Deep learning-based shallow reinforcement mesh echo suppression method and system

Abstract

The application discloses a kind of shallow reinforcement mesh echo suppression method and system based on deep learning, including shallow reinforcement mesh shielding scene is detected using ground penetrating radar, corresponding data information is acquired and training data set is obtained by processing;Based on U-net network structure, combined with convolution unit, pooling unit, attention mechanism, interpolation operation and jump connection mechanism, shallow reinforcement mesh echo suppression initial model is constructed and shallow reinforcement mesh echo suppression model is trained;The shallow reinforcement mesh echo suppression model obtained is used to suppress the echo of shallow reinforcement mesh in the actual ground penetrating radar detection process.The present application not only realizes the echo suppression of shallow reinforcement mesh based on deep learning, but also has higher reliability and better accuracy.

Description

Technical Field

[0001] This invention belongs to the field of ground-penetrating radar data processing, specifically relating to a method and system for suppressing echoes from shallow steel mesh based on deep learning. Background Technology

[0002] Ground penetrating radar (GPR), a non-destructive testing technology based on the principle of high-frequency electromagnetic wave scattering, can effectively reveal the distribution patterns and material properties of underground media, playing a crucial role in engineering geology, tunnel inspection, and underground pipeline investigation. GPR scans along a survey line on the Earth's surface; at each measurement point, broadband electromagnetic waves emitted by the transmitting antenna propagate underground. When these waves encounter heterogeneous materials with different dielectric constants, they generate scattered echoes, which are received by the receiving antenna. By analyzing the signals captured by the receiving antenna, the geometric location and physical properties of underground targets can be deduced. The combined echo signals from each measurement point form a GPR B-scan echo dataset.

[0003] However, in complex engineering inspections, shallow, high-intensity reflective materials often obscure weak signals from deeper layers. Taking roadbed inspection as an example, the surface pavement is typically covered with a dense steel mesh, while the deeper parts of the roadbed may harbor cavities. Because steel reinforcement has a strong shielding and scattering effect on electromagnetic waves, it not only significantly attenuates the downward detection signal but also causes secondary obstruction to the upward signal reflected back from deep cavities. This bidirectional attenuation results in extremely weak received cavity echoes, which are submerged in strong clutter generated by the steel reinforcement, greatly increasing the difficulty of defect identification. Therefore, suppressing echoes from shallow steel mesh has become a key research focus.

[0004] Currently, researchers mainly use deep convolutional neural networks (CNNs) and generative adversarial networks (GANs) to suppress echoes in shallow steel mesh. However, the CNN approach often suffers from detail loss and image blurring. The GAN approach, on the other hand, heavily relies on strictly paired "noisy-clean" datasets, which are extremely difficult to obtain in practical engineering applications; therefore, this approach also suffers from poor reliability and accuracy. Summary of the Invention

[0005] One of the objectives of this invention is to provide a deep learning-based method for suppressing echoes from shallow steel meshes that is highly reliable and accurate.

[0006] The second objective of this invention is to provide a system for implementing the aforementioned deep learning-based shallow steel mesh echo suppression method.

[0007] The deep learning-based shallow steel mesh echo suppression method provided by this invention includes the following steps:

[0008] S1. Ground-penetrating radar is used to detect shallow steel mesh-covered scenes and acquire corresponding data information;

[0009] S2. Based on the data obtained in step S1, perform image processing on the accumulated energy data to obtain the training dataset;

[0010] S3. Based on the U-net network structure, and combining convolutional units, pooling units, attention mechanisms, interpolation operations, and skip connection mechanisms, an initial model for shallow steel mesh echo suppression is constructed.

[0011] The constructed initial model for shallow steel mesh echo suppression includes an encoding module, a feature enhancement module, a decoding module, and an output module;

[0012] An encoding module is constructed based on convolutional and pooling units; the encoding module is used to extract features and downsample the input image data.

[0013] A feature enhancement module is constructed based on an attention mechanism; the feature enhancement module is a skip connection between the encoding module and the decoding module, used to calibrate various features in the encoding module;

[0014] A decoding module is constructed based on convolutional units and interpolation operations; the decoding module is used to fuse the features output by the encoding module and the features output by the feature enhancement module.

[0015] An output module is constructed based on convolutional units; the output module is used to map the features output by the decoding module to obtain the final ground-penetrating radar detection image with echo suppression achieved for shallow steel mesh.

[0016] S4. Using the training dataset obtained in step S2, train the initial shallow steel mesh echo suppression model constructed in step S3 to obtain the shallow steel mesh echo suppression model.

[0017] S5. Using the shallow steel mesh echo suppression model obtained in step S4, the echo suppression of the shallow steel mesh is performed in the actual ground penetrating radar detection process.

[0018] Step S1 specifically includes the following steps:

[0019] Construct a shallow steel mesh shielding scenario; the shallow steel mesh shielding scenario is a scenario containing only steel mesh.

[0020] Ground-penetrating radar was used to detect shallow steel mesh-covered scenes, obtaining several sets of B-scan echo data containing only the steel mesh; each set of data corresponds to a set of steel bar numbers and spacing between steel bars; the steel bar burial depth is the same in all data.

[0021] Step S2 specifically includes the following steps:

[0022] The several sets of B-scan echo data containing only steel mesh obtained in step S1 are converted into several B-scan images of a set size containing only steel mesh echo data.

[0023] Image enhancement operations are performed on several B-scan images containing only rebar mesh echo data to obtain several enhanced B-scan images containing only rebar mesh echo data; the image enhancement operations include image horizontal flipping and image scaling;

[0024] Using simulation software, irregular cavities and faults of randomly sized and located defective targets were constructed, and ground-penetrating radar detection simulations were performed to obtain several sets of B-scan echo data containing the targets.

[0025] The obtained B-scan echo data containing the target are converted into several B-scan images of a set size containing the target echo.

[0026] The background removal method was used to remove the direct wave from several B-scan images containing the target echo, resulting in several B-scan images containing only the target echo.

[0027] Based on the cumulative energy data of the images, several enhanced B-scan images containing only steel mesh echo data and several B-scan images containing only target echo data are fused to obtain the training dataset.

[0028] The fusion process specifically includes the following steps:

[0029] The maximum absolute amplitude of the enhanced B-scan image containing only rebar mesh echo data is calculated and multiplied by a random coefficient within a set range to obtain the simulated target amplitude.

[0030] Based on the obtained simulated target amplitude, the B-scan image containing only the target echo is normalized to obtain the target image;

[0031] The absolute value of the enhanced B-scan image containing only steel mesh echo data is taken to obtain the energy map; for the energy map, the cumulative sum is calculated along the depth direction of the image to obtain the cumulative energy distribution map;

[0032] The obtained cumulative energy distribution map is normalized, and the transmission attenuation mask is calculated based on the exponential attenuation function.

[0033] The obtained target image is multiplied pixel by pixel with the obtained transmission attenuation mask, and the result is superimposed on the enhanced B-scan image containing only steel mesh echo data;

[0034] Complete the image fusion.

[0035] The constructed encoding module specifically includes the following:

[0036] The encoding module includes an input encoding convolutional layer, a first encoding max pooling layer, a first encoding convolutional layer, a second encoding max pooling layer, a second encoding convolutional layer, a third encoding max pooling layer, a third encoding convolutional layer, a fourth encoding max pooling layer, a fourth encoding convolutional layer, and an encoding multi-scale convolutional block attention layer, all connected in sequence. The output of the encoding multi-scale convolutional block attention layer is the output of the encoding module.

[0037] Specifically, the output of the input coding convolutional layer is the first skip connection input, the output of the first coding convolutional layer is the second skip connection input, the output of the second coding convolutional layer is the third skip connection input, and the output of the third coding convolutional layer is the fourth skip connection input; all three skip connection inputs are input to the feature enhancement module.

[0038] The input encoding convolutional layer, the first encoding convolutional layer, the second encoding convolutional layer, the third encoding convolutional layer, and the fourth encoding convolutional layer all have the same structure; the encoding convolutional layer consists of two concatenated convolutional sub-layers; the convolutional sub-layers consist of sequentially concatenated... Convolutional layers, batch normalization layers, and ReLU activation function layers. The convolution kernel of the convolutional layer is The step size is 1, and the fill size is 1.

[0039] The first, second, third, and fourth encoded max pooling layers all have the same structure; the encoded max pooling layer has a pooling kernel size of [missing value]. Maximum pooling layer;

[0040] Based on channel attention and spatial attention mechanisms, we construct an attention layer that encodes multi-scale convolutional blocks.

[0041] The constructed feature enhancement module specifically includes the following:

[0042] The feature enhancement module includes a first skip connection layer, a second skip connection layer, a third skip connection layer, and a fourth skip connection layer;

[0043] The input to the first jump connection layer is the first jump connection input, and the output of the first jump connection layer is the first jump connection output; the input to the second jump connection layer is the second jump connection input, and the output of the second jump connection layer is the second jump connection output; the input to the third jump connection layer is the third jump connection input, and the output of the third jump connection layer is the third jump connection output; the input to the fourth jump connection layer is the fourth jump connection input, and the output of the fourth jump connection layer is the fourth jump connection output.

[0044] The outputs from the first jump connection to the fourth jump connection are all output to the decoding module;

[0045] The first, second, third, and fourth skip connection layers have the same structure; the skip connection layer includes a feature calibration residual dense sub-layer and an enhanced multi-scale convolutional block attention sub-layer connected in sequence.

[0046] A dense sublayer for feature calibration residuals is constructed based on convolutional and pooling units.

[0047] Based on channel attention and spatial attention mechanisms, an enhanced multi-scale convolutional block attention sublayer is constructed.

[0048] The constructed decoding module includes the following:

[0049] The decoding module includes a first decoding upsampling layer, a first decoding convolutional layer, a second decoding upsampling layer, a second decoding convolutional layer, a third decoding upsampling layer, a third decoding convolutional layer, a fourth decoding upsampling layer, and a fourth decoding convolutional layer;

[0050] The output of the encoding module is concatenated with the output of the fourth skip connection along the channel dimension, and then used as the input of the first decoding upsampling layer. The output of the first decoding upsampling layer is used as the input of the first decoding convolutional layer. The output of the first decoding convolutional layer is concatenated with the output of the third skip connection along the channel dimension, and then used as the input of the second decoding upsampling layer. The output of the second decoding upsampling layer is used as the input of the second decoding convolutional layer. The output of the second decoding convolutional layer is concatenated with the output of the second skip connection along the channel dimension, and then used as the input of the third decoding upsampling layer. The output of the third decoding upsampling layer is used as the input of the third decoding convolutional layer. The output of the third decoding convolutional layer is concatenated with the output of the first skip connection along the channel dimension, and then used as the input of the fourth decoding upsampling layer. The output of the fourth decoding upsampling layer is used as the input of the fourth decoding convolutional layer. The input of the fourth decoding convolutional layer is the output of the decoding module.

[0051] The processing procedures for the first, second, third, and fourth decoding upsampling layers are all the same; the decoding upsampling layer uses bilinear interpolation for upsampling.

[0052] The first, second, third, and fourth decoding convolutional layers all have the same structure; each decoding convolutional layer consists of two concatenated convolutional sub-layers; each convolutional sub-layer consists of sequentially concatenated... Convolutional layers, batch normalization layers, and ReLU activation function layers. The convolution kernel of the convolutional layer is The step size is 1, and the padding size is 1.

[0053] The output module constructed specifically includes the following:

[0054] The output module includes a first output convolutional layer and a second output convolutional layer connected in series.

[0055] The first output convolutional layer consists of two concatenated convolutional sub-layers; the convolutional sub-layers consist of sequentially concatenated... Convolutional layers, batch normalization layers, and ReLU activation function layers. The convolution kernel of the convolutional layer is The step size is 1, and the fill size is 1.

[0056] The second output convolutional layer has a kernel size of [size missing]. The convolutional layer.

[0057] The structure of the encoding multi-scale convolutional block attention layer and the enhanced multi-scale convolutional block attention sub-layer is the same; the multi-scale convolutional block attention layer includes the following:

[0058] The input features are processed by the channel attention sub-layer to obtain the channel attention map; the channel attention map is multiplied element-wise with the input features to obtain the intermediate feature map; the intermediate feature map is processed by the multi-scale spatial attention sub-layer to obtain the multi-scale spatial attention map; the multi-scale spatial attention map is multiplied element-wise with the intermediate feature map to obtain the output of the multi-scale convolutional block attention layer.

[0059] The processing procedure of the channel attention sublayer includes the following:

[0060] The input features of the channel attention sublayer are divided into two paths;

[0061] The first feature path passes sequentially through an average pooling layer and a convolutional kernel size of [missing information]. The size of the convolutional layer, ReLU activation function layer, and convolutional kernel is The convolutional layer processing yields the average pooling result features;

[0062] The second feature path passes sequentially through a max pooling layer and a convolutional kernel size of [missing information]. The size of the convolutional layer, ReLU activation function layer, and convolutional kernel is The convolutional layer processing yields the max pooling result features;

[0063] The obtained average pooling result features and max pooling result features are added element-wise, and then processed through a Sigmoid activation function layer to obtain the channel attention map;

[0064] The processing procedure of the multi-scale spatial attention sublayer includes the following:

[0065] The average channel feature map is obtained by averaging the input features of the multi-scale spatial attention sub-layer in the channel dimension.

[0066] The maximum value of the input features of the multi-scale spatial attention sublayer is calculated in the channel dimension to obtain the maximum channel feature map.

[0067] The average channel feature map and the maximum channel feature map are concatenated along the channel dimension to obtain the compressed spatial feature map;

[0068] The compressed spatial feature map is processed simultaneously through three convolutional branches; the first convolutional branch includes a standard convolutional layer and a batch normalization layer connected in sequence, with the kernel size of the standard convolutional layer being [missing value]. The padding is 1, and the dilation rate is 1; the second convolutional branch consists of a series of dilated convolutional layers and batch normalized layers, with the kernel size of the dilated convolutional layers being 1. The padding is 2, and the dilation rate is 2; the third convolutional branch consists of a series of dilated convolutional layers and batch normalized layers, with the kernel size of the dilated convolutional layers being [missing value]. The filler is 4, and the expansion rate is 4.

[0069] The outputs of the three convolutional branches are concatenated along the channel dimension, and then sequentially passed through convolutional kernels of size [size missing]. The convolutional layers and sigmoid activation function layers are used to process the data to obtain a multi-scale spatial attention map.

[0070] The process of processing feature-calibrated residual dense sublayers includes the following:

[0071] The input features of the feature calibration residual dense sub-layer are processed sequentially through several dense calibration units, and then through a convolutional kernel of size [size missing]. The convolutional layers are processed to obtain feature fusion sub-features;

[0072] The feature fusion sub-features are added element-wise to the input features of the feature calibration residual dense sub-layer to obtain the output features of the feature calibration residual dense sub-layer.

[0073] The processing steps of the dense calibration unit include the following:

[0074] The input features of the dense calibration unit are processed sequentially through convolutional layers and ReLU activation function layers to obtain intermediate calibration sub-features; wherein, the kernel size of the convolutional layer is [missing information]. The step size is 1, and the fill size is 1.

[0075] After the intermediate calibrator features are processed by the global average pooling layer, they are then processed sequentially by the first fully connected layer, the ReLU activation function layer, the second fully connected layer, and the Sigmoid activation function layer to obtain the channel weight vector.

[0076] The calibration feature map is obtained by multiplying the channel weight vector with the intermediate calibration sub-feature channel by channel.

[0077] The calibration feature map is concatenated with the input features of the dense calibration unit along the channel dimension to obtain the output of the dense calibration unit.

[0078] The training described in step S4 specifically includes the following steps:

[0079] The following loss function is used for training:

[0080]

[0081]

[0082]

[0083]

[0084] In the formula The value of the loss function; The height of the input image; The width of the input image; For real images The middle position is The actual pixel value at that location; For predicting images The middle position is Pixel value at; These are the brightness comparison metrics between the real and predicted images at M scales; The first index is set to adjust the relative importance of different components across M scales; The total number of scales representing the quality of the real and predicted images; This is a comparison metric between the real image and the predicted image at the k-th scale. The second index is set to adjust the relative importance of different components at the k-th scale; This is a measure of the structural comparison between the real image and the predicted image at the k-th scale. The third index is set to adjust the relative importance of different components at the k-th scale; The first parameter is set; This is the second parameter that is set; This is the third parameter that is set; The average pixel value of the real image; To predict the mean pixel value of the image; This represents the variance of the pixel values ​​in the actual image. To predict the variance of pixel values ​​in the image; This is the covariance between the pixel values ​​of the real image and the pixel values ​​of the predicted image.

[0085] This invention also provides a system for implementing the deep learning-based shallow rebar mesh echo suppression method, comprising a data acquisition module, a data processing module, a model building module, a model training module, and an echo suppression module; the data acquisition module, data processing module, model building module, model training module, and echo suppression module are connected in series; the data acquisition module is used to detect the shallow rebar mesh occlusion scene using ground penetrating radar and acquire the corresponding data information, and upload the data information to the data processing module; the data processing module is used to perform image processing based on the received data information and the acquired data information, based on the cumulative energy data, to obtain a training dataset, and upload the data information to the model building module; the model building module is used to construct an initial model for shallow rebar mesh echo suppression based on the received data information, using the U-net network structure, combined with convolutional units, pooling units, attention mechanisms, interpolation operations, and skip connection mechanisms, and upload the data information to the model training module; the constructed initial model for shallow rebar mesh echo suppression includes an encoding module, a feature enhancement module, a decoding module, and... The system comprises the following modules: an output module; an encoding module based on convolutional and pooling units; an attention-based feature enhancement module; a skip connection between the encoding and decoding modules to calibrate various features in the encoding module; a decoding module based on convolutional and interpolation operations; a fusion module that integrates the features output by the encoding and feature enhancement modules; an output module based on convolutional units; a mapping module that maps the features output by the decoding module to obtain the final ground-penetrating radar image with shallow rebar mesh echo suppression; a model training module that trains the initial shallow rebar mesh echo suppression model using the received training dataset, and uploads the data to the echo suppression module; and an echo suppression module that uses the received data and the obtained shallow rebar mesh echo suppression model to perform echo suppression of the shallow rebar mesh during actual ground-penetrating radar detection.

[0086] The present invention provides a method and system for suppressing echoes from shallow steel mesh based on deep learning. By conducting field measurements, simulations, and image processing on shallow steel mesh occlusion scenarios to construct a training dataset, and then constructing and training a shallow steel mesh echo suppression model through a deep learning scheme, the present invention not only achieves echo suppression of shallow steel mesh based on deep learning, but also has higher reliability and better accuracy. Attached Figure Description

[0087] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0088] Figure 2 This is a schematic diagram of the shallow steel mesh echo suppression model of the method of the present invention.

[0089] Figure 3 This is a schematic diagram of the structure of the multi-scale convolutional block attention layer in the method of the present invention; wherein, Figure 3 (a) is a schematic diagram of the overall structure of the attention layer of the multi-scale convolutional block. Figure 3 (b) is a schematic diagram of the channel attention sublayer. Figure 3 (c) is a schematic diagram of the structure of the multi-scale spatial attention sublayer.

[0090] Figure 4 This is a schematic diagram of the structure of the characteristic calibration residual dense sublayer of the method of the present invention, wherein, Figure 4 (a) is a schematic diagram of the overall structure of the feature calibration residual dense sublayer. Figure 4 (b) is a schematic diagram of the structure of the dense calibration unit.

[0091] Figure 5 This is a schematic diagram of B-scan data and processing results for a sparse steel mesh scenario according to an embodiment of the method of the present invention; wherein, Figure 5 (a) is a composite image of a sparse steel mesh scene. Figure 5 (b) is a B-scan image of a weak target in a sparse steel mesh scene. Figure 5 (c) is a schematic diagram of the processing results of the method of the present invention in a sparse steel mesh scenario. Figure 5 (d) is a schematic diagram of the processing results of the comparison scheme in the sparse steel mesh scenario.

[0092] Figure 6 This is a schematic diagram of B-scan data and processing results for a relatively dense steel mesh scenario according to an embodiment of the method of the present invention; wherein, Figure 6 (a) is a composite image of a scene with relatively dense steel mesh. Figure 6 (b) is a B-scan image of a weak target in a scene with a relatively dense steel mesh. Figure 6 (c) is a schematic diagram of the processing results of the method of the present invention in a scenario with relatively dense steel mesh. Figure 6(d) is a schematic diagram of the processing results of the comparison scheme in a scenario with a relatively dense steel mesh.

[0093] Figure 7 This is a schematic diagram of B-scan data and processing results for a dense steel mesh scenario according to an embodiment of the method of the present invention; wherein, Figure 7 (a) is a composite image of a scene with dense steel mesh. Figure 7 (b) is a B-scan image of a weak target in a scene with dense steel mesh. Figure 7 (c) is a schematic diagram of the processing results of the method of the present invention in a scenario with dense steel mesh. Figure 7 (d) is a schematic diagram of the processing results of the comparison scheme for the dense steel mesh scenario.

[0094] Figure 8 This is a schematic diagram of simulated B-scan data and processing results containing a steel mesh and the target below, according to an embodiment of the method of the present invention; wherein, Figure 8 (a) is a schematic diagram of the simulated steel mesh target image. Figure 8 (b) is a schematic diagram of the processing results of the present invention on the simulation data. Figure 8 (c) is a schematic diagram of the processing results of the simulation data comparison scheme.

[0095] Figure 9 This is a schematic diagram of measured B-scan data and processing results of a method embodiment of the present invention, including a steel mesh and the target below it; wherein, Figure 9 (a) is a schematic diagram of the measured target image of the steel mesh strip. Figure 9 (b) is a schematic diagram of the processing results of the present invention for measured data. Figure 9 (c) is a schematic diagram of the processing results of the comparison scheme for measured data.

[0096] Figure 10 This is a schematic diagram of the functional modules of the system of the present invention. Detailed Implementation

[0097] like Figure 1 The diagram shown is a flowchart of the method of the present invention: The shallow steel mesh echo suppression method based on deep learning disclosed in this invention includes the following steps:

[0098] S1. Ground-penetrating radar is used to detect shallow reinforced mesh-covered scenes and acquire corresponding data; specifically, the following steps are included:

[0099] Construct a shallow steel mesh shielding scenario; the shallow steel mesh shielding scenario is a scenario containing only steel mesh.

[0100] Ground-penetrating radar was used to detect shallow steel mesh-covered scenes, and several sets of B-scan echo data containing only steel mesh were obtained; each set of data corresponds to a set of steel bar numbers and spacing between steel bars; the steel bar burial depth is the same in all data.

[0101] S2. Based on the data obtained in step S1, and using the accumulated energy data, perform image processing to obtain the training dataset; specifically, this includes the following steps:

[0102] The several sets of B-scan echo data containing only steel mesh obtained in step S1 are converted into several frames of a set size (e.g., B-scan images containing only rebar mesh echo data;

[0103] Image enhancement operations are performed on several B-scan images containing only rebar mesh echo data to obtain several enhanced B-scan images containing only rebar mesh echo data; the image enhancement operations include image horizontal flipping and image scaling;

[0104] Using simulation software (such as gprMax software), irregular cavities and faults of random size and location of defect targets are constructed, and ground-penetrating radar detection simulation is carried out to obtain several sets of B-scan echo data containing targets.

[0105] The obtained B-scan echo data containing the target are converted into several frames of a set size (e.g., ...). B-scan image containing target echo;

[0106] The background removal method was used to remove the direct wave from several B-scan images containing the target echo, resulting in several B-scan images containing only the target echo.

[0107] Based on the cumulative energy data of the images, several enhanced B-scan images containing only steel mesh echo data and several B-scan images containing only target echo data are fused to obtain the training dataset.

[0108] In practice, the integration process includes the following steps:

[0109] The maximum absolute amplitude of the enhanced B-scan image containing only rebar mesh echo data is calculated and multiplied by a random coefficient within a set range (the random coefficient is set to 0.1~0.4) to obtain the simulated target amplitude.

[0110] Based on the obtained simulated target amplitude, the B-scan image containing only the target echo is normalized to obtain the target image;

[0111] The absolute value of the enhanced B-scan image containing only steel mesh echo data is taken to obtain the energy map; for the energy map, the cumulative sum is calculated along the depth direction of the image to obtain the cumulative energy distribution map;

[0112] The obtained cumulative energy distribution map is normalized, and the transmission attenuation mask is calculated based on the exponential attenuation function; where the attenuation coefficient of the exponential function simulates the degree of absorption and scattering of electromagnetic waves by the medium.

[0113] The obtained target image is multiplied pixel by pixel with the obtained transmission attenuation mask (the purpose is to simulate the attenuation response of the target signal after passing through the steel mesh), and the result is superimposed on the enhanced B-scan image containing only steel mesh echo data.

[0114] Complete image fusion;

[0115] The fusion process designed in this invention accurately reproduces the energy loss that occurs when electromagnetic waves penetrate the steel mesh, realistically reproducing the physical phenomenon that the signal of the target weakens sharply directly below the steel mesh. Moreover, this process uses the depth-accumulated energy of the measured echo to dynamically calculate the mask, so that the attenuation degree can be adaptively matched according to the intensity and depth of local steel mesh clutter, which is more in line with complex underground environments. At the same time, in actual engineering, it is difficult to obtain paired measured data of "with / without steel mesh interference" at the same location, and the problem of "domain difference" caused by the overly idealized nature of pure simulation data is addressed by the process of physically fusing the real measured steel mesh background with the simulated target. This not only avoids the physical limitations of obtaining paired measured labels, but also injects highly realistic complex environmental features into the training set.

[0116] S3. Based on the U-net network structure, and combining convolutional units, pooling units, attention mechanisms, interpolation operations, and skip connection mechanisms, an initial model for shallow steel mesh echo suppression is constructed.

[0117] The constructed initial model for shallow steel mesh echo suppression includes an encoding module, a feature enhancement module, a decoding module, and an output module; its structure diagram is shown below. Figure 2 As shown;

[0118] An encoding module is constructed based on convolutional and pooling units; the encoding module is used to extract features and downsample the input image data.

[0119] A feature enhancement module is constructed based on an attention mechanism; the feature enhancement module is a skip connection between the encoding module and the decoding module, used to calibrate various features in the encoding module;

[0120] A decoding module is constructed based on convolutional units and interpolation operations; the decoding module is used to fuse the features output by the encoding module and the features output by the feature enhancement module.

[0121] An output module is constructed based on convolutional units; the output module is used to map the features output by the decoding module to obtain the final ground-penetrating radar detection image with echo suppression achieved for shallow steel mesh.

[0122] In practical implementation, the constructed encoding module specifically includes the following:

[0123] The encoding module includes an input encoding convolutional layer, a first encoding max pooling layer, a first encoding convolutional layer, a second encoding max pooling layer, a second encoding convolutional layer, a third encoding max pooling layer, a third encoding convolutional layer, a fourth encoding max pooling layer, a fourth encoding convolutional layer, and an encoding multi-scale convolutional block attention layer, all connected in sequence. The output of the encoding multi-scale convolutional block attention layer is the output of the encoding module.

[0124] Specifically, the output of the input coding convolutional layer is the first skip connection input, the output of the first coding convolutional layer is the second skip connection input, the output of the second coding convolutional layer is the third skip connection input, and the output of the third coding convolutional layer is the fourth skip connection input; all three skip connection inputs are input to the feature enhancement module.

[0125] The input encoding convolutional layer, the first encoding convolutional layer, the second encoding convolutional layer, the third encoding convolutional layer, and the fourth encoding convolutional layer all have the same structure; the encoding convolutional layer consists of two concatenated convolutional sub-layers; the convolutional sub-layers consist of sequentially concatenated... Convolutional layers, batch normalization layers, and ReLU activation function layers. The convolution kernel of the convolutional layer is The step size is 1, and the fill size is 1.

[0126] The first, second, third, and fourth encoded max pooling layers all have the same structure; the encoded max pooling layer has a pooling kernel size of [missing value]. Maximum pooling layer;

[0127] Based on channel attention and spatial attention mechanisms, a multi-scale convolutional block attention layer is constructed; the multi-scale convolutional block attention layer is used to perform global attention weighting on deep semantic features.

[0128] In practical implementation, the feature enhancement module includes the following components:

[0129] The feature enhancement module includes a first skip connection layer, a second skip connection layer, a third skip connection layer, and a fourth skip connection layer;

[0130] The input to the first jump connection layer is the first jump connection input, and the output of the first jump connection layer is the first jump connection output; the input to the second jump connection layer is the second jump connection input, and the output of the second jump connection layer is the second jump connection output; the input to the third jump connection layer is the third jump connection input, and the output of the third jump connection layer is the third jump connection output; the input to the fourth jump connection layer is the fourth jump connection input, and the output of the fourth jump connection layer is the fourth jump connection output.

[0131] The outputs from the first jump connection to the fourth jump connection are all output to the decoding module;

[0132] The first, second, third, and fourth skip connection layers have the same structure; the skip connection layer includes a feature calibration residual dense sub-layer and an enhanced multi-scale convolutional block attention sub-layer connected in sequence.

[0133] Based on convolutional and pooling units, a feature calibration residual dense sub-layer is constructed. The feature calibration residual dense sub-layer does not blindly extract features from the encoded feature map, but performs calibration immediately after each layer of convolution generates features to avoid noise accumulating and amplifying layer by layer.

[0134] Based on channel attention and spatial attention mechanisms, an enhanced multi-scale convolutional block attention sublayer is constructed. The enhanced multi-scale convolutional block attention sublayer plays a secondary filtering role, using multi-scale perspectives to further eliminate background interference, ensuring that only the most significant steel mesh and defect features are transmitted to the decoding part of the network.

[0135] In practical implementation, the decoding module includes the following:

[0136] The decoding module includes a first decoding upsampling layer, a first decoding convolutional layer, a second decoding upsampling layer, a second decoding convolutional layer, a third decoding upsampling layer, a third decoding convolutional layer, a fourth decoding upsampling layer, and a fourth decoding convolutional layer;

[0137] The output of the encoding module is concatenated with the output of the fourth skip connection along the channel dimension, and then used as the input of the first decoding upsampling layer. The output of the first decoding upsampling layer is used as the input of the first decoding convolutional layer. The output of the first decoding convolutional layer is concatenated with the output of the third skip connection along the channel dimension, and then used as the input of the second decoding upsampling layer. The output of the second decoding upsampling layer is used as the input of the second decoding convolutional layer. The output of the second decoding convolutional layer is concatenated with the output of the second skip connection along the channel dimension, and then used as the input of the third decoding upsampling layer. The output of the third decoding upsampling layer is used as the input of the third decoding convolutional layer. The output of the third decoding convolutional layer is concatenated with the output of the first skip connection along the channel dimension, and then used as the input of the fourth decoding upsampling layer. The output of the fourth decoding upsampling layer is used as the input of the fourth decoding convolutional layer. The input of the fourth decoding convolutional layer is the output of the decoding module.

[0138] The processing procedures for the first, second, third, and fourth decoding upsampling layers are all the same; the decoding upsampling layer uses bilinear interpolation for upsampling.

[0139] The first, second, third, and fourth decoding convolutional layers all have the same structure; each decoding convolutional layer consists of two concatenated convolutional sub-layers; each convolutional sub-layer consists of sequentially concatenated... Convolutional layers, batch normalization layers, and ReLU activation function layers. The convolution kernel of the convolutional layer is The step size is 1, and the fill size is 1.

[0140] In practice, the output module includes the following:

[0141] The output module includes a first output convolutional layer and a second output convolutional layer connected in series.

[0142] The first output convolutional layer consists of two concatenated convolutional sub-layers; the convolutional sub-layers consist of sequentially concatenated... Convolutional layers, batch normalization layers, and ReLU activation function layers. The convolution kernel of the convolutional layer is The step size is 1, and the fill size is 1.

[0143] The second output convolutional layer has a kernel size of [size missing]. The first convolutional layer; the second output convolutional layer is used to map the multi-channel feature map into a single-channel output, thereby obtaining the final target image after steel mesh clutter suppression;

[0144] In practice, the structures of the encoding multi-scale convolutional block attention layer and the enhanced multi-scale convolutional block attention sub-layer are the same; the structure of the multi-scale convolutional block attention layer is as follows: Figure 3 As shown, it includes the following:

[0145] The input features are processed by the channel attention sub-layer to obtain the channel attention map; the channel attention map is multiplied element-wise with the input features to obtain the intermediate feature map; the intermediate feature map is processed by the multi-scale spatial attention sub-layer to obtain the multi-scale spatial attention map; the multi-scale spatial attention map is multiplied element-wise with the intermediate feature map to obtain the output of the multi-scale convolutional block attention layer.

[0146] The processing procedure of the channel attention sublayer includes the following:

[0147] The input features of the channel attention sublayer are divided into two paths;

[0148] The first feature path passes sequentially through an average pooling layer and a convolutional kernel size of [missing information]. The size of the convolutional layer, ReLU activation function layer, and convolutional kernel is The convolutional layer processing yields the average pooling result features;

[0149] The second feature path passes sequentially through a max pooling layer and a convolutional kernel size of [missing information]. The size of the convolutional layer, ReLU activation function layer, and convolutional kernel is The convolutional layer processing yields the max pooling result features;

[0150] The obtained average pooling result features and max pooling result features are added element-wise, and then processed through a Sigmoid activation function layer to obtain the channel attention map;

[0151] The processing procedure of the multi-scale spatial attention sublayer includes the following:

[0152] The average channel feature map is obtained by averaging the input features of the multi-scale spatial attention sub-layer in the channel dimension.

[0153] The maximum value of the input features of the multi-scale spatial attention sublayer is calculated in the channel dimension to obtain the maximum channel feature map.

[0154] The average channel feature map and the maximum channel feature map are concatenated along the channel dimension to obtain the compressed spatial feature map;

[0155] The compressed spatial feature map is processed simultaneously through three convolutional branches; the first convolutional branch includes a standard convolutional layer and a batch normalization layer connected in sequence, with the kernel size of the standard convolutional layer being [missing value]. The padding is 1, and the dilation rate is 1; the second convolutional branch consists of a series of dilated convolutional layers and batch normalized layers, with the kernel size of the dilated convolutional layers being 1. The padding is 2, and the dilation rate is 2; the third convolutional branch consists of a series of dilated convolutional layers and batch normalized layers, with the kernel size of the dilated convolutional layers being [missing value]. The padding is 4 and the expansion rate is 4; the multi-path convolution branch scheme can extract spatial context information of different receptive fields;

[0156] The outputs of the three convolutional branches are concatenated along the channel dimension, and then sequentially passed through convolutional kernels of size [size missing]. The convolutional layers and sigmoid activation function layers are used to process the data to obtain a multi-scale spatial attention map;

[0157] In practical implementation, the structure of the feature calibration residual dense sublayer is as follows: Figure 4 As shown; the processing procedure for the feature calibration residual dense sublayer includes the following:

[0158] The input features of the feature calibration residual dense sub-layer are processed sequentially through several dense calibration units, and then through a convolutional kernel of size [size missing]. The convolutional layer is used to process the features to obtain the feature fusion sub-features; this convolutional layer is used to map the cascaded high-dimensional features back to the original input channel dimension.

[0159] The feature fusion sub-features are added element-wise to the input features of the feature calibration residual dense sub-layer to obtain the output features of the feature calibration residual dense sub-layer.

[0160] The processing steps of the dense calibration unit include the following:

[0161] The calibration dense unit is used to receive preceding features, generate new features, perform channel-dimensional weight calibration on the new features, and concatenate the calibrated features with the input features along the channel dimension to obtain the output of the dense calibration unit.

[0162] The input features of the dense calibration unit are processed sequentially through convolutional layers and ReLU activation function layers to obtain intermediate calibration sub-features; wherein, the kernel size of the convolutional layer is [missing information]. The step size is 1, and the fill size is 1.

[0163] The intermediate calibrator features are processed through a global average pooling layer, then sequentially through a first fully connected layer, a ReLU activation function layer, a second fully connected layer, and a Sigmoid activation function layer to obtain the channel weight vector; the global average pooling layer is used to process the channel weight vector with a spatial size of The intermediate feature map is compressed into The channel descriptor; a fully connected layer + ReLU activation function combination is used to compress the channel bit depth; a fully connected layer + Sigmoid activation function combination is used to restore the channel bit depth;

[0164] The calibration feature map is obtained by multiplying the channel weight vector with the intermediate calibration sub-features channel by channel to suppress invalid features and enhance valid features.

[0165] The calibration feature map is concatenated with the input features of the dense calibration unit along the channel dimension to obtain the output of the dense calibration unit.

[0166] S4. Using the training dataset obtained in step S2, train the initial shallow steel mesh echo suppression model constructed in step S3 to obtain the shallow steel mesh echo suppression model.

[0167] During training, the following loss function is used:

[0168]

[0169]

[0170]

[0171]

[0172] In the formula The value of the loss function; The height of the input image; The width of the input image; For real images The middle position is The actual pixel value at that location; For predicting images The middle position is Pixel value at; These are the brightness comparison metrics between the real and predicted images at M scales; The first index is set to adjust the relative importance of different components across M scales; The total number of scales representing the quality of the real and predicted images; This is a comparison metric between the real image and the predicted image at the k-th scale. The second index is set to adjust the relative importance of different components at the k-th scale; This is a measure of the structural comparison between the real image and the predicted image at the k-th scale. The third index is set to adjust the relative importance of different components at the k-th scale; The first parameter is set; This is the second parameter that is set; This is the third parameter that is set; The average pixel value of the real image; To predict the mean pixel value of the image; This represents the variance of the pixel values ​​in the actual image. To predict the variance of pixel values ​​in the image; The covariance between the pixel values ​​of the real image and the pixel values ​​of the predicted image;

[0173] S5. Using the shallow steel mesh echo suppression model obtained in step S4, the echo suppression of the shallow steel mesh is performed in the actual ground penetrating radar detection process.

[0174] The effects of the method of the present invention will be illustrated by the following embodiment:

[0175] The present invention is compared with existing solutions; the existing solution is the one proposed by Hai-Han Sun in the paper "Learning to Remove Clutter in Real-World GPR Images Using Hybrid Data";

[0176] In clutter suppression models, metrics for evaluating network performance include: Mean Absolute Error (MAE), Mean Square Error (MSE), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Multi-scale Structural Similarity Index Measure (MS-SSIM). MAE measures the average absolute difference between the predicted image and the ground truth; a lower value is better. MSE measures the average of the squared differences between the predicted image and the ground truth; a lower value is better. PSNR measures the reconstruction quality of the image; a higher value indicates better clutter removal. SSIM and MS-SSIM measure the structural similarity between the predicted image and the ground truth; higher values ​​indicate greater similarity.

[0177] The proposed solution and existing solutions were trained on the same training set and tested on the same test set. The performance indicators obtained during the testing process are shown in Table 1.

[0178]

[0179] As shown in Table 1, the proposed solution outperforms existing solutions in all five performance indicators: MAE is reduced by 26.13%, indicating that the proposed solution has a smaller average deviation at the pixel level and the target echo recovery result is closer to the true value; MSE is reduced by 38.10%, indicating that the proposed solution greatly reduces the occurrence of significant outliers or large deviations; PSNR is improved by 3.04dB. In the field of image processing, a 3dB improvement is usually considered a significant quality leap, meaning that the proposed solution has a substantial improvement in the ability to suppress steel mesh echoes and recover target echoes; SSIM and MS-SSIM are closer to 1, proving that the proposed solution has better consistency and recovery ability when processing image details at different scales.

[0180] Depend on Figures 5-7 As can be seen, under sparse rebar mesh conditions, both networks can effectively suppress rebar mesh echoes and recover the echoes of targets obscured by the rebar mesh. Under denser rebar mesh conditions, compared with existing solutions, the present invention can significantly reduce clutter residue while recovering the echoes of targets obscured by the rebar mesh. Under dense rebar mesh conditions, although the target echoes recovered by both networks will have a certain degree of distortion and clutter residue, the comparison shows that the rebar mesh echo suppression effect of the present invention is better: the recovered target echo is closer to the true value, and there is less clutter residue.

[0181] To further verify the generalizability of the present invention, a simulation model was constructed using gprMax: four steel bars simulated a steel mesh, and an irregular void was constructed below the steel mesh. B-scan images of the void below the steel mesh were obtained through simulation calculations, and then processed using both the present invention's method and existing methods. The results are as follows. Figure 8 As shown, compared with existing solutions, the present invention can effectively suppress steel mesh echoes and reduce clutter residue and artifacts; moreover, the present invention has a stronger ability to recover weak targets, the recovered void targets are more obvious, and their multiple echo characteristics are clearer.

[0182] Further experiments were conducted in a laboratory sandbox: three steel bars were used to simulate a steel mesh, and a fault defect target was pre-embedded below the steel mesh. B-scan images of the fault target under the steel mesh were obtained, and then processed using both the present invention's method and existing methods. The results are as follows. Figure 9 As shown. Through comparison, it can be seen that: the present invention can not only suppress the rebar mesh echo, but also recover the double-layer echo characteristics of the fault defect, which is convenient for subsequent defect identification and location, and has certain engineering significance; although the existing solution can also suppress the rebar mesh echo, its recovery ability is not as good as the present invention, and the result after processing by the existing solution still has a large number of artifacts, which seriously affects the subsequent defect identification and location.

[0183] In summary, the present invention can effectively separate the rebar mesh echo from the target signal. Its rebar mesh echo suppression performance and target imaging quality are superior to existing solutions, significantly improving GPR's ability to detect weak targets in shallow rebar mesh occlusion scenarios.

[0184] like Figure 10 The diagram shows the functional modules of the system of this invention: The system for implementing the shallow rebar mesh echo suppression method based on deep learning disclosed in this invention includes a data acquisition module, a data processing module, a model building module, a model training module, and an echo suppression module; the data acquisition module, data processing module, model building module, model training module, and echo suppression module are connected in series; the data acquisition module is used to detect the shallow rebar mesh occlusion scene using ground penetrating radar and acquire the corresponding data information, and upload the data information to the data processing module; the data processing module is used to perform image processing based on the received data information and the acquired data information, based on the cumulative energy data, to obtain a training dataset, and upload the data information to the model building module; the model building module is used to construct an initial model for shallow rebar mesh echo suppression based on the received data information, using the U-net network structure, combined with convolutional units, pooling units, attention mechanisms, interpolation operations, and skip connection mechanisms, and upload the data information to the model training module; the constructed initial model for shallow rebar mesh echo suppression includes an encoding module and feature enhancement. The system comprises a module, a decoding module, and an output module. An encoding module is constructed based on convolutional and pooling units. This module extracts and downsamples the input image data. A feature enhancement module is constructed based on an attention mechanism. This module acts as a jump connection between the encoding and decoding modules to calibrate various features within the encoding module. A decoding module is constructed based on convolutional units and interpolation operations. This module fuses the features output by the encoding and enhancement modules. An output module is constructed based on convolutional units. This module maps the features output by the decoding module to obtain the final ground-penetrating radar image with echo suppression achieved for shallow rebar mesh. A model training module trains the initial shallow rebar mesh echo suppression model using the received data and training dataset, obtaining the shallow rebar mesh echo suppression model. The training module then uploads the data to the echo suppression module. Finally, the echo suppression module uses the obtained shallow rebar mesh echo suppression model to perform echo suppression of the shallow rebar mesh during actual ground-penetrating radar detection.

Claims

1. A method for suppressing echoes from shallow steel mesh based on deep learning, characterized in that... Includes the following steps: S1. Ground-penetrating radar is used to detect shallow steel mesh-covered scenes and acquire corresponding data information; S2. Based on the data obtained in step S1, perform image processing on the accumulated energy data to obtain the training dataset; S3. Based on the U-net network structure, and combining convolutional units, pooling units, attention mechanisms, interpolation operations, and skip connection mechanisms, an initial model for shallow steel mesh echo suppression is constructed. The constructed initial model for shallow steel mesh echo suppression includes an encoding module, a feature enhancement module, a decoding module, and an output module; An encoding module is constructed based on convolutional and pooling units; the encoding module is used to extract features and downsample the input image data. A feature enhancement module is constructed based on the attention mechanism; The feature enhancement module is a skip connection between the encoding and decoding modules, used to calibrate various features in the encoding module; A decoding module is constructed based on convolutional units and interpolation operations; the decoding module is used to fuse the features output by the encoding module and the features output by the feature enhancement module. An output module is constructed based on convolutional units; the output module is used to map the features output by the decoding module to obtain the final ground-penetrating radar detection image with echo suppression achieved for shallow steel mesh. S4. Using the training dataset obtained in step S2, train the initial shallow steel mesh echo suppression model constructed in step S3 to obtain the shallow steel mesh echo suppression model. S5. Using the shallow steel mesh echo suppression model obtained in step S4, the echo suppression of the shallow steel mesh is performed in the actual ground penetrating radar detection process.

2. The method for suppressing echoes from shallow steel mesh based on deep learning according to claim 1, characterized in that... Step S1 specifically includes the following steps: Construct a shallow steel mesh shielding scenario; the shallow steel mesh shielding scenario is a scenario containing only steel mesh. Ground-penetrating radar was used to detect shallow steel mesh-covered scenes, obtaining several sets of B-scan echo data containing only the steel mesh; each set of data corresponds to a set of steel bar numbers and spacing between steel bars; the steel bar burial depth is the same in all data.

3. The method for suppressing echoes from shallow steel mesh based on deep learning according to claim 2, characterized in that... Step S2 specifically includes the following steps: The several sets of B-scan echo data containing only steel mesh obtained in step S1 are converted into several B-scan images of a set size containing only steel mesh echo data. Image enhancement operations are performed on several B-scan images containing only rebar mesh echo data to obtain several enhanced B-scan images containing only rebar mesh echo data; the image enhancement operations include image horizontal flipping and image scaling; Using simulation software, irregular cavities and faults of randomly sized and located defective targets were constructed, and ground-penetrating radar detection simulations were performed to obtain several sets of B-scan echo data containing the targets. The obtained B-scan echo data containing the target are converted into several B-scan images of a set size containing the target echo. The background removal method was used to remove the direct wave from several B-scan images containing the target echo, resulting in several B-scan images containing only the target echo. Based on the cumulative energy data of the images, several enhanced B-scan images containing only steel mesh echo data and several B-scan images containing only target echo data are fused to obtain the training dataset. The fusion process specifically includes the following steps: The maximum absolute amplitude of the enhanced B-scan image containing only rebar mesh echo data is calculated and multiplied by a random coefficient within a set range to obtain the simulated target amplitude. Based on the obtained simulated target amplitude, the B-scan image containing only the target echo is normalized to obtain the target image; The absolute value of the enhanced B-scan image containing only steel mesh echo data is taken to obtain the energy map; for the energy map, the cumulative sum is calculated along the depth direction of the image to obtain the cumulative energy distribution map; The obtained cumulative energy distribution map is normalized, and the transmission attenuation mask is calculated based on the exponential attenuation function. The obtained target image is multiplied pixel by pixel with the obtained transmission attenuation mask, and the result is superimposed on the enhanced B-scan image containing only steel mesh echo data; Complete the image fusion.

4. The method for suppressing echoes from shallow steel mesh based on deep learning according to claim 3, characterized in that... The constructed encoding module specifically includes the following: The encoding module includes an input encoding convolutional layer, a first encoding max pooling layer, a first encoding convolutional layer, a second encoding max pooling layer, a second encoding convolutional layer, a third encoding max pooling layer, a third encoding convolutional layer, a fourth encoding max pooling layer, a fourth encoding convolutional layer, and an encoding multi-scale convolutional block attention layer, all connected in sequence. The output of the encoding multi-scale convolutional block attention layer is the output of the encoding module. Specifically, the output of the input coding convolutional layer is the first skip connection input, the output of the first coding convolutional layer is the second skip connection input, the output of the second coding convolutional layer is the third skip connection input, and the output of the third coding convolutional layer is the fourth skip connection input; all three skip connection inputs are input to the feature enhancement module. The input encoding convolutional layer, the first encoding convolutional layer, the second encoding convolutional layer, the third encoding convolutional layer, and the fourth encoding convolutional layer all have the same structure; the encoding convolutional layer consists of two concatenated convolutional sub-layers; the convolutional sub-layers consist of sequentially concatenated... Convolutional layers, batch normalization layers, and ReLU activation function layers. The convolution kernel of the convolutional layer is The step size is 1, and the fill size is 1. The first, second, third, and fourth encoded max pooling layers all have the same structure; the encoded max pooling layer has a pooling kernel size of [missing value]. Maximum pooling layer; Based on channel attention and spatial attention mechanisms, we construct an attention layer that encodes multi-scale convolutional blocks.

5. The method for suppressing echoes from shallow steel mesh based on deep learning according to claim 4, characterized in that... The constructed feature enhancement module specifically includes the following: The feature enhancement module includes a first skip connection layer, a second skip connection layer, a third skip connection layer, and a fourth skip connection layer; The input to the first hop connection layer is the first hop connection input, and the output of the first hop connection layer is the first hop connection output. The input to the second jump connection layer is the second jump connection input, and the output of the second jump connection layer is the second jump connection output; the input to the third jump connection layer is the third jump connection input, and the output of the third jump connection layer is the third jump connection output; the input to the fourth jump connection layer is the fourth jump connection input, and the output of the fourth jump connection layer is the fourth jump connection output; The outputs from the first jump connection to the fourth jump connection are all output to the decoding module; The first, second, third, and fourth skip connection layers have the same structure; the skip connection layer includes a feature calibration residual dense sub-layer and an enhanced multi-scale convolutional block attention sub-layer connected in sequence. A dense sublayer for feature calibration residuals is constructed based on convolutional and pooling units. Based on channel attention and spatial attention mechanisms, an enhanced multi-scale convolutional block attention sublayer is constructed.

6. The method for suppressing echoes from shallow steel mesh based on deep learning according to claim 5, characterized in that... The constructed decoding module includes the following: The decoding module includes a first decoding upsampling layer, a first decoding convolutional layer, a second decoding upsampling layer, a second decoding convolutional layer, a third decoding upsampling layer, a third decoding convolutional layer, a fourth decoding upsampling layer, and a fourth decoding convolutional layer; The output of the encoding module is concatenated with the output of the fourth skip connection along the channel dimension, and then used as the input of the first decoding upsampling layer. The output of the first decoding upsampling layer is used as the input of the first decoding convolutional layer. The output of the first decoding convolutional layer is concatenated with the output of the third skip connection along the channel dimension, and then used as the input of the second decoding upsampling layer. The output of the second decoding upsampling layer is used as the input of the second decoding convolutional layer. The output of the second decoding convolutional layer is concatenated with the output of the second skip connection along the channel dimension, and then used as the input of the third decoding upsampling layer. The output of the third decoding upsampling layer is used as the input of the third decoding convolutional layer. The output of the third decoding convolutional layer is concatenated with the output of the first skip connection along the channel dimension, and then used as the input of the fourth decoding upsampling layer. The output of the fourth decoding upsampling layer is used as the input of the fourth decoding convolutional layer. The input of the fourth decoding convolutional layer is the output of the decoding module. The processing procedures for the first, second, third, and fourth decoding upsampling layers are all the same; the decoding upsampling layer uses bilinear interpolation for upsampling. The first, second, third, and fourth decoding convolutional layers all have the same structure; each decoding convolutional layer consists of two concatenated convolutional sub-layers; each convolutional sub-layer consists of sequentially concatenated... Convolutional layers, batch normalization layers, and ReLU activation function layers. The convolution kernel of the convolutional layer is The step size is 1, and the padding size is 1.

7. The method for suppressing echoes from shallow steel mesh based on deep learning according to claim 6, characterized in that... The output module constructed specifically includes the following: The output module includes a first output convolutional layer and a second output convolutional layer connected in series. The first output convolutional layer consists of two concatenated convolutional sub-layers; the convolutional sub-layers consist of sequentially concatenated... Convolutional layers, batch normalization layers, and ReLU activation function layers. The convolution kernel of the convolutional layer is The step size is 1, and the fill size is 1. The second output convolutional layer has a kernel size of [size missing]. The convolutional layer.

8. The method for suppressing echoes from shallow steel mesh based on deep learning according to claim 7, characterized in that... The structure of the encoding multi-scale convolutional block attention layer and the enhanced multi-scale convolutional block attention sub-layer is the same; the multi-scale convolutional block attention layer includes the following: The input features are processed by the channel attention sub-layer to obtain the channel attention map; the channel attention map is multiplied element-wise with the input features to obtain the intermediate feature map; the intermediate feature map is processed by the multi-scale spatial attention sub-layer to obtain the multi-scale spatial attention map; the multi-scale spatial attention map is multiplied element-wise with the intermediate feature map to obtain the output of the multi-scale convolutional block attention layer. The processing procedure of the channel attention sublayer includes the following: The input features of the channel attention sublayer are divided into two paths; The first feature path passes sequentially through an average pooling layer and a convolutional kernel size of [missing information]. The size of the convolutional layer, ReLU activation function layer, and convolutional kernel is The convolutional layer processing yields the average pooling result features; The second feature path passes sequentially through a max pooling layer and a convolutional kernel size of [missing information]. The size of the convolutional layer, ReLU activation function layer, and convolutional kernel is The convolutional layer processing yields the max pooling result features; The obtained average pooling result features and max pooling result features are added element-wise, and then processed through a Sigmoid activation function layer to obtain the channel attention map; The processing procedure of the multi-scale spatial attention sublayer includes the following: The average channel feature map is obtained by averaging the input features of the multi-scale spatial attention sub-layer in the channel dimension. The maximum value of the input features of the multi-scale spatial attention sublayer is calculated in the channel dimension to obtain the maximum channel feature map. The average channel feature map and the maximum channel feature map are concatenated along the channel dimension to obtain the compressed spatial feature map; The compressed spatial feature map is processed simultaneously through three convolutional branches; the first convolutional branch includes a standard convolutional layer and a batch normalization layer connected in sequence, with the kernel size of the standard convolutional layer being [missing value]. The padding is 1, and the dilation rate is 1; the second convolutional branch consists of a series of dilated convolutional layers and batch normalized layers, with the kernel size of the dilated convolutional layers being 1. The padding is 2, and the dilation rate is 2; the third convolutional branch consists of a series of dilated convolutional layers and batch normalized layers, with the kernel size of the dilated convolutional layers being [missing value]. The filler is 4, and the expansion rate is 4. The outputs of the three convolutional branches are concatenated along the channel dimension, and then sequentially passed through convolutional kernels of size [size missing]. The convolutional layers and sigmoid activation function layers are used to process the data to obtain a multi-scale spatial attention map.

9. The method for suppressing echoes from shallow steel mesh based on deep learning according to claim 8, characterized in that... The process of calibrating dense sublayers of residuals includes the following: The input features of the feature calibration residual dense sub-layer are processed sequentially through several dense calibration units, and then through a convolutional kernel of size [size missing]. The convolutional layers are processed to obtain feature fusion sub-features; The feature fusion sub-features are added element-wise to the input features of the feature calibration residual dense sub-layer to obtain the output features of the feature calibration residual dense sub-layer. The processing steps of the dense calibration unit include the following: The input features of the dense calibration unit are processed sequentially through convolutional layers and ReLU activation function layers to obtain intermediate calibration sub-features; wherein, the kernel size of the convolutional layer is [missing information]. The step size is 1, and the fill size is 1. The intermediate calibrator features are processed by a global average pooling layer, and then sequentially processed by a first fully connected layer, a ReLU activation function layer, a second fully connected layer, and a Sigmoid activation function layer to obtain the channel weight vector. The calibration feature map is obtained by multiplying the channel weight vector with the intermediate calibration sub-feature channel by channel. The calibration feature map is concatenated with the input features of the dense calibration unit along the channel dimension to obtain the output of the dense calibration unit.

10. A system for implementing the deep learning-based shallow steel mesh echo suppression 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, and an echo suppression module; the data acquisition module, data processing module, model building module, model training module, and echo suppression module are connected in series; the data acquisition module is used to use ground penetrating radar to detect shallow steel mesh-covered scenes, acquire corresponding data information, and upload the data information to the data processing module; The data processing module is used to perform image processing based on the received data and the acquired data, and based on the cumulative energy data, to obtain the training dataset, and then upload the data to the model building module. The model building module is used to construct an initial model for shallow rebar mesh echo suppression based on the received data and the U-net network structure, combining convolutional units, pooling units, attention mechanisms, interpolation operations, and skip connection mechanisms. The module then uploads the data to the model training module. The constructed initial model includes an encoding module, a feature enhancement module, a decoding module, and an output module. The encoding module is built based on convolutional and pooling units. It performs feature extraction and downsampling on the input image data. The feature enhancement module is built based on an attention mechanism. The feature enhancement module is a skip connection between the encoding and decoding modules, used to calibrate various features in the encoding module. A decoding module is constructed based on convolutional units and interpolation operations. This module fuses the features output from the encoding and feature enhancement modules. An output module is constructed based on convolutional units. This module maps the features output from the decoding module to obtain the final ground-penetrating radar image with echo suppression achieved for shallow rebar mesh. The model training module trains the initial shallow rebar mesh echo suppression model using the received data and the obtained training dataset, obtaining the shallow rebar mesh echo suppression model, and uploads the data to the echo suppression module. The echo suppression module, based on the received data and the obtained shallow rebar mesh echo suppression model, performs echo suppression of the shallow rebar mesh during actual ground-penetrating radar detection.

Images