model obtained in step S4 for processing to obtain the inversion slice data.

[0036] S6. Stitch and three-dimensionally reconstruct the inversion slice data obtained in step S5 to complete the C-scan inversion of the ground penetrating radar; among them, during three-dimensional reconstruction, the Marching Cubes algorithm is used for three-dimensional reconstruction of the stitched data.

[0037] The method of the present invention can directly use a two-dimensional inversion network to perform three-dimensional shape inversion on C-scan data, greatly alleviating the problem of the huge number of parameters in the three-dimensional inversion network. Therefore, the operating memory occupancy rate of the present invention is lower. At the same time, the method of the present invention designs a method for extracting three-view data of potential targets, and uses the three-view data to provide more prior information for the two-dimensional inversion network. Therefore, the accuracy of shape inversion is greatly improved.

[0038] The following further illustrates the method of the present invention in conjunction with an embodiment: This embodiment is carried out on a workstation equipped with a 13th Gen Intel(R) Core(TM) i9-13900K CPU, 32 GB of RAM, and a GTX 4090 GPU. The method of the present invention is implemented based on the Python and PyTorch frameworks, and RMSprop is used as the weight optimization algorithm, with the learning rate set to 1e-5.

[0039] First, a three-dimensional target model was randomly generated by 3DsMax software and saved as an STL format file. Subsequently, when creating the input file, a dielectric constant model with a size of 2.0 m in length and 1 m in depth was constructed, and the shape of the three-dimensional model slice corresponding to the survey line was selected as the target at the survey line. The transmitter and receiver of the ground-penetrating radar were set at a height of 0.9 m from the ground and continuously moved from left to right along the scanning trajectory to obtain B-scan data. The antenna was emitted using a Ricker wavelet with a center frequency of 500 MHz. The distance between the transmitter and the receiver was set to 0.02 m, and the antenna moved 0.02 m per step. The sampling time window length was 26 ns. The finally generated GPR B-scan data was a matrix with a size of In the FDTD forward simulation, the spatial spacing of the grid was 0.01 m, and 10 grid points were set around each side as the perfectly matched layer (PML) absorption boundary condition. Each dielectric constant model contained grid points. Finally, a simulation data set was generated by applying the finite-difference time-domain (FDTD) algorithm through Gprmax software.

[0040] As Figure 3 shown is a model diagram of the detection scenario of the embodiment of the present invention. In this example, two targets are used as an example. There are two cylindrical targets in the detection scenario. 20 horizontal survey lines are set in the Z-axis direction to detect the B-scan data of the XY plane scenario. The 20 groups of B-scan data are as Figure 4 shown. Among them, in the B-scan data obtained on the first survey line, it can be found that there are 2 hyperbolic targets. In this example, the area where the scattering curve on the right is located is selected as the initial tracking target, as Figure 4As shown by the dashed box in (1); then, the subsequent B-scan is tracked by the KCF algorithm to obtain the position information of the region where the target scattering curve is located in each frame of the C-scan, and then cropping and extraction are performed based on the target position information. By following the above steps, the data of the region where the tracked target is located in each B-scan can be obtained. For the convenience of training the subsequent inversion algorithm, the target scattering curve region data needs to undergo operations such as padding, cropping, and scaling before being input into the inversion network to ensure that the aspect ratio of the selected potential target does not change.

[0041] After the region where the target scattering curve is located in the B-scan undergoes operations such as padding, cropping, and scaling, the result is as Figure 5 shown; then, slices are taken for different-dimensional perspectives to obtain data for three perspectives. After undergoing operations such as padding, cropping, and scaling in the Top-scan, the result is as Figure 6 shown; after undergoing operations such as padding, cropping, and scaling in the Crossline-scan, the result is as Figure 7 shown; Next, the label of the inversion network is constructed, and after passing through the inversion network, the shape slices of the selected targets at each survey line are stitched together to obtain volume data. Finally, Marching cubes is used to perform three-dimensional reconstruction on the slice data, and the result is as Figure 8 shown; In addition, the similarity between the two is evaluated by calculating the error value of ICP registration and the slope distance between them. The calculation process of the ICP registration error is as follows: First, the surface point data is converted into the point cloud format, and the best rigid body transformation (rotation and translation) between the two point sets is found to maximize the coincidence degree of the registered point cloud in space. Finally, the error value between the two is obtained. The smaller the error value, the more similar the appearance shapes are. In Figure 8 it is Figure 8 (1) and Figure 8 (4), the normalized ICP value between the three-dimensional inversion results after reconstruction is 0.02152. The theoretical maximum value of the normalized ICP index is , and the larger the value, the farther the distance and the less similar the two shapes are. Through this embodiment, it is shown that the three-dimensional inversion result of the present invention has a high similarity with the label.

[0042] In addition, the similarity of the three-dimensional model is judged by calculating the slope distance. The two vertex sets are traversed bidirectionally, and the following formula is used to calculate the average minimum neighborhood distance: This formula calculates the mean of the nearest point distances from each point x in the set to , and then reversely calculates the corresponding mean of the distances from each point y in to . Finally, the sum of the two is used as the similarity measure.Figure 8 (1) The normalized slope distance calculation value between the reconstructed three-dimensional inversion result of Figure 8 and (4) is 0.0332. The theoretical maximum value of the normalized slope distance index is 3. The larger the value, the farther the distance and the less similar the two shapes. Through this index, it can be shown that the present invention can use a two-dimensional inversion network to achieve shape inversion of the target.

[0043] As Figure 9 shown in the schematic diagram of the functional modules of the system of the present invention: The system for implementing the ground penetrating radar C-scan inversion method disclosed in the present invention includes a data acquisition module, a data processing module, a model construction module, a model training module, a model inversion module, and a three-dimensional reconstruction module; the data acquisition module, the data processing module, the model construction module, the model training module, the model inversion module, and the three-dimensional reconstruction module are connected in series in sequence; the data acquisition module is used to obtain the XZ slice, YZ slice, and XY slice data of the target to be inverted from the ground penetrating radar C-scan data in a three-dimensional detection scenario, and upload the data information to the data processing module; the data processing module is used to perform data preprocessing on the obtained three-view detection data according to the received data information to construct a training data set, and upload the data information to the model construction module; the model construction module is used to construct an initial model for inverting the ground penetrating radar slice data based on a convolutional layer, a linear layer, a Patch Merging layer, a Patch Expanding layer, a projection layer, a residual connection, and a skip connection according to the received data information, and upload the data information to the model training module; the constructed initial model for inverting the ground penetrating radar slice data includes an embedding layer, an encoder, a decoder, and a projection layer; the embedding layer adjusts the channels of the input data; the encoder is used to perform feature extraction and downsampling operations on the input data with adjusted channels; the decoder is used to perform feature decoding and upsampling operations on the features output by the encoder; the projection layer is used to restore the features output by the decoder; the model training module is used to train the constructed initial model for inverting the ground penetrating radar slice data with the obtained training data set according to the received data information to obtain a model for inverting the ground penetrating radar slice data, and upload the data information to the model inversion module; the model inversion module is used to input the detection data obtained in the actual three-dimensional detection scenario into the obtained model for inverting the ground penetrating radar slice data for processing according to the received data information to obtain inverted slice data, and upload the data information to the three-dimensional reconstruction module; the three-dimensional reconstruction module is used to splice and three-dimensionally reconstruct the obtained inverted slice data according to the received data information to complete the C-scan inversion of the ground penetrating radar.

Claims

1. A ground penetrating radar C-scan inversion method, characterized in that The steps include: S1. In a three-dimensional detection scenario, obtain XZ slice, YZ slice and XY slice data of the target to be inverted from the ground penetrating radar C-scan data; S2. Preprocessing the three-view detection data obtained in step S1 to construct a training data set; S3. Based on convolutional layer, linear layer, patch merging layer, patch expanding layer, projection layer, residual connection and skip connection, the initial model of GPR slice data inversion is constructed; The constructed initial model for inversion of GPR slice data includes an embedding layer, an encoder, a decoder and a projection layer; the embedding layer adjusts the channels of the input data; the encoder is used to perform feature extraction and downsampling operations on the input data after the channels are adjusted; The decoder is used to decode and upsample the features output by the encoder; the projection layer is used to restore the features output by the decoder; S4. Using the training data set obtained in step S2, the initial ground penetrating radar slice data inversion model constructed in step S3 is trained to obtain a ground penetrating radar slice data inversion model; S5. The detection data obtained in the actual three-dimensional detection scene is input into the ground penetrating radar slice data inversion model obtained in step S4 for processing to obtain inversion slice data; S6. The inversion slice data obtained in step S5 are spliced ​​and three-dimensionally reconstructed to complete the C-scan inversion of the ground penetrating radar.

2. The ground penetrating radar C-scan inversion method according to claim 1, characterized in that The step S1 specifically includes the following steps: In the three-dimensional detection scene, several parallel detection lines are set on the OXZ plane where the surface is located; Based on the set parallel detection lines, several groups of B-scan data are collected to form C-scan data; In the B-scan data collected at the first parallel detection line, the scattering curve area of ​​the potential target to be inverted is selected and used as the target to be tracked; Using the tracking algorithm, the target to be tracked is tracked in the B-scan collected at the subsequent parallel detection line, and the target frame of the target scattering curve area in the B-scan is obtained; the length of the target frame is , the width of the target box is , the coordinates of the upper left corner of the target box are , the target box is represented as ; B-scan data corresponds to XY slice data; Slice the view in different dimensions to obtain XZ slice data and YZ slice data: In the Top-scan data, the length of the target box is , the width of the target box is N, and the coordinates of the upper left corner of the target box are , the coordinates of the lower right corner of the target box are , the target box is represented as , where Top-scan data corresponds to XZ slice data; In the Crossline-scan data, the length of the target box is N and the width of the target box is , the coordinates of the upper left corner of the target box are , the coordinates of the lower right corner of the target box are , the target box is represented as , where the Crossline-scan data corresponds to the YZ slice data.

3. The ground penetrating radar C-scan inversion method according to claim 2, characterized in that The step S2 specifically includes the following steps: The XZ slice data, YZ slice data and XY slice data obtained in step S1 are adjusted to the set size through image size operation. , A is the side length of a square of set size; the image size operation includes filling, cropping and scaling; First, the expansion parameter p is calculated as ; Then calculate the side length L of the square area to be cut out as ; Next, the scaling factor s is calculated as ; When clipping, determine whether it exceeds the data boundary based on the target's center position and side length information: If it exceeds the data boundary, the excess part will be filled with the average pixel value of the image; If the data does not exceed the boundary, no padding is performed; Adjust the size of the square area to the set size according to the scaling factor s ; The three-dimensional spatial structure model of the target is sliced ​​based on the position of each survey line; to ensure that the size of the target model remains unchanged, a matrix of a set size is set, the center of the slice data and the label matrix are aligned, the slice data is assigned to the label matrix accordingly, and the value of the target area in the label matrix is ​​set to 1, and the value of the background area in the label matrix is ​​set to 0; finally, the label matrix is ​​used as the inversion label.

4. The ground penetrating radar C-scan inversion method according to claim 3, characterized in that The step S3 comprises the following steps: The Patch Embedding layer is used as the embedding layer to adjust the channels of the input data; Based on the linear layer, patch merging layer and residual connection, an encoder is constructed to perform feature extraction and downsampling operations on the input data; Based on the Patch Expanding layer and the linear layer, a decoder is constructed to perform feature decoding and upsampling operations on the features output by the encoder; a skip connection is used between the encoder and the decoder; The projection layer is used to restore the features output by the decoder.

5. The ground penetrating radar C-scan inversion method according to claim 4, characterized in that The initial model for inversion of GPR slice data includes the following: The input of the model includes preprocessed XZ slice data, YZ slice data and XY slice data; The preprocessed XZ slice data, YZ slice data, and XY slice data are processed by their respective Patch Embedding layers and then input to the encoder; The encoder includes an XZ slice data sub-encoder, a YZ slice data sub-encoder and an XY slice data sub-encoder; the structures of the XZ slice data sub-encoder, the YZ slice data sub-encoder and the XY slice data sub-encoder are the same; the input of the XZ slice data sub-encoder is the XZ slice data processed by the Patch Embedding layer, the input of the YZ slice data sub-encoder is the YZ slice data processed by the Patch Embedding layer, and the input of the XY slice data sub-encoder is the XY slice data processed by the Patch Embedding layer; The processing process of the sub-encoder includes: the input slice data is processed by the first feature extraction module, the first Patch Merging layer and the first linear layer to obtain the first sub-feature; the first sub-feature is processed by the second feature extraction module, the second Patch Merging layer and the second linear layer to obtain the second sub-feature; the second sub-feature is processed by the third feature extraction module, the third Patch Merging layer and the third linear layer to obtain the third sub-feature; the third sub-feature is processed by the fourth feature extraction module to obtain the fourth sub-feature; wherein the Patch Merging layer and the linear layer are both used for downsampling operation; The fourth sub-feature of the XZ slice data, the fourth sub-feature of the YZ slice data, and the fourth sub-feature of the XY slice data output by the encoder are added together and used as the input of the decoder; The processing process of the decoder includes: after the input of the decoder is processed by the fifth feature extraction module, the decoder is added with the third sub-feature of the XZ slice data, the third sub-feature of the YZ slice data and the third sub-feature of the XY slice data output by the encoder to obtain the fifth sub-feature; after the fifth sub-feature is processed by the sixth Patch Expanding layer, the sixth linear layer and the sixth feature extraction module, the decoder is added with the second sub-feature of the XZ slice data, the second sub-feature of the YZ slice data and the second sub-feature of the XY slice data output by the encoder to obtain the sixth sub-feature; after the sixth sub-feature is processed by the seventh Patch Expanding layer, the seventh linear layer and the seventh feature extraction module, the decoder is added with the first sub-feature of the XZ slice data, the first sub-feature of the YZ slice data and the first sub-feature of the XY slice data output by the encoder to obtain the seventh sub-feature; the seventh sub-feature is processed by the eighth PatchExpanding layer, the eighth linear layer and the eighth feature extraction module to obtain the output of the encoder; wherein the PatchExpanding layer and the linear layer are both used for upsampling operations; The output of the encoder is processed through the projection layer to restore the feature channels and finally obtain the inverted slice data.

6. The ground penetrating radar C-scan inversion method according to claim 5, characterized in that The structures of the first feature extraction module to the eighth feature extraction module are the same; the processing process of the feature extraction module includes the following steps: The input feature data is processed by the normalization layer and divided into G paths; Each channel of data is subjected to feature extraction through the feature extraction submodule; The obtained G-path feature extraction data is concatenated and then processed through a normalization layer and a linear layer to obtain a feature extraction module.

7. The ground penetrating radar C-scan inversion method according to claim 6, characterized in that The process of extracting submodules includes the following steps: After the input feature data is processed by the first normalization layer, it is divided into two paths: The first path of data is processed by a first linear layer and a first activation function layer to obtain first branch data; The second path of data is processed in sequence by a second linear layer, a depth-separable convolutional layer, a two-dimensional selective scanning module SS2D, and a second normalization layer to obtain second branch data; The first branch data and the second branch data are element-wise multiplied and then processed through a third linear layer to obtain third branch data; The third branch data is added to the input feature data of the feature extraction submodule to serve as the output of the feature extraction submodule.

8. The ground penetrating radar C-scan inversion method according to claim 7, characterized in that The training described in step S4 specifically includes the following steps: The following formula is used as the loss function of the training process: Where L is the loss function value; n is the number of samples; is the actual label of the i-th sample; is the predicted label of the i-th sample.

9. The ground penetrating radar C-scan inversion method according to claim 8, characterized in that The three-dimensional reconstruction described in step S6 specifically includes the following steps: The Marching Cubes algorithm is used to reconstruct the three-dimensional data obtained after splicing.

10. A system for implementing the ground penetrating radar C-scan inversion method according to any one of claims 1 to 9, characterized in that It includes a data acquisition module, a data processing module, a model building module, a model training module, a model inversion module and a three-dimensional reconstruction module; the data acquisition module, the data processing module, the model building module, the model training module, the model inversion module and the three-dimensional reconstruction module are connected in series in sequence; the data acquisition module is used to obtain XZ slices, YZ slices and XY slices of the target to be inverted from the ground penetrating radar C-scan data in a three-dimensional detection scene, and upload the data information to the data processing module; The data processing module is used to perform data preprocessing on the acquired three-view detection data according to the received data information to construct a training data set, and upload the data information to the model construction module; The model building module is used to build an initial model for inversion of ground penetrating radar slice data based on the received data information, based on the convolution layer, linear layer, patch merging layer, patch expanding layer, projection layer, residual connection and skip connection, and upload the data information to the model training module; The constructed initial model for inversion of GPR slice data includes an embedding layer, an encoder, a decoder and a projection layer; the embedding layer adjusts the channels of the input data; the encoder is used to perform feature extraction and downsampling operations on the input data after the channels are adjusted; The decoder is used to decode and upsample the features output by the encoder; the projection layer is used to restore the features output by the decoder; The model training module is used to train the constructed GPR slice data inversion initial model based on the received data information and the obtained training data set, obtain the GPR slice data inversion model, and upload the data information to the model inversion module; The model inversion module is used to input the detection data obtained in the actual three-dimensional detection scene into the obtained ground penetrating radar slice data inversion model for processing according to the received data information, obtain the inverted slice data, and upload the data information to the three-dimensional reconstruction module; the three-dimensional reconstruction module is used to splice and three-dimensionally reconstruct the obtained inversion slice data according to the received data information to complete the C-scan inversion of the ground penetrating radar.

Images