Ground penetrating radar data real-time processing method and system and medium

Abstract

The invention discloses a ground penetrating radar data real-time processing method and system and a medium, and belongs to the field of underground detection, and the method comprises the steps: carrying out the real-time processing of ground penetrating radar data, inputting collected B-scan image data into a preset defect discrimination model, carrying out the automatic real-time discrimination, and outputting a defect recognition result; performing image processing on the B-scan image data according to a preset image processing model, and displaying the B-scan image data to a user; wherein the defect discrimination model and the image processing model are obtained by migrating and training an encoder in an auto-encoder model; therefore, by implementing the method, high-precision and automatic underground disease real-time identification can be realized under the condition of not depending on a large number of manually labeled samples, including effective processing of strong noise interference and electromagnetic wave scattering effect in GPR data, improvement of GPR image identification data discrimination precision and real-time performance, and improvement of the generalization ability of the model.

Description

A method, system and medium for real-time processing of ground-penetrating radar data Technical Field

[0001] This invention belongs to the field of underground detection, specifically relating to a method, system, and medium for real-time processing of ground-penetrating radar data. Background Technology

[0002] Currently, the interpretation of Ground Penetrating Radar (GPR) data mainly relies on manual interpretation, which involves observing the reflected wave characteristics in B-scan (two-dimensional profile) images to identify underground targets or defects. However, manual interpretation has drawbacks such as strong subjectivity, low efficiency, and high dependence on personnel experience, making it difficult to cope with large-scale and long-term detection tasks.

[0003] To overcome the limitations of manual interpretation, some existing technologies have developed methods based on image processing and traditional machine learning to automatically identify targets in GPR images. However, these methods typically suffer from the following shortcomings: 1. Strong feature dependency: Features need to be designed and extracted manually, and the quality of the features directly determines the accuracy of the identification, resulting in poor generalization ability; 2. Limited accuracy: For GPR data with low signal-to-noise ratio, unclear target features, or complex backgrounds, the accuracy and robustness of traditional methods are often unsatisfactory; 3. Non-end-to-end: Data processing, feature extraction, and classification are usually separate steps, resulting in a complex process that makes true real-time processing difficult; 4. Ignoring global contextual relationships: Traditional methods mostly focus on local features and struggle to effectively capture long-range dependencies across distance and depth in B-scan images.

[0004] In recent years, deep learning technology, especially convolutional neural networks, has achieved great success in the field of image recognition. Although some studies have attempted to apply CNN (convolutional neural network) to GPR target recognition, most of them directly use general image classification networks. This not only fails to fully consider the unique sequential nature and global contextual correlation of GPR data, resulting in insufficient feature extraction and affecting the discrimination accuracy and real-time performance in complex scenes, but also fails to solve the problems of strong noise interference and difficulty in feature extraction in GPR data, and lacks the ability to process the unique electromagnetic wave scattering characteristics of GPR data. Summary of the Invention

[0005] This invention provides a method, system, and medium for real-time processing of ground-penetrating radar (GPR) data, which can improve the accuracy and real-time performance of GPR image recognition data.

[0006] This invention provides a real-time processing method for ground-penetrating radar data, comprising: acquiring first B-scan image data in real time; inputting the first B-scan image data into a preset defect discrimination model, so that the defect discrimination model processes strong noise interference and electromagnetic wave scattering effects in the first B-scan image data through a bidirectional attention mechanism, and automatically and in real-time discriminates the first B-scan image data, outputting defect identification results; wherein the defect discrimination model is obtained by transferring and training the first encoder in an autoencoder model; the autoencoder model is trained based on historical B-scan image data; when the defect identification results meet the manual verification triggering conditions, performing image processing on the first B-scan image data according to a preset image processing model to obtain second B-scan image data, and displaying the second B-scan image data to the user, so that the user can perform visual verification and decision support based on the second B-scan image data; wherein the image processing model is obtained by transferring and training the first encoder in an autoencoder model.

[0007] This invention employs a multi-model collaborative architecture based on unsupervised learning training and model transfer strategies, achieving an end-to-end processing flow and significantly improving the real-time performance of GPR data processing. By utilizing an encoder-decoder structure and a bidirectional attention mechanism to extract and identify features from B-scan image data, it addresses the challenges of strong noise interference and difficult feature extraction in GPR data. This enhances the processing capability for the unique electromagnetic wave scattering characteristics of GPR data, effectively capturing long-range dependencies and global contextual information, thus greatly improving the discrimination accuracy of GPR data processing. The defect discrimination model can output pixel-level segmentation maps for precise defect location or image-level classification results for rapid screening, depending on actual needs, combining accuracy and efficiency with wide applicability. The image processing model serves as an auxiliary verification tool, providing enhanced and clear visualization of defect identification results, reducing the burden of manual verification, and forming a closed-loop detection process based on human-machine mutual trust.

[0008] Furthermore, the autoencoder model is trained based on historical B-scan image data, including: constructing an initial model based on an encoder-decoder architecture, wherein the encoder of the initial model includes a multi-layer two-dimensional convolutional neural network and a bidirectional attention mechanism module, and the decoder of the initial model includes a deconvolution layer or an upsampling layer symmetrical to the encoder; acquiring historical B-scan image data, and training the initial model in unsupervised learning based on the historical B-scan image data until the initial model meets the training conditions, and outputting the autoencoder model; the first encoder in the autoencoder model extracts the spatial hierarchical features of the historical B-scan image data through a CNN layer, processes the strong noise interference and electromagnetic wave scattering effect in the historical B-scan image data through a bidirectional attention mechanism, and uses a two-dimensional calibration feature response to suppress noise interference and enhance electromagnetic wave scattering features, thereby obtaining compressed features containing long-range dependencies and global contextual information in the historical B-scan image data, and gradually reconstructing the obtained compressed feature mapping into a B-scan image with the same size as the input.

[0009] This ensures that the first encoder can effectively extract spatial hierarchical features such as edges, textures, and shapes of the image, while suppressing noise interference and enhancing electromagnetic wave scattering features. It can effectively capture long-range dependencies and global context information in GPR data, and enhance the feature extraction capability and generalization performance in complex noise environments. This is a key execution step to enable the first encoder to extract robust and general features of GPR data.

[0010] Furthermore, the defect discrimination model is obtained by transferring and training the first encoder in the autoencoder model. Specifically, the encoder of the defect discrimination model transfers the first encoder in the autoencoder model, and the decoder includes a segmentation and localization output layer and a classification image output layer. Historical B-scan image data and labeled pixel-level binary segmentation masks are collected as segmentation supervision data, and the segmentation task part of the defect discrimination model is optimized and trained using gradient descent combined with loss function BCE Loss or Dice Loss based on the segmentation supervision data. Historical B-scan image data and corresponding image-level category labels are collected as classification supervision data, and the classification task part of the defect discrimination model is optimized and trained using gradient descent combined with loss function classification cross-entropy loss based on the classification supervision data.

[0011] In this way, different supervised data are used to train the defect discrimination model according to different application scenarios. By minimizing the difference between the predicted heatmap output by the defect discrimination model and the real segmentation mask, and minimizing the difference between the predicted probability distribution output by the defect discrimination model and the real category label, the output result of the defect discrimination model converges to the human recognition result, thereby improving the recognition accuracy and recognition efficiency of the defect discrimination model.

[0012] Furthermore, the image processing model is obtained by transferring and training based on the encoder, specifically: the encoder of the image processing model transfers the first encoder in the autoencoder model, and the decoder configures the image output layer; historical B-scan image data and manually optimized historical B-scan image data are collected as image processing supervision data for training, and the image processing supervision data is used to optimize the training by employing the loss function mean square error or perceptual loss to obtain the trained image processing model.

[0013] By performing unsupervised learning training on the image processing model, the difference between the image output by the image processing model and the result after manual processing is minimized, so that the image processing model's image processing converges to the result of manual processing, thereby improving the image processing capability and efficiency of the image processing model.

[0014] Further, the automated real-time discrimination of the first B-scan image data and the output of defect recognition results specifically involves: sequentially inputting the first B-scan image data into the defect discrimination model for real-time discrimination, and obtaining a high-dimensional feature vector after data processing; sequentially inputting the first B-scan image data into the defect discrimination model for real-time discrimination, obtaining a high-dimensional feature vector after data processing, and generating a defect probability heatmap by processing the high-dimensional feature vector; binarizing the defect probability heatmap according to an initial confidence threshold to generate a final defect binary mask, which is then overlaid and displayed on the first B-scan image data in real time to obtain a defect recognition result with pixel-level scene localization; or sequentially inputting the first B-scan image data into the defect discrimination model for real-time discrimination, obtaining a high-dimensional feature vector after data processing, generating a category label and confidence level corresponding to the high-dimensional feature vector, and judging and classifying the first B-scan image data according to the initial confidence threshold to obtain a defect recognition result with image-level scene classification.

[0015] The high-dimensional feature vector output in real-time is decoded to produce a defect probability heatmap. This heatmap is then binarized using an initial confidence threshold. The resulting binary defect mask is overlaid on the first B-scan image data in real-time, accurately delineating the shape, size, and location of defects, resulting in pixel-level defect identification. Alternatively, it can output category labels and their confidence scores. When the confidence score of a category exceeds the initial confidence threshold, an audio-visual alarm is displayed on the interface for rapid screening and early warning. This defect discrimination model can achieve both pixel-level defect segmentation and image-level defect classification according to different scenario requirements, combining accuracy and efficiency, and has a wide range of applications.

[0016] Furthermore, when the defect identification result meets the manual verification triggering condition, the first B-scan image data is processed according to a preset image processing model to obtain second B-scan image data, and the second B-scan image data is displayed to the user. Specifically, when the defect identification result contains a defect and the confidence level is higher than a preset threshold, the image processing model is automatically triggered to start working, and the first B-scan image data corresponding to the triggering time is input into the image processing model; the image processing model performs forward inference and image processing on the first B-scan image data to obtain second B-scan image data, and displays the second B-scan image data to the user so that the user can confirm or correct the defect identification result based on the second B-scan image data.

[0017] In this way, the image processing model processes the B-scan image data and outputs a B-scan image with high signal-to-noise ratio and obvious features to the user. The image processing model serves as an auxiliary verification tool, providing enhanced and clear visual evidence for manual verification of defect identification results, forming a closed-loop detection process with mutual trust between humans and machines, effectively improving the efficiency of human-machine collaboration, and ultimately improving the discrimination accuracy and real-time performance of the entire system.

[0018] Furthermore, the unsupervised learning trains the initial model until the initial model meets the training conditions. Specifically, it optimizes the training by using mean squared error or mean absolute error as the loss function to minimize the reconstruction error, and optimizes the model parameters through backpropagation algorithm so that the output of the initial model converges to the input, thus obtaining the trained initial model.

[0019] In this way, through unsupervised learning training, the output of the initial model converges to the input, allowing the model to learn the feature representation of GPR data, and finally obtain a high-performance initial model. The initial model can achieve feature recognition of a large amount of GPR data without relying on a large number of manually labeled samples.

[0020] Furthermore, the decoder of the defect discrimination model includes a segmentation and localization output layer and a classification image output layer. Specifically, the segmentation and localization output layer includes a segmentation decoder, where the historical B-scan image data is processed and combined with a Sigmoid activation function to output a defect probability heatmap with the same size as the input. The classification image output layer includes a classification head composed of a global pooling layer and several fully connected layers, where the historical B-scan image data is processed and combined with a Softmax activation function to output a multi-class probability distribution vector.

[0021] This approach connects a segmentation decoder after the encoder, employing upsampling and skip connections to ultimately output a defect probability heatmap of the same size as the input through a sigmoid activation function. Alternatively, a classification head consisting of a global pooling layer and several fully connected layers can be connected after the encoder, ultimately outputting a multi-class probability distribution vector through a softmax activation function. Combining these two output layers allows the defect discrimination model to output corresponding defect identification results based on different scenarios, thus improving the generalization ability of the defect discrimination model.

[0022] Another embodiment of the present invention provides a real-time ground-penetrating radar data processing system, including: a defect discrimination module and an image processing module; the defect discrimination module is used to acquire first B-scan image data in real time, input the first B-scan image data into a preset defect discrimination model, automatically and in real time discriminate the first B-scan image data, and output defect identification results; the image processing module is used to perform image processing on the first B-scan image data according to the defect identification results to obtain second B-scan image data, and display the second B-scan image data to the user.

[0023] The defect discrimination module of this invention can output pixel-level segmentation maps to accurately locate defects, or output image-level classification results to achieve rapid screening, depending on actual needs. It combines accuracy and efficiency and has a wide range of applications. The image processing module provides enhanced and clear visualization of the defect recognition results, reducing the burden of manual verification and forming a closed-loop detection process with mutual trust between humans and machines.

[0024] Another embodiment of the present invention provides a data processing apparatus, including: a module for performing the real-time processing method for ground-penetrating radar data as described in this application. Attached Figure Description

[0025] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0026] Figure 1 is a flowchart illustrating one embodiment of the real-time ground-penetrating radar data processing method provided by the present invention; Figure 2 is a structural diagram of model A (autoencoder model) in another embodiment provided by the present invention; Figure 3 is a structural diagram of model B (defect discrimination model) in another embodiment provided by the present invention; Figure 4 is a structural diagram of model C (image processing model) in another embodiment provided by the present invention; Figure 5 is a schematic diagram of the overall architecture and flowchart of another embodiment of the real-time ground-penetrating radar data processing method provided by the present invention; Figure 6 is a module schematic diagram of another embodiment of the real-time ground-penetrating radar data processing system provided by the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0028] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0029] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0030] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0031] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0032] Referring to Figure 1, to solve the problem of real-time processing of ground-penetrating radar data in the prior art, this embodiment of the invention provides a real-time processing method for ground-penetrating radar data, including steps S1 to S2, each step being as follows: S1, acquiring first B-scan image data in real time, inputting the first B-scan image data into a preset defect discrimination model, so that the defect discrimination model processes strong noise interference and electromagnetic wave scattering effects in the first B-scan image data through a bidirectional attention mechanism, and automatically and in real-time discriminates the first B-scan image data, outputting defect identification results, wherein the defect discrimination model is a self- The first encoder in the encoder model is obtained by transfer and training; the autoencoder model is obtained by training based on historical B-scan image data; S2, when the defect identification result meets the manual verification triggering condition, the first B-scan image data is processed according to the preset image processing model to obtain the second B-scan image data, and the second B-scan image data is displayed to the user so that the user can perform visual verification and auxiliary decision-making based on the second B-scan image data, wherein the image processing model is obtained by transfer and training the first encoder in the autoencoder model.

[0033] As an example of an embodiment of the present invention, as shown in Figure 5, a general first encoder for extracting features from ground-penetrating radar data is first pre-trained using unsupervised learning. Then, the first encoder is transferred to two different downstream task models, where supervised fine-tuning is performed to form a defect discrimination model and an image processing model. Finally, in real-time application scenarios, the defect discrimination model and the image processing model work collaboratively to sequentially complete automated defect identification and result verification, forming a closed-loop detection process with human-machine mutual trust. This process effectively reduces the dependence on expensive labeled data and improves the model's generalization ability and practicality for engineering applications.

[0034] In one embodiment, as shown in Figure 2, the first encoder of the autoencoder model is trained based on historical B-scan image data, including steps S201 to S203, each step being as follows: S201, constructing an initial model based on an encoder-decoder architecture, wherein the encoder of the initial model includes a multi-layer two-dimensional convolutional neural network and a bidirectional attention mechanism module, and the decoder of the initial model includes a deconvolution layer or an upsampling layer symmetrical to the encoder.

[0035] The decoder in the autoencoder model is designed to progressively reconstruct an image of the same size as the input by mapping the compressed features output by the encoder.

[0036] S202. Collect historical B-scan image data, and train the initial model in unsupervised learning based on the historical B-scan image data until the initial model meets the training conditions, and output the autoencoder model.

[0037] The original B-scan image data was used to construct a dataset. ,in Indicates the first The dataset contains raw radar data images; at this stage, no manual annotation is required. The raw data in the model undergoes standardized preprocessing, mainly including operations such as mean removal and variance normalization, to accelerate model convergence.

[0038] The training objective of the autoencoder model is to minimize the reconstruction error, that is, to make the output of the autoencoder model converge to the input; the mean squared error or mean absolute error is used as the loss function, and the formula for calculating the mean squared error is as follows: Where N is the input data The total number of pixels.

[0039] After training is completed, the model parameters are optimized by backpropagation algorithm. Only the structure and parameters of the first encoder of the autoencoder model are retained and solidified. At this time, the first encoder has become a feature extractor that can extract strong robust general features of GPR data.

[0040] S203. The first encoder in the autoencoder model extracts the spatial hierarchical features of the historical B-scan image data through a CNN layer, processes the strong noise interference and electromagnetic wave scattering effect in the historical B-scan image data through a bidirectional attention mechanism, and uses a two-dimensional calibration feature response to suppress noise interference and enhance electromagnetic wave scattering features, thereby obtaining compressed features containing long-range dependencies and global context information in the historical B-scan image data, and gradually reconstructs the obtained compressed feature mapping into a B-scan image with the same size as the input.

[0041] The CNN layers are used to extract spatial hierarchical features (such as edges, textures, and shapes) of the image layer by layer. The bidirectional attention mechanism is used to adaptively recalibrate features in both channel and spatial dimensions. Spatial attention addresses the long-range spatial dependency problem of defects (such as hyperbolic reflections) in GPR images. By calculating the correlation strength between any two spatial locations on the feature map, the feature map is reshaped and then a spatial attention map is generated through query and key-value operations. This allows each pixel in the image to be directly associated with all other pixels in the image. Traditional CNNs have limited receptive fields and struggle to capture the association between the vertex of a complete hyperbola and its distant flanks. Channel attention addresses the uneven importance of features across different frequency channels in GPR data, suppressing noise and enhancing effective signals. Defect information in GPR data may be concentrated in certain frequency components (channels), while other channels may be filled with environmental noise or clutter. The channel attention mechanism can adaptively amplify the contribution of key channels and suppress secondary or interfering channels, essentially acting as a self-learning, intelligent frequency filter. The outputs of the spatial and channel attention modules are fused to form the final optimized features. This process enables in-depth feature mining within a single image, focusing on the complete spatial structure of defects, selecting the channel information that best represents defects, and addressing noise and scattering issues in GPR data.

[0042] This embodiment ensures that the encoder can effectively extract spatial hierarchical features such as edges, textures, and shapes of images, while suppressing noise interference and enhancing electromagnetic wave scattering features. It can effectively capture long-range dependencies and global contextual information in GPR data, enhancing feature extraction capabilities and generalization performance in complex noisy environments. This is a key execution step for enabling the first encoder to extract robust and general features from GPR data. Feature extraction and recognition through the encoder-decoder structure and bidirectional attention mechanism have advantages such as not requiring a large amount of manually labeled data, high training efficiency, and good recognition accuracy. The bidirectional attention mechanism introduced in the pre-trained encoder can effectively capture long-range dependencies and global contextual information in GPR data, enhancing the model's feature extraction capabilities and generalization performance in complex noisy environments.

[0043] In one embodiment, as shown in Figure 3, the defect discrimination model is obtained by transferring and training the first encoder in the autoencoder model, including steps S301 to S303, each step being as follows: S301, the encoder of the defect discrimination model transfers the first encoder in the autoencoder model, and the decoder includes a segmentation and localization output layer and a classification image output layer.

[0044] The segmentation and localization output layer is connected to the segmentation decoder (such as the decoding part of a U-Net structure) after the encoder. Through operations such as upsampling and skip connections, it finally outputs a defect probability heatmap with the same size as the input through the Sigmoid activation function. The classification image output layer is connected to the classification head after the encoder. The classification head is usually composed of a global pooling layer and several fully connected layers. Finally, it outputs a multi-class probability distribution vector through the Softmax activation function.

[0045] S302. Collect historical B-scan image data and labeled pixel-level binary segmentation mask images as segmentation supervision data, and optimize the training by using gradient descent combined with loss function BCE Loss or Dice Loss based on the segmentation supervision data to obtain the segmentation task part of the trained defect discrimination model.

[0046] The segmentation supervision data is constructed into a dataset. .in For historical B-scan image data, This is the corresponding labeled pixel-level binary segmentation mask (the pixel value in the defect area is 1, and the pixel value in the background area is 0). For training data, gradient descent is used to optimize model parameters. The loss function chosen is either Binary Cross-Entropy Loss (BCE Loss) or Dice Loss, with the goal of minimizing the difference between the predicted heatmap and the actual segmentation mask. The differences between them.

[0047] S303. Collect historical B-scan image data and corresponding image-level category labels as classification supervision data, and optimize the training by using gradient descent combined with the loss function classification cross-entropy loss based on the classification supervision data to obtain the classification task part of the defect discrimination model after training.

[0048] The classification supervision data is constructed into a dataset. .in For historical B-scan image data, For the corresponding image-level category labels (e.g., 0 represents "no defects", 1 represents "void", 2 represents "water-filled voids", etc.). For the training data, gradient descent is used to optimize the model parameters. The loss function chosen is the categorical cross-entropy loss, which aims to minimize the difference between the predicted probability distribution and the true class label. The differences between them.

[0049] This embodiment trains the defect discrimination model using different supervised data according to different application scenarios. By minimizing the difference between the predicted heatmap output by the defect discrimination model and the real segmentation mask, and minimizing the difference between the predicted probability distribution output by the defect discrimination model and the real category label, the output result of the defect discrimination model converges to the human recognition result, thereby improving the recognition accuracy and recognition efficiency of the defect discrimination model.

[0050] In one embodiment, as shown in FIG4, the first encoder in the autoencoder model is transferred and trained to obtain an image processing model, including steps S401 to S402, each step being as follows: S401, the encoder of the image processing model transfers the first encoder in the autoencoder model, and the decoder configures the image output layer.

[0051] The decoder structure is similar to that of the defect discrimination model, but the ultimate goal of the image processing model is to reconstruct a high-quality processed image, rather than the segmentation map of the defect discrimination model.

[0052] S402. Collect historical B-scan image data and manually optimized historical B-scan image data as image processing supervision data for training, and optimize the training using the loss function mean square error or perceptual loss based on the image processing supervision data to obtain the trained image processing model.

[0053] Specifically, the image processing supervision data is constructed into a supervised dataset. ,in For historical B-scan image data, For signal processing experts The resulting image is clearer and easier for human interpretation after a series of advanced processing steps (such as precise filtering, gain control, background removal, etc.).

[0054] by For training data, mean squared error (MSE) or perceptual loss is used as the loss function to train the image processing model so that the output of the model converges to the result after manual processing. .

[0055] This embodiment trains the image processing model through unsupervised learning, minimizing the difference between the image output by the model and the result after manual processing. This makes the image processing model's image processing converge to the result of manual processing, thereby improving the image processing capability and efficiency of the model.

[0056] In one embodiment, the defect discrimination model automatically and in real-time discriminates the first B-scan image data, and the image processing model is used to assist manual verification, including steps S501 to S504. The specific steps are as follows: S501: The first B-scan image data is sequentially input into the defect discrimination model for real-time discrimination. After data processing, a high-dimensional feature vector is obtained. The high-dimensional feature vector is then processed to generate a defect probability heatmap. The defect probability heatmap is binarized according to an initial confidence threshold to generate a final binary defect mask, which is then overlaid and displayed in real-time on the first B-scan image data to obtain a defect recognition result with pixel-level scene localization. Alternatively, the first B-scan image data is sequentially input into the defect discrimination model for real-time discrimination. After data processing, a high-dimensional feature vector is obtained, and the category label and confidence level corresponding to the high-dimensional feature vector are generated. The first B-scan image data is judged and classified according to an initial confidence threshold to obtain a defect recognition result with image-level scene classification.

[0057] During the detection process, the ground-penetrating radar (GPR) equipment generates B-scan image data streams in real time. Each newly generated image frame undergoes the same standardized preprocessing as during the training phase to obtain the first B-scan image data. The standardized first B-scan image data is then sequentially input into the defect discrimination model for real-time discrimination. After operations such as convolution, pooling, and bidirectional attention mechanisms in the encoder of the defect discrimination model, a high-dimensional feature vector is output.

[0058] The high-dimensional feature vector is decoded and processed through upsampling / deconvolution, fusion of encoder features, pointwise convolution, and sigmoid activation to output a defect probability heatmap. The system sets a confidence threshold (e.g., 0.5) for binarization, generating a final defect binary mask, which can be overlaid on the original B-scan image in real time, accurately outlining the shape, size, and location of the defect. Alternatively, the high-dimensional feature vector is decoded and processed through pooling, fully connected layers, and sigmoid activation to output a category label and its confidence level. Combined with the set initial confidence threshold (e.g., 0.8), when the confidence level of a certain category exceeds the threshold, the interface displays "Detected XX defect, confidence level YY%" and triggers an audible and visual alarm. This mode is suitable for rapid screening and early warning.

[0059] S502. When the defect identification result contains a defect and the confidence level is higher than a preset threshold, the image processing model is automatically triggered to start working, and the first B-scan image data corresponding to the trigger time is input into the image processing model.

[0060] S503. The image processing model performs forward reasoning and image processing on the first B-scan image data to obtain the second B-scan image data, and displays the second B-scan image data to the user so that the user can confirm or correct the defect identification result based on the second B-scan image data.

[0061] Specifically, the first B-scan image data corresponding to the trigger time is input into the image processing model. The image processing model performs forward inference, and the first B-scan image data is processed by the encoder and decoder of the image processing model to output enhanced, high-definition second B-scan image data. The second B-scan image data simulates the effect after expert processing, with a high signal-to-noise ratio and obvious features. The second B-scan image data is provided to the operator (e.g., displayed on another monitoring screen or in a pop-up window) as a visual verification and auxiliary decision-making basis for the automatic judgment result of the defect discrimination model. The operator can quickly confirm or correct the automatic judgment result based on this.

[0062] The defect discrimination model in this embodiment can output pixel-level segmentation maps to accurately locate defects or output image-level classification results to achieve rapid screening, depending on actual needs. It combines accuracy and efficiency and has a wide range of applications. The image processing model serves as an auxiliary verification tool, providing enhanced and clear visualization evidence for the automatic discrimination results, forming a reliable closed loop of "machine initial judgment and manual confirmation", which effectively improves the credibility and practicality of the overall process.

[0063] As shown in Figure 6, which illustrates real-time processing of ground-penetrating radar data, a corresponding system embodiment is provided based on the above method embodiment. This embodiment of the invention provides a real-time ground-penetrating radar data processing system, comprising: a defect discrimination module 601 and an image processing module 602. The defect discrimination module 601 is used to acquire first B-scan image data in real time, input the first B-scan image data into a preset defect discrimination model, perform automated real-time discrimination on the first B-scan image data, and output defect identification results. The image processing module 602 is used to perform image processing on the first B-scan image data according to the defect identification results to obtain second B-scan image data, and display the second B-scan image data to the user.

[0064] The defect discrimination module of this invention can output pixel-level segmentation maps to accurately locate defects, or output image-level classification results to achieve rapid screening, depending on actual needs. It combines accuracy and efficiency and has a wide range of applications. The image processing module provides enhanced and clear visualization of the defect recognition results, reducing the burden of manual verification and forming a closed-loop detection process with mutual trust between humans and machines.

[0065] It is understood that the above system embodiments correspond to the method embodiments of the present invention, and can implement the ground-penetrating radar data real-time processing method provided by any of the above method embodiments of the present invention.

[0066] It should be noted that the system embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0067] For ease of description and brevity, the system embodiments of the present invention include all the implementation methods described in the above embodiments of the real-time processing method based on ground penetrating radar data, and will not be repeated here.

[0068] Based on the above-described method embodiments, another embodiment of the present invention provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the real-time ground-penetrating radar data processing method described in any of the above-described method embodiments of the present invention.

[0069] The modules / units integrated in the device / terminal equipment, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0070] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for real-time processing of ground-penetrating radar data, characterized in that, include: The system acquires first B-scan image data in real time and inputs it into a preset defect discrimination model. The defect discrimination model uses a bidirectional attention mechanism to process strong noise interference and electromagnetic wave scattering effects in the first B-scan image data, and performs automated real-time discrimination, outputting defect identification results. The defect discrimination model is obtained by transferring and training the first encoder in an autoencoder model, which is trained based on historical B-scan image data. When the defect identification result meets the manual verification trigger condition, the system processes the first B-scan image data according to a preset image processing model to obtain second B-scan image data. This second B-scan image data is then displayed to the user for visual verification and decision support based on the second B-scan image data. The image processing model is obtained by transferring and training the first encoder in an autoencoder model.

2. The real-time processing method for ground-penetrating radar data as described in claim 1, characterized in that, The autoencoder model is trained based on historical B-scan image data, including: constructing an initial model based on an encoder-decoder architecture, wherein the encoder of the initial model includes a multi-layer two-dimensional convolutional neural network and a bidirectional attention mechanism module, and the decoder of the initial model includes a deconvolution layer or an upsampling layer symmetrical to the encoder; acquiring historical B-scan image data, and training the initial model in unsupervised learning based on the historical B-scan image data until the initial model meets the training conditions, and outputting the autoencoder model; the first encoder in the autoencoder model extracts the spatial hierarchical features of the historical B-scan image data through a CNN layer, processes the strong noise interference and electromagnetic wave scattering effect in the historical B-scan image data through a bidirectional attention mechanism, and uses a two-dimensional calibration feature response to suppress noise interference and enhance electromagnetic wave scattering features, thereby obtaining compressed features containing long-range dependencies and global contextual information in the historical B-scan image data, and gradually reconstructing the obtained compressed feature mapping into a B-scan image with the same size as the input.

3. The real-time processing method for ground-penetrating radar data as described in claim 1, characterized in that, The defect discrimination model is obtained by transferring and training the first encoder in the autoencoder model. Specifically, the encoder of the defect discrimination model transfers the first encoder in the autoencoder model, and the decoder includes a segmentation and localization output layer and a classification image output layer. Historical B-scan image data and labeled pixel-level binary segmentation masks are collected as segmentation supervision data, and the segmentation task part of the defect discrimination model is optimized and trained by using gradient descent combined with loss function BCE Loss or Dice Loss based on the segmentation supervision data to obtain the segmentation task part of the trained defect discrimination model. Historical B-scan image data and corresponding image-level category labels are collected as classification supervision data. Based on the classification supervision data, gradient descent combined with the classification cross-entropy loss function is used for optimization training to obtain the classification task part of the defect discrimination model after training.

4. The real-time processing method for ground-penetrating radar data as described in claim 2, characterized in that, The image processing model is obtained by transferring and training the encoder, specifically: the encoder of the image processing model transfers the first encoder in the autoencoder model, and the decoder configures the image output layer; historical B-scan image data and manually optimized historical B-scan image data are collected as image processing supervision data for training, and the image processing supervision data is used to optimize the training by means of mean square error or perceptual loss to obtain the trained image processing model.

5. The real-time processing method for ground-penetrating radar data as described in claim 1, characterized in that, The automated real-time discrimination of the first B-scan image data and the output of defect recognition results specifically involves: inputting the first B-scan image data sequentially into the defect discrimination model for real-time discrimination; obtaining a high-dimensional feature vector after data processing; generating a defect probability heatmap by processing the high-dimensional feature vector; binarizing the defect probability heatmap according to an initial confidence threshold to generate a final defect binary mask; and then overlaying and displaying it on the first B-scan image data in real time to obtain a defect recognition result with pixel-level scene localization; or inputting the first B-scan image data sequentially into the defect discrimination model for real-time discrimination; obtaining a high-dimensional feature vector after data processing; generating a category label and confidence level corresponding to the high-dimensional feature vector; and judging and classifying the first B-scan image data according to an initial confidence threshold to obtain a defect recognition result with image-level scene classification.

6. The real-time processing method for ground-penetrating radar data as described in claim 1, characterized in that, When the defect identification result meets the manual verification trigger condition, the first B-scan image data is processed according to a preset image processing model to obtain second B-scan image data, and the second B-scan image data is displayed to the user. Specifically, when the defect identification result contains a defect and the confidence level is higher than a preset threshold, the image processing model is automatically triggered to start working, and the first B-scan image data corresponding to the trigger time is input into the image processing model; the image processing model performs forward inference and image processing on the first B-scan image data to obtain second B-scan image data, and displays the second B-scan image data to the user so that the user can confirm or correct the defect identification result based on the second B-scan image data.

7. The real-time processing method for ground-penetrating radar data as described in claim 2, characterized in that, The unsupervised learning process trains the initial model until it meets the training conditions. Specifically, it optimizes the training by using mean squared error or mean absolute error as the loss function to minimize the reconstruction error, and optimizes the model parameters through backpropagation to make the output of the initial model converge to the input, thus obtaining the trained initial model.

8. The real-time processing method for ground-penetrating radar data as described in claim 3, characterized in that, The decoder of the defect discrimination model includes a segmentation and localization output layer and a classification image output layer. Specifically, the segmentation and localization output layer includes a segmentation decoder, where the historical B-scan image data is processed and combined with a Sigmoid activation function to output a defect probability heatmap with the same size as the input. The classification image output layer includes a classification head composed of a global pooling layer and several fully connected layers, where the historical B-scan image data is processed and combined with a Softmax activation function to output a multi-class probability distribution vector.

9. A real-time data processing system for ground-penetrating radar, characterized in that, include: Defect detection module, image processing module; The defect discrimination module is used to acquire first B-scan image data in real time, input the first B-scan image data into a preset defect discrimination model, perform automatic real-time discrimination on the first B-scan image data, and output defect recognition results; the image processing module is used to perform image processing on the first B-scan image data according to the defect recognition results to obtain second B-scan image data, and display the second B-scan image data to the user.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program or instructions, which, when executed by a data processing device, implement the real-time processing method for ground-penetrating radar data as described in any one of claims 1 to 8.

Images