Reconstruction method and system for cavity under shallow reinforcement net shielding

Abstract

The application discloses a reconstruction method and system for a cavity under a shallow reinforcement net shelter, and comprises the following steps: constructing a layered medium scene containing a shallow reinforcement net and a cavity; acquiring image data information with different shelter states and preprocessing to construct a reconstruction data set; constructing a reconstruction initial model for the cavity under the shallow reinforcement net shelter and training to obtain a reconstruction model for the cavity under the shallow reinforcement net shelter; and using the obtained reconstruction model for the cavity under the shallow reinforcement net shelter to perform actual reconstruction for the cavity under the shallow reinforcement net shelter. The application acquires multiple types of training data through the constructed layered medium scene containing the shallow reinforcement net and the cavity, and trains the reconstruction model for the cavity under the shallow reinforcement net shelter which comprises a U-shaped network, a Sobel operator, an attention mechanism, a residual structure, an average mechanism and a skip connection scheme. Therefore, the application can not only realize the reconstruction for the cavity under the shallow reinforcement net shelter, but also has higher reliability, better accuracy and higher reconstruction efficiency.

Description

Technical Field

[0001] This invention belongs to the field of ground-penetrating radar detection, specifically relating to a method and system for reconstructing cavities under shallow steel mesh shielding. Background Technology

[0002] Ground penetrating radar (GPR) is a non-destructive testing technique based on the principle of high-frequency electromagnetic wave reflection. It is suitable for detecting changes in the medium and abnormal reflectors within underground structures or underground environments. The working principle of GPR involves deploying an antenna system on the Earth's surface, emitting high-frequency pulsed electromagnetic waves, and receiving the signals reflected or scattered at the boundaries of different media, thereby obtaining information about the spatial structure underground or within the medium. GPR technology features moderate penetration depth, high resolution, and non-destructive operation, and is widely used in numerous fields.

[0003] However, in typical scenarios such as municipal roads and underground engineering projects, shallow layers often contain steel reinforcement mesh. As a strong reflector, the scattered signal of the steel reinforcement mesh significantly masks the echoes of weak targets beneath it, especially the reflection signals from cavities. Currently, researchers have proposed corresponding schemes for reconstructing underground structure images from acquired GPR B-scan images, mainly including migration algorithms, tomographic imaging schemes, full waveform inversion schemes, and deep learning-based schemes. For example, Ding Yilin et al. proposed a YOLO network detection algorithm, which combines simulated and real datasets to detect various defects even with steel reinforcement mesh occlusion; however, this scheme ignores the location and size of the defects. Zhao Huimin et al. used a pix2pix parameter inversion network to directly invert defects on a 3D tunnel similarity model containing shallow steel reinforcement mesh; the interference of the steel reinforcement mesh affected the inversion results to some extent.

[0004] In summary, the existing scheme does not take into account the interference suppression problem of steel mesh echoes in the layered background medium scenario on the inversion of underground cavities, resulting in low accuracy and reliability of cavity inversion. Summary of the Invention

[0005] One of the objectives of this invention is to provide a highly reliable, accurate, and efficient method for reconstructing voids under the cover of shallow steel mesh.

[0006] The second objective of this invention is to provide a system for reconstructing the voids under the cover of the shallow steel mesh.

[0007] The reconstruction method for concealing cavities under shallow steel mesh provided by this invention includes the following steps:

[0008] S1. Construct a layered medium scenario containing shallow steel mesh and voids;

[0009] S2. Based on the scene constructed in step S1, obtain image data information with different occlusion states;

[0010] S3. Preprocess the image data information obtained in step S2 to construct a reconstructed dataset;

[0011] S4. Construct an initial reconstruction model of the voids concealed by shallow steel mesh;

[0012] The initial model for reconstructing cavities under shallow steel mesh includes a first-stage model and a second-stage model;

[0013] A first-stage model is constructed based on a U-shaped network, the Sobel operator, residual structures, and a skip connection scheme; the first-stage model is used to suppress the echo of the steel mesh.

[0014] A second-stage model is constructed based on U-shaped networks, residual structures, attention mechanisms, averaging mechanisms, multi-scale edge fusion, and skip connection schemes. The second-stage model is used to reconstruct holes based on the output of the first-stage model.

[0015] S5. Using the reconstructed dataset obtained in step S3, train the initial reconstruction model of the void under the shallow steel mesh constructed in step S4 to obtain the reconstruction model of the void under the shallow steel mesh.

[0016] S6. Using the reconstruction model of the void under the shallow steel mesh obtained in step S5, the actual reconstruction of the void under the shallow steel mesh is carried out.

[0017] Step S1 specifically includes the following steps:

[0018] The constructed layered medium scenario consists of four layers, from top to bottom: an air layer, a surface layer, a base layer containing steel mesh echoes, and a soil base layer;

[0019] The surface layer is made of asphalt concrete, and is divided into three layers according to the different aggregate sizes, shapes, and random distribution of aggregate positions.

[0020] The base material is set to reinforced concrete, and the spacing and number of steel bars are set. The base is divided into two layers according to the different aggregate sizes and shapes.

[0021] The subgrade is set to a randomly mixed soil, and the soil surface is set to an uneven, undulating surface.

[0022] Step S2 specifically includes the following steps:

[0023] Based on the scenario constructed in step S1, several forward model files containing hollow targets are generated.

[0024] The types of cavities include rectangular cavities without water, rectangular cavities with water, irregularly shaped cavities without water, and irregularly shaped cavities with water; the cavities are set in the soil base layer; the size and depth of the cavities are randomly generated.

[0025] The acquired forward model files include B-scan images of underground cavities with shallow steel mesh covering them, B-scan images of underground cavities without shallow steel mesh covering them, and black-and-white mask images of the cavity targets.

[0026] Step S3 specifically includes the following steps:

[0027] Set the maximum reference threshold for the B-scan image;

[0028] Signal suppression processing: In the image data information obtained in step S2, all data higher than the maximum reference threshold are replaced with the maximum reference threshold to complete the suppression of strong signals;

[0029] Signal enhancement processing: Construct an exponential gain sequence that increases with depth, and multiply the data after signal suppression by the constructed exponential gain sequence to complete the signal enhancement processing;

[0030] Image scaling: The image data after signal enhancement is uniformly scaled to a set size;

[0031] Complete the preprocessing of the image data information obtained in step S2.

[0032] The first-stage model is constructed, including the first-stage encoder, the first-stage decoder, and the first-stage filtering module;

[0033] The first-stage encoder includes a first encoding first residual module, a first encoding first edge module, a first encoding first max pooling module, a first encoding second residual module, a first encoding second edge module, a first encoding second max pooling module, a first encoding third residual module, a first encoding third edge module, a first encoding third max pooling module, a first encoding fourth residual module, a first encoding fourth edge module, a first encoding fourth max pooling module, and a first encoding fifth residual module, which are connected in series.

[0034] The first-stage decoder comprises, in sequence, a first decoding first transposed convolutional module, a first decoding first residual module, a first decoding first edge module, a first decoding second transposed convolutional module, a first decoding second residual module, a first decoding second edge module, a first decoding third transposed convolutional module, a first decoding third residual module, a first decoding third edge module, a first decoding fourth transposed convolutional module, a first decoding fourth residual module, and a first decoding fourth edge module. Simultaneously, the output of the first encoding fourth edge module and the output of the first decoding first residual module are concatenated and used as the input of the first decoding first edge module; the output of the first encoding third edge module and the output of the first decoding second residual module are concatenated and used as the input of the first decoding second edge module; the output of the first encoding second edge module and the output of the first decoding third residual module are concatenated and used as the input of the first decoding third edge module; and the output of the first encoding first edge module and the output of the first decoding fourth residual module are concatenated and used as the input of the first decoding fourth edge module.

[0035] The first-stage filtering module includes a first-stage convolutional layer and a first-stage activation function layer connected in series, and the output of the first-stage decoder is the input of the first-stage filtering module;

[0036] The first-stage model is used to suppress the echo of the rebar mesh and output a GPR image without the echo of the rebar mesh.

[0037] The first stage of the convolutional layer is... The 2D convolutional layer; the first-stage activation function layer is a ReLU activation function layer; the first decoding first transposed convolutional module to the first decoding fourth transposed convolutional module are all... The transposed convolutional layer is used to amplify the resolution; the first encoding first max pooling module to the first encoding fourth max pooling module are all... A max-pooling layer with a stride of 2 is used to halve the size of the input feature map.

[0038] The first encoding first edge module to the first encoding fourth edge module, and the first decoding first edge module to the first decoding fourth edge module all have the same structure;

[0039] The processing of the edge module includes the following steps:

[0040] The Sobel operator is used to detect the horizontal and vertical edge information of the input image, and the absolute values ​​are taken to obtain the absolute horizontal and absolute vertical edge information respectively.

[0041] After concatenating the obtained absolute horizontal edge information and absolute vertical edge information with the input image information, ... The 2D convolutional layers are fused together, and then processed sequentially through a normalization layer and a ReLU activation function layer to obtain the output of the edge module.

[0042] The constructed second-stage model includes a second-stage encoder, a second-stage decoder, a second-stage averaging module, and a second-stage reconstruction module;

[0043] The second-stage encoder includes a second encoding first residual module, a second encoding first max pooling module, a second encoding second residual module, a second encoding second max pooling module, a second encoding third residual module, a second encoding third max pooling module, a second encoding fourth residual module, a second encoding fourth max pooling module, and a second encoding fifth residual module, which are connected in series.

[0044] The second-stage encoder comprises, in sequence, a second decoding first transposed convolutional module, a second decoding first residual module, a second decoding second transposed convolutional module, a second decoding second residual module, a second decoding third transposed convolutional module, a second decoding third residual module, a second decoding fourth transposed convolutional module, and a second decoding fourth residual module. Simultaneously, the output of the second decoding fourth residual module is processed by a fourth attention module and then concatenated with the output of the second decoding fifth residual module as the input to the second decoding first transposed convolutional module; the output of the second decoding third residual module is processed by a third attention module and then concatenated with the output of the second decoding first residual module as the input to the second decoding second transposed convolutional module; the output of the second decoding second residual module is processed by a second attention module and then concatenated with the output of the second decoding second residual module as the input to the second decoding third transposed convolutional module; and the output of the second decoding first residual module is processed by a first attention module and then concatenated with the output of the second decoding third residual module as the input to the second decoding fourth transposed convolutional module.

[0045] The second-stage averaging module includes a second-averaging first convolutional layer, a second-averaging second convolutional layer, a second-averaging third convolutional layer, and a second-averaging fourth convolutional layer. The output of the second encoding first residual module is processed by the second-averaging first convolutional layer to obtain a second-averaged first signal. The output of the second encoding second residual module is processed by the second-averaging second convolutional layer to obtain a second-averaged second signal. The output of the second encoding third residual module is processed by the second-averaging third convolutional layer to obtain a second-averaged third signal. The output of the second encoding fourth residual module is processed by the second-averaging fourth convolutional layer to obtain a second-averaged fourth signal. The average of the obtained second-averaged first signal to second-averaged fourth signal is calculated to obtain the output of the second-stage averaging module. Here, the second-averaged first signal to second-averaged fourth signal are all B-scan two-dimensional images of equal size. The average of these four images is used to obtain the output of the second-stage averaging module. The output of the second-stage averaging module is also a B-scan two-dimensional image of the same size.

[0046] The second-stage reconstruction module includes reconstructing the convolutional layer; the output of the second-stage decoder is processed by the reconstructed convolutional layer and then summed with the output of the second-stage averaging module to obtain the final dilated reconstruction result.

[0047] Reconstruct the convolutional layer as The 2D convolutional layer; the reconstructed activation function layer is a Sigmoid activation function layer; the second encoding first max pooling module to the second encoding fourth max pooling module are all... The first module is a max pooling module with a stride of 2; the second decoding first transposed convolution module and the second decoding first transposed convolution module are both... The transposed convolutional layer; the second average first convolutional layer to the second average fourth convolutional layer are all... 2D convolutional layers.

[0048] The structures of the first attention module through the fourth attention module are all the same;

[0049] The attention module's processing steps include the following:

[0050] Obtain the gated feature map of the input image and jump feature map Gated feature map The feature map output by the second decoding first residual module, the skip feature map The feature map output by the second encoding and third residual modules;

[0051] Will Passing in sequence The image is processed by 2D convolutional layers and normalization layers to obtain the first image component; Passing in sequence The second image component is obtained by processing 2D convolutional layers and normalization layers;

[0052] After summing the first and second image components, the image is then passed sequentially through a ReLU activation function layer. The image is processed by 2D convolutional layers, normalization layers, and sigmoid activation function layers to obtain the third image component;

[0053] Combine the third image component with After multiplying by pixels, the output of the attention module is obtained.

[0054] The structures of the first encoding first residual module to the first encoding fifth residual module, the first decoding first residual module to the first decoding fourth residual module, the second encoding first residual module to the second encoding fifth residual module, and the second decoding first residual module to the second decoding fourth residual module are all the same;

[0055] The residual module processing procedure includes the following steps:

[0056] The input image data is divided into two paths: the first path passes through sequentially... 2D convolutional layers, normalization layers, ReLU activation function layers, The first path of features is obtained by processing 2D convolutional layers and normalization layers; the second path is then processed sequentially... The second path features are obtained by processing the 2D convolutional layer and normalization layer;

[0057] The first and second features are added together, and then processed through a ReLU activation function layer to obtain the output of the residual module.

[0058] The training described in step S5 specifically includes the following steps:

[0059] During training, the pre-processed B-scan image of underground cavities with shallow steel mesh covering them was used as the input of the first-stage model, and the pre-processed B-scan image of underground cavities without shallow steel mesh covering them was used as the label for the input of the first-stage model. The first-stage model was then trained.

[0060] Then, the output of the first-stage model is used as the input of the second-stage model, and the preprocessed black-and-white mask image of the hole target is used as the label of the input of the second-stage model to train the second-stage model.

[0061] The loss function of the first-stage model is a weighted sum of the L1 loss function and the SSIM loss function;

[0062] The loss function of the second-stage model is a weighted sum of the BCE loss function, Dice loss function, Boundary loss function, and Focal loss function;

[0063] During training, the weighted sum of the loss functions of the first-stage model and the second-stage model is used as the total loss function.

[0064] During training, the Adam optimizer is used, and an adaptive learning rate mechanism and an early stopping strategy are employed to improve the stability and efficiency of the training process.

[0065] This invention also provides a system for reconstructing voids under the shading of shallow steel mesh, comprising a scene construction module, a data acquisition module, a data processing module, a model construction module, a model training module, and a void reconstruction module; the scene construction module, data acquisition module, data processing module, model construction module, model training module, and void reconstruction module are connected in series; the scene construction module is used to construct a layered medium scene containing shallow steel mesh and voids, and upload the data information to the data acquisition module; the data acquisition module is used to acquire image data information with different shading states based on the constructed scene according to the received data information, and upload the data information to the data processing module; the data processing module is used to preprocess the acquired image data information according to the received data information to construct a reconstruction dataset, and upload the data information to the model construction module; the model construction module is used to construct an initial reconstruction model of voids under the shading of shallow steel mesh according to the received data information, and... The data information upload and model training module includes a first-stage model and a second-stage model for reconstructing voids under shallow rebar mesh cover. The first-stage model is constructed based on a U-shaped network, Sobel operator, residual structure, and skip connection scheme. This first-stage model is used to suppress the echo of the rebar mesh. The second-stage model is constructed based on a U-shaped network, residual structure, attention mechanism, averaging mechanism, multi-scale edge fusion, and skip connection scheme. This second-stage model is used to reconstruct voids based on the output of the first-stage model. The model training module trains the initial model for reconstructing voids under shallow rebar mesh cover using the received data information and the obtained reconstruction dataset, obtaining a reconstruction model for voids under shallow rebar mesh cover. This data information is then uploaded to the void reconstruction module. The void reconstruction module uses the obtained reconstruction model for voids under shallow rebar mesh cover, based on the received data information, to perform actual reconstruction of voids under shallow rebar mesh cover.

[0066] The present invention provides a method and system for reconstructing cavities under shallow steel mesh shading. By constructing a layered medium scene containing shallow steel mesh and cavities, it acquires multiple types of training data and trains a reconstruction model for cavities under shallow steel mesh shading, including U-shaped networks, Sobel operators, attention mechanisms, residual structures, averaging mechanisms, and skip connection schemes. Therefore, the present invention can not only reconstruct cavities under shallow steel mesh shading, but also has higher reliability, better accuracy, and higher reconstruction efficiency. Attached Figure Description

[0067] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0068] Figure 2 This is a schematic diagram of the structure of the first stage model of the method of the present invention.

[0069] Figure 3 This is a schematic diagram of the edge module structure of the method of the present invention.

[0070] Figure 4 This is a schematic diagram of the structure of the second-stage model of the method of the present invention.

[0071] Figure 5 This is a schematic diagram of the attention module in the method of the present invention.

[0072] Figure 6 This is a schematic diagram of the residual module of the method of the present invention.

[0073] Figure 7 This is a schematic diagram of the rectangular hole reconstruction result according to an embodiment of the method of the present invention.

[0074] Figure 8 This is a schematic diagram comparing the network results with real data in the first stage of an embodiment of the method of the present invention; wherein, Figure 8 (a) is a schematic diagram of the output results of the first-stage network. Figure 8 (b) is a schematic diagram of the actual first-stage network labels obtained through simulation.

[0075] Figure 9 This is a schematic diagram comparing the second-stage network results with real data in an embodiment of the method of the present invention; wherein, Figure 9 (a) is a schematic diagram of the output results of the second-stage network. Figure 9 (b) is a schematic diagram of the actual second-stage network labels obtained through simulation.

[0076] Figure 10 This is a schematic diagram of the functional modules of the system of the present invention. Detailed Implementation

[0077] like Figure 1The diagram shown is a flowchart of the method of the present invention: The method for reconstructing cavities under shallow steel mesh concealment disclosed in this invention includes the following steps:

[0078] S1. Construct a layered medium scene containing shallow steel mesh and voids; specifically including the following steps:

[0079] The constructed layered medium scenario consists of four layers, from top to bottom: an air layer, a surface layer, a base layer containing steel mesh echoes, and a soil base layer;

[0080] The surface layer is made of asphalt concrete, and is divided into three layers according to the different aggregate sizes, shapes, and random distribution of aggregate positions.

[0081] The base material is set as reinforced concrete, and the spacing of the reinforcing bars (preferably 70cm, 40cm, 28cm and 24cm) and the number of bars (preferably 3, 5, 7 and 8 bars) are set. The base is divided into two layers according to the different aggregate particle size and the different aggregate shape.

[0082] The subgrade is set to a randomly mixed soil, and the soil surface is set to an uneven, undulating surface.

[0083] In a preferred embodiment, the constructed scene has a horizontal length of 2.0 meters and a vertical depth of 1.6 meters. The overall medium is divided into four layers: the first layer is an air layer with a thickness of 0.1 meters; the second layer is a surface layer with a thickness of 0.18 meters; the third layer is a base layer with a thickness of 0.52 meters; and the fourth layer is a soil base layer with a thickness of 0.80 meters. An irregular undulating surface is set above the roadbed, with a maximum vertical undulation height of 0.06 meters. The surface layer is divided into three sub-layers based on the characteristics of different aggregate particle sizes, shapes, and random distribution, with each sub-layer having a thickness of 0.04 meters, 0.06 meters, and 0.08 meters. The base layer is divided into two sub-layers based on the characteristics of different aggregate particle sizes, shapes, and random distribution, with each sub-layer having a thickness of 0.36 meters and 0.18 meters.

[0084] The reinforcing bars are laid horizontally at the junction of the asphalt and the base layer, i.e., in the surface layer, with a fixed radius of 0.005m, which does not change during the sample generation process; radar reflection background modeling of different complexities is achieved by controlling the number and spacing of the reinforcing bars in the reinforcing mesh.

[0085] In the model, the horizontal range is 0.5 to 1.5 meters and the vertical range is 0.9 to 1.08 meters. The center of the void is randomly embedded in the soil base layer and does not physically overlap with the steel mesh components; this ensures a certain degree of diversity and randomness in different positions, sizes and shapes.

[0086] Regarding the reconstruction of the location and size of the voids, considering the complexity of the dielectric constant of the scene and the uniformity of the dielectric constant of the void target, the layered medium scene is set to 1 and the voids are set to 0. The corresponding black and white mask image for reconstructing the location and size of the voids is obtained through the corresponding forward model file.

[0087] In the forward modeling parameter settings, the radar transmit waveform is selected as Ricker wavelet, the center frequency is set to 800MHz, and the time window is set to 27ns to ensure effective detection of deep targets. The antenna transmit / receive type is selected as solid aperture antenna.

[0088] S2. Based on the scene constructed in step S1, obtain image data information with different occlusion states; specifically including the following steps:

[0089] Based on the scenario constructed in step S1, several forward model files containing hollow targets are generated.

[0090] The types of cavities include rectangular cavities without water, rectangular cavities with water, irregularly shaped cavities without water, and irregularly shaped cavities with water; the cavities are set in the soil base layer; the size and depth of the cavities are randomly generated.

[0091] The acquired forward model files include B-scan images of underground cavities with shallow steel mesh covering them, B-scan images of underground cavities without shallow steel mesh covering them, and black-and-white mask images of the cavity targets.

[0092] S3. Preprocess the image data information obtained in step S2 to construct the reconstructed dataset; specifically including the following steps:

[0093] Set the maximum reference threshold for the B-scan image;

[0094] Signal suppression processing: Due to the strong reflection of the steel mesh, the echo of the weak target signal under the steel mesh is not obvious; therefore, in the image data information obtained in step S2, all data higher than the maximum reference threshold are replaced with the maximum reference threshold to complete the suppression of strong signals.

[0095] Signal enhancement processing: Holes are usually located in deeper layers with low relative permittivity, resulting in weak target echo signals; therefore, an exponential gain sequence that increases with depth is constructed, and the data after signal suppression is multiplied by the constructed exponential gain sequence to complete the signal enhancement processing.

[0096] Image scaling: The image data after signal enhancement is uniformly scaled to a set size;

[0097] Complete the preprocessing of the image data information obtained in step S2.

[0098] S4. Construct an initial reconstruction model of the voids concealed by shallow steel mesh;

[0099] Since the strong reflection of the steel mesh has a significant impact on the reconstruction accuracy of the voids, a two-stage model is constructed. The first stage is to suppress the steel mesh echo; the second stage is to reconstruct the void targets in the GPR B-scan image with the steel mesh echo suppressed in the first stage.

[0100] The initial model for reconstructing cavities under shallow steel mesh includes a first-stage model and a second-stage model;

[0101] A first-stage model is constructed based on a U-shaped network, the Sobel operator, residual structures, and a skip connection scheme; the first-stage model is used to suppress the echo of the steel mesh.

[0102] A second-stage model is constructed based on U-shaped networks, residual structures, attention mechanisms, averaging mechanisms, multi-scale edge fusion, and skip connection schemes. The second-stage model is used to reconstruct holes based on the output of the first-stage model.

[0103] In practical implementation, the first-stage model includes a first-stage encoder, a first-stage decoder, and a first-stage filtering module, with the specific structure as follows: Figure 2 As shown;

[0104] The first-stage encoder includes a first encoding first residual module, a first encoding first edge module, a first encoding first max pooling module, a first encoding second residual module, a first encoding second edge module, a first encoding second max pooling module, a first encoding third residual module, a first encoding third edge module, a first encoding third max pooling module, a first encoding fourth residual module, a first encoding fourth edge module, a first encoding fourth max pooling module, and a first encoding fifth residual module, which are connected in series.

[0105] The first-stage decoder comprises, in sequence, a first decoding first transposed convolutional module, a first decoding first residual module, a first decoding first edge module, a first decoding second transposed convolutional module, a first decoding second residual module, a first decoding second edge module, a first decoding third transposed convolutional module, a first decoding third residual module, a first decoding third edge module, a first decoding fourth transposed convolutional module, a first decoding fourth residual module, and a first decoding fourth edge module. Simultaneously, the output of the first encoding fourth edge module and the output of the first decoding first residual module are concatenated and used as the input of the first decoding first edge module; the output of the first encoding third edge module and the output of the first decoding second residual module are concatenated and used as the input of the first decoding second edge module; the output of the first encoding second edge module and the output of the first decoding third residual module are concatenated and used as the input of the first decoding third edge module; and the output of the first encoding first edge module and the output of the first decoding fourth residual module are concatenated and used as the input of the first decoding fourth edge module.

[0106] The first-stage filtering module includes a first-stage convolutional layer and a first-stage activation function layer connected in series, and the output of the first-stage decoder is the input of the first-stage filtering module;

[0107] The first-stage model is used to suppress the echo of the rebar mesh and output a GPR image without the echo of the rebar mesh.

[0108] In the first-stage model, shallow edge features of the encoder channel are spliced ​​with features of the same scale as the decoder through skip connections, thereby enhancing the preservation of weak hyperbolic signals after the reinforcement mesh suppression.

[0109] The first stage of the convolutional layer is... The 2D convolutional layer; the first-stage activation function layer is a ReLU activation function layer; the first decoding first transposed convolutional module to the first decoding fourth transposed convolutional module are all... The transposed convolutional layer is used to amplify the resolution; the first encoding first max pooling module to the first encoding fourth max pooling module are all... A max-pooling layer with a stride of 2 is used to halve the size of the input feature map.

[0110] The first encoding first edge module to the first encoding fourth edge module, and the first decoding first edge module to the first decoding fourth edge module all have the same structure;

[0111] The processing of the edge module includes the following steps (structure as follows): Figure 3 (as shown)

[0112] The Sobel operator is used to detect the horizontal and vertical edge information of the input image, and the absolute horizontal and vertical edge information are obtained by taking their absolute values ​​respectively. In specific implementation, two sets of fixed weights are used... The convolutional kernels perform horizontal edge detection and vertical edge detection respectively, with the weight matrix of the horizontal convolutional kernel fixed at [value]. The vertical convolution kernel weight matrix is ​​fixed as follows: Furthermore, all convolutional kernel weights are set to an untrainable state;

[0113] After concatenating the obtained absolute horizontal edge information and absolute vertical edge information with the input image information, ... The 2D convolutional layers are fused together, and then processed sequentially through a normalization layer and a ReLU activation function layer to obtain the output of the edge module;

[0114] The edge module retains the characteristics of classic edge detection by fixing the convolution kernel, while enhancing the edge information in the original features through the feature fusion mechanism, effectively improving the feature representation capability.

[0115] In practical implementation, the constructed second-stage model includes a second-stage encoder, a second-stage decoder, a second-stage averaging module, and a second-stage reconstruction module, with the specific structure as follows: Figure 4 As shown;

[0116] The second-stage encoder includes a second encoding first residual module, a second encoding first max pooling module, a second encoding second residual module, a second encoding second max pooling module, a second encoding third residual module, a second encoding third max pooling module, a second encoding fourth residual module, a second encoding fourth max pooling module, and a second encoding fifth residual module, which are connected in series.

[0117] The second-stage encoder comprises, in sequence, a second decoding first transposed convolutional module, a second decoding first residual module, a second decoding second transposed convolutional module, a second decoding second residual module, a second decoding third transposed convolutional module, a second decoding third residual module, a second decoding fourth transposed convolutional module, and a second decoding fourth residual module. Simultaneously, the output of the second decoding fourth residual module is processed by a fourth attention module and then concatenated with the output of the second decoding fifth residual module as the input to the second decoding first transposed convolutional module; the output of the second decoding third residual module is processed by a third attention module and then concatenated with the output of the second decoding first residual module as the input to the second decoding second transposed convolutional module; the output of the second decoding second residual module is processed by a second attention module and then concatenated with the output of the second decoding second residual module as the input to the second decoding third transposed convolutional module; and the output of the second decoding first residual module is processed by a first attention module and then concatenated with the output of the second decoding third residual module as the input to the second decoding fourth transposed convolutional module.

[0118] The second-stage averaging module includes a second-averaging first convolutional layer, a second-averaging second convolutional layer, a second-averaging third convolutional layer, and a second-averaging fourth convolutional layer. The output of the second encoding first residual module is processed by the second-averaging first convolutional layer to obtain a second-averaged first signal. The output of the second encoding second residual module is processed by the second-averaging second convolutional layer to obtain a second-averaged second signal. The output of the second encoding third residual module is processed by the second-averaging third convolutional layer to obtain a second-averaged third signal. The output of the second encoding fourth residual module is processed by the second-averaging fourth convolutional layer to obtain a second-averaged fourth signal. The average of the obtained second-averaged first signal to second-averaged fourth signal is calculated to obtain the output of the second-stage averaging module. Here, the second-averaged first signal to second-averaged fourth signal are all B-scan two-dimensional images of equal size. The average of these four images is used to obtain the output of the second-stage averaging module. The output of the second-stage averaging module is also a B-scan two-dimensional image of the same size.

[0119] The second-stage reconstruction module includes reconstructing the convolutional layer; the output of the second-stage decoder is processed by the reconstructed convolutional layer and then summed with the output of the second-stage averaging module to obtain the final dilated reconstruction result.

[0120] In the second-stage model, attention-gated skip connections are introduced at the channels to fuse the shallow texture of the encoder with the high-level semantics of the decoder. This enables the network to have both fine boundaries and global context when reconstructing holes, significantly enhancing its ability to capture and locate weak reflection signals.

[0121] Reconstruct the convolutional layer as The 2D convolutional layer; the reconstructed activation function layer is a Sigmoid activation function layer; the second encoding first max pooling module to the second encoding fourth max pooling module are all... The first module is a max pooling module with a stride of 2; the second decoding first transposed convolution module and the second decoding first transposed convolution module are both... The transposed convolutional layer; the second average first convolutional layer to the second average fourth convolutional layer are all... 2D convolutional layers.

[0122] The structures of the first attention module through the fourth attention module are all the same;

[0123] The processing procedure of the attention module is structured as follows: Figure 5 As shown, it includes the following steps:

[0124] Obtain the gated feature map of the input image and jump feature map Gated feature map The feature map output by the second decoding first residual module, the skip feature map The feature map output by the second encoding and third residual modules;

[0125] Will Passing in sequence The image is processed by 2D convolutional layers and normalization layers to obtain the first image component; Passing in sequence The second image component is obtained by processing 2D convolutional layers and normalization layers;

[0126] After summing the first and second image components, the image is then passed sequentially through a ReLU activation function layer. The image is processed by 2D convolutional layers, normalization layers, and sigmoid activation function layers to obtain the third image component;

[0127] Combine the third image component with After multiplying by pixels, the output of the attention module is obtained.

[0128] The attention module effectively identifies and enhances spatial feature regions related to the current decoding stage in skip connections through dual-path feature compression and co-activation mechanisms, significantly improving the accuracy of feature fusion.

[0129] The structures of the first encoding first residual module to the first encoding fifth residual module, the first decoding first residual module to the first decoding fourth residual module, the second encoding first residual module to the second encoding fifth residual module, and the second decoding first residual module to the second decoding fourth residual module are all the same;

[0130] The processing procedure of the residual module is structured as follows: Figure 6As shown, it includes the following steps:

[0131] The input image data is divided into two paths: the first path passes through sequentially... 2D convolutional layers (expanding input channels to the target output dimension while maintaining spatial dimensions), normalization layers, ReLU activation function layers, The first path of features is obtained by processing the 2D convolutional layer (for deep feature extraction while preserving channel dimensions) and the normalization layer; the second path is then processed sequentially through... The second feature path is obtained by processing the 2D convolutional layer (with channel number adjustment) and normalization layer;

[0132] After the first feature and the second feature are added together, they are processed through a ReLU activation function layer to obtain the output of the residual module;

[0133] The residual module maintains spatial dimension invariance through equal-width padding, achieves adaptive adjustment of channel dimension through conditional shortcut connections, and preserves original feature information by utilizing the identity mapping mechanism, effectively solving the gradient vanishing problem in deep neural network training.

[0134] S5. Using the reconstructed dataset obtained in step S3, train the initial reconstruction model of the void under the shallow steel mesh constructed in step S4 to obtain the reconstruction model of the void under the shallow steel mesh.

[0135] During training, the pre-processed B-scan image of underground cavities with shallow steel mesh covering them was used as the input of the first-stage model, and the pre-processed B-scan image of underground cavities without shallow steel mesh covering them was used as the label for the input of the first-stage model. The first-stage model was then trained.

[0136] Then, the output of the first-stage model is used as the input of the second-stage model, and the preprocessed black-and-white mask image of the hole target is used as the label of the input of the second-stage model to train the second-stage model.

[0137] During training, the loss function of the model in the first stage is a weighted sum of the L1 loss function and the SSIM loss function;

[0138] The loss function of the second-stage model is a weighted sum of the BCE loss function, Dice loss function, Boundary loss function, and Focal loss function;

[0139] During training, the weighted sum of the loss functions of the first-stage model and the second-stage model is used as the total loss function.

[0140] During training, the Adam optimizer is used, and an adaptive learning rate mechanism and an early stopping strategy are employed to improve the stability and efficiency of the training process.

[0141] S6. Using the reconstruction model of the void under the shallow steel mesh obtained in step S5, the actual reconstruction of the void under the shallow steel mesh is carried out.

[0142] This invention can automatically output high-precision underground cavity distribution maps, with fast processing capabilities and good robustness, and is suitable for underground cavity detection and safety assessment tasks in various layered geological or urban infrastructure scenarios.

[0143] The effects of the method of the present invention will be illustrated below with reference to an embodiment:

[0144] Figure 7 This is a schematic diagram showing the reconstruction result of a rectangular cavity with a rebar spacing of 70cm. Figure 7 The input is fed into the two-stage network trained by the scheme of this invention.

[0145] The diagram illustrating the comparison between the network output in the first stage and the actual data is shown below. Figure 8 As shown; via Figure 8 As can be seen, the first-stage network effectively suppresses the clutter from the reinforcing mesh while preserving the echoes from the void targets, resulting in clear target boundary contours and distinguishable weak echo structures. This demonstrates that the proposed first-stage network can weaken the interference from the strongly scattering reinforcing mesh while preserving as much effective information as possible about the voids, providing reliable input data for the second-stage reconstruction of target location and size.

[0146] A diagram illustrating the comparison between the network output results and the actual data in the second stage is shown below. Figure 9 As shown; via Figure 9 As can be seen, the second-stage network can accurately identify the location of the hole target and recover its geometric shape and size features well. For rectangular holes, the reconstruction result is highly consistent with the true mask boundary, and the target area is complete with clear boundaries.

[0147] like Figure 10The diagram shows the functional modules of the system of this invention: The system for reconstructing voids under the shading of shallow steel mesh disclosed in this invention includes a scene construction module, a data acquisition module, a data processing module, a model construction module, a model training module, and a void reconstruction module; the scene construction module, data acquisition module, data processing module, model construction module, model training module, and void reconstruction module are connected in series; the scene construction module is used to construct a layered medium scene containing shallow steel mesh and voids, and upload the data information to the data acquisition module; the data acquisition module is used to acquire image data information with different shading states based on the constructed scene according to the received data information, and upload the data information to the data processing module; the data processing module is used to preprocess the acquired image data information according to the received data information to construct a reconstruction dataset, and upload the data information to the model construction module; the model construction module is used to construct the shallow steel mesh according to the received data information. An initial model for reconstructing cavities under shallow rebar mesh shading is generated, and the data is uploaded to the model training module. This initial model comprises a first-stage model and a second-stage model. The first-stage model is constructed based on a U-shaped network, Sobel operator, residual structure, and skip connection scheme to suppress the echo of the rebar mesh. The second-stage model is constructed based on the U-shaped network, residual structure, attention mechanism, and skip connection scheme to reconstruct cavities based on the output of the first-stage model. The model training module trains the initial model using the received data and the obtained reconstruction dataset to obtain a reconstruction model for cavities under shallow rebar mesh shading, and uploads the data to the cavity reconstruction module. The cavity reconstruction module then uses the obtained reconstruction model to reconstruct cavities under shallow rebar mesh shading based on the received data.

Claims

1. A reconstruction method of a cavity under a shallow reinforcement screen, characterized in that Includes the following steps: S1. Construct a layered medium scenario containing shallow steel mesh and voids; S2. Based on the scene constructed in step S1, obtain image data information with different occlusion states; S3. Preprocess the image data information obtained in step S2 to construct a reconstructed dataset; S4. Construct an initial reconstruction model of the voids concealed by shallow steel mesh; The initial model for reconstructing cavities under shallow steel mesh includes a first-stage model and a second-stage model; A first-stage model is constructed based on a U-shaped network, the Sobel operator, residual structures, and a skip connection scheme; the first-stage model is used to suppress the echo of the steel mesh. The second-stage model is constructed based on U-shaped network, residual structure, attention mechanism, averaging mechanism, multi-scale edge fusion and skip connection scheme; The second-stage model is used to reconstruct the holes based on the output of the first-stage model; The first-stage model is constructed, including the first-stage encoder, the first-stage decoder, and the first-stage filtering module; The first-stage encoder includes a first encoding first residual module, a first encoding first edge module, a first encoding first max pooling module, a first encoding second residual module, a first encoding second edge module, a first encoding second max pooling module, a first encoding third residual module, a first encoding third edge module, a first encoding third max pooling module, a first encoding fourth residual module, a first encoding fourth edge module, a first encoding fourth max pooling module, and a first encoding fifth residual module, which are connected in series. The first-stage decoder comprises, in sequence, a first decoding first transposed convolutional module, a first decoding first residual module, a first decoding first edge module, a first decoding second transposed convolutional module, a first decoding second residual module, a first decoding second edge module, a first decoding third transposed convolutional module, a first decoding third residual module, a first decoding third edge module, a first decoding fourth transposed convolutional module, a first decoding fourth residual module, and a first decoding fourth edge module. Simultaneously, the output of the first encoding fourth edge module and the output of the first decoding first residual module are concatenated and used as the input of the first decoding first edge module; the output of the first encoding third edge module and the output of the first decoding second residual module are concatenated and used as the input of the first decoding second edge module; the output of the first encoding second edge module and the output of the first decoding third residual module are concatenated and used as the input of the first decoding third edge module; and the output of the first encoding first edge module and the output of the first decoding fourth residual module are concatenated and used as the input of the first decoding fourth edge module. The first-stage filtering module includes a first-stage convolutional layer and a first-stage activation function layer connected in series, and the output of the first-stage decoder is the input of the first-stage filtering module; The first-stage model is used to suppress the echo of the rebar mesh and output a GPR image without the echo of the rebar mesh. The first encoding first edge module to the first encoding fourth edge module, and the first decoding first edge module to the first decoding fourth edge module all have the same structure; The processing of the edge module includes the following steps: The Sobel operator is used to detect the horizontal and vertical edge information of the input image, and the absolute values ​​are taken to obtain the absolute horizontal and absolute vertical edge information respectively. After splicing the obtained absolute horizontal edge information, absolute vertical edge information and input image information, the spliced information is fused through a 2d convolution layer, and then sequentially processed through a normalization layer and a ReLU activation function layer to obtain an output of the edge module. the output of the edge module. The constructed second-stage model includes a second-stage encoder, a second-stage decoder, a second-stage averaging module, and a second-stage reconstruction module; The second-stage encoder includes a second encoding first residual module, a second encoding first max pooling module, a second encoding second residual module, a second encoding second max pooling module, a second encoding third residual module, a second encoding third max pooling module, a second encoding fourth residual module, a second encoding fourth max pooling module, and a second encoding fifth residual module, which are connected in series. The second-stage encoder includes a second decoding first transposed convolutional module, a second decoding first residual module, a second decoding second transposed convolutional module, a second decoding second residual module, a second decoding third transposed convolutional module, a second decoding third residual module, a second decoding fourth transposed convolutional module, and a second decoding fourth residual module, all connected in series. Simultaneously, the output of the second decoding fourth residual module is processed by a fourth attention module and then concatenated with the output of the second decoding fifth residual module, serving as the input to the second decoding first transposed convolutional module; the output of the second decoding third residual module is processed by a third attention module and then concatenated with the output of the second decoding first residual module, serving as the input to the second decoding second transposed convolutional module. The output of the second encoding second residual module is processed by the second attention module and then concatenated with the output of the second decoding second residual module as the input of the second decoding third transposed convolution module; the output of the second encoding first residual module is processed by the first attention module and then concatenated with the output of the second decoding third residual module as the input of the second decoding fourth transposed convolution module. The second-stage averaging module includes a second-averaging first convolutional layer, a second-averaging second convolutional layer, a second-averaging third convolutional layer, and a second-averaging fourth convolutional layer; the output of the second encoding first residual module is processed by the second-averaging first convolutional layer to obtain the second-averaged first signal; The output of the second encoding second residual module is processed by the second averaging second convolutional layer to obtain the second average second signal; The output of the second encoding third residual module is processed by the second averaging third convolutional layer to obtain the second average third signal; the output of the second encoding fourth residual module is processed by the second averaging fourth convolutional layer to obtain the second average fourth signal; the average of the obtained second average first signal to second average fourth signal is calculated to obtain the output of the second stage averaging module; The second-stage reconstruction module includes a reconstruction convolutional layer; the output of the second-stage decoder is processed by the reconstruction convolutional layer and then summed with the output of the second-stage averaging module to obtain the final dilated reconstruction result. S5. Using the reconstructed dataset obtained in step S3, train the initial reconstruction model of the void under the shallow steel mesh constructed in step S4 to obtain the reconstruction model of the void under the shallow steel mesh. S6. Using the reconstruction model of the void under the shallow steel mesh obtained in step S5, the actual reconstruction of the void under the shallow steel mesh is carried out.

2. The reconstruction method of shallow reinforcement screen-shielded under voids according to claim 1, characterized in that Step S1 specifically includes the following steps: The constructed layered medium scenario consists of four layers, from top to bottom: an air layer, a surface layer, a base layer containing steel mesh echoes, and a soil base layer; The surface layer is made of asphalt concrete, and is divided into three layers according to the different aggregate sizes, shapes, and random distribution of aggregate positions. The base material is set to reinforced concrete, and the spacing and number of steel bars are set. The base is divided into two layers according to the different aggregate sizes and shapes. The subgrade is set to a randomly mixed soil, and the soil surface is set to an uneven, undulating surface.

3. The method of claim 2, wherein Step S2 specifically includes the following steps: Based on the scenario constructed in step S1, several forward model files containing hollow targets are generated. The types of cavities include rectangular cavities without water, rectangular cavities with water, irregularly shaped cavities without water, and irregularly shaped cavities with water; the cavities are set in the soil base layer; the size and depth of the cavities are randomly generated. The obtained forward model files include B-scan images of underground cavities with shallow steel mesh covering them, B-scan images of underground cavities without shallow steel mesh covering them, and black-and-white mask images of the cavity targets. Step S3 specifically includes the following steps: Set the maximum reference threshold for the B-scan image; Signal suppression processing: In the image data information obtained in step S2, all data higher than the maximum reference threshold are replaced with the maximum reference threshold to complete the suppression of strong signals; Signal enhancement processing: Construct an exponential gain sequence that increases with depth, and multiply the data after signal suppression by the constructed exponential gain sequence to complete the signal enhancement processing; Image scaling: The image data after signal enhancement is uniformly scaled to a set size; Complete the preprocessing of the image data information obtained in step S2.

4. The reconstruction method of shallow reinforcement screen-shielded under voids according to claim 3, characterized in that The structures of the first attention module through the fourth attention module are all the same; The attention module's processing steps include the following: Obtain the gated feature map of the input image and jump feature map Gated feature map The feature map output by the second decoding first residual module, the skip feature map The feature map output by the second encoding and third residual modules; Will Passing in sequence The image is processed by 2D convolutional layers and normalization layers to obtain the first image component; Passing in sequence The second image component is obtained by processing 2D convolutional layers and normalization layers; After summing the first and second image components, the image is then passed sequentially through a ReLU activation function layer. The image is processed by 2D convolutional layers, normalization layers, and sigmoid activation function layers to obtain the third image component; The third image component is multiplied with After pixel-wise multiplication, the output of the attention module is obtained.

5. The method of claim 4, wherein The structures of the first encoding first residual module to the first encoding fifth residual module, the first decoding first residual module to the first decoding fourth residual module, the second encoding first residual module to the second encoding fifth residual module, and the second decoding first residual module to the second decoding fourth residual module are all the same; The residual module processing procedure includes the following steps: The input image data is divided into two paths: the first path passes through sequentially... 2D convolutional layers, normalization layers, ReLU activation function layers, The first path features are obtained by processing the 2D convolutional layer and normalization layer; The second path sequentially passes through a 2d convolution layer and a normalization layer to obtain a second path feature; The first and second features are added together, and then processed through a ReLU activation function layer to obtain the output of the residual module.

6. The method of claim 5, wherein The first stage of the convolutional layer is... The 2D convolutional layer; the first-stage activation function layer is a ReLU activation function layer; the first decoding first transposed convolutional module to the first decoding fourth transposed convolutional module are all... The transposed convolutional layer is used to amplify the resolution; the first encoding first max pooling module to the first encoding fourth max pooling module are all... A max-pooling layer with a stride of 2 is used to halve the size of the input feature map; Reconstruct the convolutional layer as The 2D convolutional layer; the reconstructed activation function layer is a Sigmoid activation function layer; the second encoding first max pooling module to the second encoding fourth max pooling module are all... The first module is a max pooling module with a stride of 2; the second decoding first transposed convolution module and the second decoding first transposed convolution module are both... The transposed convolutional layer; the second average first convolutional layer to the second average fourth convolutional layer are all... 2D convolutional layers; The training described in step S5 specifically includes the following steps: During training, the pre-processed B-scan image of underground cavities with shallow steel mesh covering them was used as the input of the first-stage model, and the pre-processed B-scan image of underground cavities without shallow steel mesh covering them was used as the label for the input of the first-stage model. The first-stage model was then trained. Then, the output of the first-stage model is used as the input of the second-stage model, and the preprocessed black-and-white mask image of the hole target is used as the label of the input of the second-stage model to train the second-stage model. The loss function of the first-stage model is a weighted sum of the L1 loss function and the SSIM loss function; The loss function of the second-stage model is a weighted sum of the BCE loss function, Dice loss function, Boundary loss function, and Focal loss function; During training, the weighted sum of the loss functions of the first-stage model and the second-stage model is used as the total loss function. During training, the Adam optimizer is used, and an adaptive learning rate mechanism and an early stopping strategy are employed to improve the stability and efficiency of the training process.

7. A system for implementing the reconstruction method of the cavity under the masking of the shallow reinforcement mesh according to one of claims 1 to 6, characterized in that It includes a scene construction module, a data acquisition module, a data processing module, a model construction module, a model training module, and a void reconstruction module; these modules are connected in series. The scene construction module is used to construct a layered medium scene containing shallow steel mesh and voids, and uploads the data information to the data acquisition module. The data acquisition module is used to acquire image data information with different occlusion states based on the constructed scene and the received data information, and uploads the data information to the data processing module. The data processing module is used to preprocess the acquired image data based on the received data information to construct a reconstructed dataset and upload the data information to the model building module; The model building module is used to construct an initial reconstruction model of voids under shallow rebar mesh shading based on the received data information, and upload the data information to the model training module. The initial reconstruction model of voids under shallow rebar mesh shading includes a first-stage model and a second-stage model. The first-stage model is constructed based on a U-shaped network, Sobel operator, residual structure, and skip connection scheme; the first-stage model is used to suppress the echo of the rebar mesh. The second-stage model is constructed based on a U-shaped network, residual structure, attention mechanism, averaging mechanism, multi-scale edge fusion, and skip connection scheme; the second-stage model is used to reconstruct voids based on the output of the first-stage model. The model training module is used to train the initial reconstruction model of the voids under the shallow steel mesh based on the received data information and the obtained reconstruction dataset, so as to obtain the reconstruction model of the voids under the shallow steel mesh and upload the data information to the void reconstruction module; the void reconstruction module is used to reconstruct the actual voids under the shallow steel mesh based on the received data information and the obtained reconstruction model of the voids under the shallow steel mesh.

Images