Ground penetrating radar deep learning data denoising method based on generative adversarial network
A deep learning data denoising model for ground-penetrating radar (GPR) constructed using generative adversarial networks (GANs) solves the problem of poor image quality in GPR, achieving clearer image presentation and more efficient detection capabilities.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTH CHINA UNIV OF WATER RESOURCES & ELECTRIC POWER
- Filing Date
- 2026-02-12
- Publication Date
- 2026-06-19
AI Technical Summary
Ground penetrating radar images suffer from poor quality due to the complex underground environment and various noise sources, making it difficult to present sufficient effective information and affecting interpretation and judgment.
A data denoising model is constructed using a generative adversarial network (GAN). Training and test sets are established using the finite-difference temporal method (FDTD). The network model combines Transformer and CNN modules to perform image denoising. Key details are preserved by using feature pyramid fusion and residual connection mechanisms.
It improved the quality of radar images, enhanced the interpretation capabilities of detection personnel, and improved the intelligence and efficiency of detection.
Smart Images
Figure CN122243791A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to geological testing data processing technology, and in particular to a method for denoising ground-penetrating radar deep learning data based on generative adversarial networks. Background Technology
[0002] Highways built in permafrost regions frequently suffer from roadbed and pavement defects. The combined effects of heat absorption by the road surface and climate warming, leading to rising permafrost temperatures, decreasing permafrost upper limits, and reduced bearing capacity of permafrost foundations, result in continuous uneven thaw settlement, long-distance longitudinal cracks, and slope slippage, severely impacting highway service performance. Ground-penetrating radar (GPR) is an important electromagnetic non-destructive testing technology that can accurately and quickly image shallow surfaces, acquiring spatial and geometric information of buried objects. It plays a crucial role in geophysical exploration applications such as pipeline surveys and detection of hidden road defects. However, due to the complex and unknown underground environment and various noise sources affecting the interpretation of GPR profile images, the acquired radar images often fail to present sufficient effective information, obscuring or obscuring important details, resulting in poor image quality and hindering interpretation and judgment by researchers and exploration personnel.
[0003] Therefore, to address the noise reduction problem of ground-penetrating radar data, this invention uses deep learning methods, based on generative adversarial networks, to establish a data noise reduction network model, generating clearer and more recognizable radar images. Summary of the Invention
[0004] The purpose of this invention is to provide a deep learning data denoising method for ground penetrating radar based on generative adversarial networks. By constructing a specific denoising model to denoise ground penetrating radar image data, this method solves the problem that existing radar data images are of poor quality and fail to present sufficient effective information, thus affecting the interpretation and judgment of researchers or exploration personnel.
[0005] This invention is implemented as follows: a method for denoising ground-penetrating radar deep learning data based on generative adversarial networks, the construction method comprising the following steps:
[0006] S1. The finite-difference time-domain (FDTD) method is used to preprocess the raw geological data to obtain the training set and the test set.
[0007] S2. Construct a network model based on Generative Adversarial Networks (GANs) that includes Transformer and CNN modules (preferably, the network model in this part includes a generator network and a discriminator network. The encoder in the generator network adopts a four-level downsampling structure, with each level consisting of a Transformer-CNN coupled module and a convolutional layer alternately. The decoder converts low-resolution latent features into high-resolution semantic features through four upsampling operations. To compensate for information loss during downsampling, the network adopts a feature pyramid fusion strategy, which achieves information complementarity through cross-layer feature concatenation, and then completes channel recombination through 1×1 convolution. The final output stage adopts a residual connection mechanism to ensure that key detail features are preserved. The discriminator network consists of four convolutional layers and one fully connected layer). Then, import the training set into the network model for training to obtain the trained model.
[0008] S3. The training model is continuously tested and optimized using a test set until the simulated denoised image data generated by the generator can be discriminated by the discriminator. This yields a denoising model for denoising ground-penetrating radar data.
[0009] A further technical solution of the present invention is: step S1 specifically includes:
[0010] S101. By dynamically adjusting at least one variable among the background medium electrical parameters, stratigraphic spatial distribution, target geometric characteristics, and antenna center frequency (e.g., the relative permittivity of the background medium is adjusted from 3 to 10; the target pipe diameter is adjusted from DN300 to DN400; the side length of the rectangular culvert is adjusted from 10 to 40 cm; the number of strata is adjusted from 1 to 3; and the antenna center frequency is adjusted from 100 MHz, 200 MHz, 400 MHz, 1000 MHz, etc.), a geological data sample library based on random parameters is constructed.
[0011] S102. Use the FDTD algorithm to perform forward modeling on each sample data in the above geological data sample library to obtain a preprocessed dataset consisting of noisy and noiseless B-scan images; randomly allocate the preprocessed dataset to obtain a training set and a test set.
[0012] S103. To improve the model's generalization ability, random spatial cropping preprocessing is performed on the original GPR images, uniformly adjusting them to a standard size with the same number of pixels. In other words, step S103 unifies the data scale. The original GPR images are the training and test set data from step S102. Cropping is random; the cropped pixels only need to be uniform, for example, they can be uniformly adjusted to any one of the standard sizes of 64×64 pixels, 128×128 pixels, or 256×256 pixels.
[0013] A further technical solution of the present invention is: in step S101, the geometric features of the target body include at least one of the following: the number of target bodies, the size of the target body, and the burial depth of the target body.
[0014] In step S102, the number of noisy and noiseless B-scan images in the preprocessed dataset is no less than 1000 sets. The ratio of the number of noisy and noiseless B-scan image sets in the training set and the test set is 9~8:1~2.
[0015] A further technical solution of the present invention is: step S2 specifically includes:
[0016] S201. The input training set image is segmented (randomly cropped) into overlapping block regions. Then, shallow features are extracted through convolutional layers, and deep semantic information is gradually recovered using an encoder-decoder structure. Finally, in the feature reconstruction stage, channel dimension compression is achieved through cross-layer feature fusion and 1×1 convolution operation, while preserving key detail information.
[0017] S202. Input the data in the training set into the generator, and the decoder in the generator converts the input data into simulated image data. Then, the discriminator judges and scores the generated simulated image data.
[0018] It should be noted that in step S202, the score refers to the degree of similarity between the model image generated by the generator and the original image. For example, it outputs a single scalar value between 0 and 1. The closer the value is to 1, the closer the discriminator judges the simulated image to the real image.
[0019] S203. First, fix the generator parameters and train the discriminator parameters. Then, fix the discriminator parameters and train the generator parameters again. Repeat this alternating process until a generator and discriminator parameter combination that meets the scoring requirements is obtained (i.e., alternately fix and train the generator and discriminator parameters, iteratively optimize, until the optimal generator and discriminator parameter combination is obtained. For example, first fix the generator parameters (such as the weights and biases of each network layer), train the discriminator parameters separately, until the optimal discriminator parameters under the fixed generator parameters are obtained; then fix the obtained discriminator parameters, and train the generator parameters in reverse, until the optimal generator parameters under the fixed discriminator parameters are obtained. Alternately fix and train the generator and discriminator parameters in this way, iteratively optimize, until the optimal generator and discriminator parameter combination is obtained).
[0020] A further technical solution of the present invention is: in step S202, a composite loss function is used to train the generator network; the composite loss function consists of mean square error, absolute error and adversarial loss.
[0021] A further technical solution of the present invention is: step S3 specifically includes:
[0022] S301. Adjust the model hyperparameters (e.g., adjust the initial learning rate, number of iterations, and number of images per training session (Batchsize) for model training) and continuously test and optimize the trained model using a test set. Generate a batch of simulated denoised image data every 5 to 15 iterations.
[0023] S302. Compare the generated simulated denoised image data with the original real data, and verify the network model performance by comprehensively considering the quality of the generated data and the speed of model convergence.
[0024] A further technical solution of the present invention is: in step S301, the model feature parameters include the initial learning rate, the number of iterations, and the number of images (Batchsize); the optimization is performed using the Adam optimizer to minimize the generator and maximize the discriminator.
[0025] A further technical solution of the present invention is: the initial learning rate of the generator and the discriminator in the training model is set to 0.0001~0.0002, the number of iterations is set to 100~200 times, the number of images trained each time is 8~16, and the model parameters are saved once after every 5~15 iterations.
[0026] A further technical solution of the present invention is: in step S301, the main basis for adjusting the model hyperparameters is the model loss function and the model convergence time when different parameters are set, and the hyperparameter settings corresponding to the optimal model performance are saved.
[0027] A further technical solution of the present invention is: the phrase "until the simulated denoised image data generated by the generator can be identified by the discriminator" means that the pass rate of the simulated denoised image data generated by the generator being identified by the discriminator as real denoised image data is not less than 90%.
[0028] In step S1 of this invention, the Finite-Difference Time-Domain (FDTD) method is used as the core simulation tool to establish random geological samples and construct noisy-noise image pairs as training samples for model training. Specifically, this includes: 1.1) Generating a geological model library based on random parameters: By dynamically adjusting key variables such as background medium electrical parameters, stratigraphic spatial distribution, target geometric features (e.g., quantity, size, burial depth), and antenna center frequency, diverse underground structure samples are constructed. 1.2) Using the FDTD algorithm, forward modeling is performed on each sample, and pairs of noisy and noiseless B-scan images are output in batches. The constructed dataset contains at least 1000 sets of B-scan samples, of which 80-90% are divided into training sets (e.g., 800-900 sets), and the remaining 10-20% are used as test sets (e.g., 100-200 sets). 1.3) To improve the model's generalization ability, random spatial cropping preprocessing is performed on the original GPR images, for example, uniformly adjusting them to a standard size of 256×256 pixels.
[0029] In step S2 of this invention, a network model based on Generative Adversarial Network (GAN) is constructed, integrating the global modeling advantages of Transformer and the local feature extraction characteristics of CNN. Transfer learning is used to initialize the network model, and the training set is imported into the model for training. Specifically, this includes: 2.1) Segmenting the input image into overlapping block regions to effectively alleviate the common boundary artifact problem when Transformer processes images; secondly, after extracting shallow features through convolutional layers, an encoder-decoder structure is used to gradually recover deep semantic information; finally, in the feature reconstruction stage, channel dimension compression is achieved through cross-layer feature fusion and 1×1 convolution operations while preserving key detail information; furthermore, the network model described in this part includes a generator network and a discriminator network. The generator network integrates a Transformer-CNN coupling module. The input noisy image first undergoes overlapping block processing, and then initial features are extracted through convolutional layers; the encoder part adopts a four-level downsampling structure, with each level consisting of a Transformer-CNN... The network consists of alternating coupling modules and convolutional layers, gradually generating feature maps containing multi-scale dependencies. The decoder converts low-resolution latent features into high-resolution semantic features through four upsampling operations. To compensate for information loss during downsampling, the network employs a feature pyramid fusion strategy, achieving information complementarity through cross-layer feature concatenation, followed by channel recombination via 1×1 convolution. The final output stage uses a residual connection mechanism to perform difference fusion between the generated image and the original noisy image, ensuring that key detail features are preserved. For example, its discriminator network consists of four convolutional layers and one fully connected layer. The generator converts random data into virtual images, the discriminator judges the virtual images based on real image data, and finally optimizes the network parameters of the generator G and the discriminator D based on the discrimination results. 2.2) After random data is input into the generator, the decoder in the generator converts the random data into image format data. The discriminator judges the generated data (generally to determine whether it is real data) and obtains a score. Further, the network model described in this part uses a composite loss function to train the generator network, and achieves generation quality optimization by integrating pixel-level reconstruction error and adversarial constraints. Specifically, the generator's total loss function consists of three parts: mean squared error, absolute error, and adversarial loss. 2.3) First, fix the parameters of the generator G and train the parameters of the discriminator D. Then fix the parameters of the discriminator D and train the parameters of the generator G again. This process is repeated alternately.
[0030] In step S3 of this invention, the model is tested and optimized using the training set. Based on the training effect and performance of the network model, parameters are adjusted, and the generator G and discriminator D are continuously tuned. Once the global optimum is reached, the model is saved. Specifically, this includes generating a batch of denoised radar image data every few iterations (e.g., 10 times) to observe the model training progress and effect. The trained model is used to generate denoised radar image data, and the generated data is compared with the original real data. The quality of the generated data and the model convergence speed are considered comprehensively to verify the network performance. Further, the initial learning rate of the generator and discriminator of the network model is set to 0.0001~0.0002, the number of iterations is set to 100~200, the number of images trained each time (Batchsize) is 10~30 (e.g., 16), and the model parameters are saved every 5~10 iterations. Further, the optimization algorithm uses the Adam optimizer, and the main parameters adjusted are the initial learning rate, the number of iterations, and the batch size. Furthermore, the main basis for adjusting the parameters is the model loss function when setting different parameters, and the time taken for the model to converge. Furthermore, the essence of the generative adversarial network training process is to continuously adjust and optimize the parameters of the generator G and the discriminator D. The optimization method is to minimize the generator and maximize the discriminator, with the global optimum being that the denoised image generated by the generator can fool the discriminator, ultimately achieving a highly realistic effect.
[0031] The beneficial effects of this invention are as follows: This invention provides a method for constructing a deep learning data denoising model for ground-penetrating radar based on generative adversarial networks (GANs). By establishing an FDTD simulation dataset of noisy-noise pairs, random samples are generated for actual detection scenarios, resulting in higher quality radar data. The simulation dataset can achieve accurate matching of noisy-noise pairs, providing ideal conditions for the network to learn noise distribution characteristics; the parameter randomization generation mechanism ensures data diversity, which helps improve the model's generalization ability; the strict division between the test set and the training set follows machine learning standards and specifications, laying the foundation for objectively evaluating model performance.
[0032] Furthermore, this invention utilizes a Transformer-CNN hybrid module to optimize the generator network and construct a network model. Through a dual-path feature fusion mechanism, it efficiently captures fundamental spatial features using depthwise separable convolutions, while simultaneously resolving semantic relationships between feature channels through a channel attention mechanism. This also alleviates the memory consumption issues associated with traditional global attention computation. While ensuring high-quality data denoising, it reduces the time cost of data generation, which is beneficial for improving the intelligence level, detection capabilities, and efficiency of the engineering inspection industry. Attached Figure Description
[0033] Figure 1This is a flowchart of the deep learning data denoising method for ground-penetrating radar based on generative adversarial networks according to the present invention.
[0034] Figure 2 This is a generator network structure diagram of the deep learning data denoising method for ground penetrating radar based on generative adversarial networks according to the present invention. Detailed Implementation
[0035] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention.
[0036] Example 1:
[0037] Figure 1 and Figure 2 This paper presents a method for denoising ground-penetrating radar deep learning data based on generative adversarial networks. The method includes the following steps:
[0038] S1. The raw geological data is preprocessed using the Finite-Difference Time-Domain (FDTD) method to obtain training and test sets. Specifically, FDTD is used as the core simulation tool to establish random geological samples and construct noisy-noise-free image pairs as training samples for model training. The specific process includes:
[0039] S101. By dynamically adjusting at least one variable among the background medium electrical parameters, stratigraphic spatial distribution, target geometric features, and antenna center frequency, a geological data sample library based on random parameters is constructed: that is, a geological model library based on random parameters is constructed: by dynamically adjusting key variables such as background medium electrical parameters, stratigraphic spatial distribution, target geometric features (including quantity, size, and burial depth), and antenna center frequency, a diverse underground structure sample is constructed.
[0040] S102. Using the FDTD algorithm, perform forward modeling on each sample in the aforementioned geological data sample database to obtain a preprocessed dataset consisting of noisy and noise-free B-scan images. Randomly allocate the preprocessed dataset to obtain a training set and a test set: that is, use the FDTD algorithm to perform forward modeling on each sample, and output pairs of noisy and noise-free B-scan images in batches. The constructed dataset contains 1000 sets of B-scan samples, of which 90% are divided into a training set (900 sets), and the remaining 10% are used as a test set (100 sets).
[0041] S103. To improve the generalization ability of the model, random spatial cropping preprocessing is performed on the original GPR image, and it is uniformly adjusted to the standard size of 256×256 pixels.
[0042] Specifically, the randomization strategy for sample parameters covers the following dimensions: the conductivity and dielectric constant of the background medium are dynamically selected within typical geological ranges; the stratigraphic interface adopts a random undulation design, with the undulation amplitude controlled within the range of 0.05~0.5m; the number of target objects is set to 1~3, and their geometric dimensions follow a uniform distribution (length, width, and height range of 0.2~1m); the spatial position of the target objects is achieved through random arrangement of three-dimensional coordinates; the antenna center frequency is randomly selected within 100MHz, 200MHz, 400MHz, and 1000MHz. This parameter configuration scheme can effectively cover common geological exploration scenarios and provide representative data support for network training.
[0043] S2. Construct a network model based on Generative Adversarial Networks (GANs) that includes Transformer and CNN modules, and then import the training set into the network model for training to obtain the trained model: that is, construct a network model based on Generative Adversarial Networks (GANs), combine the global modeling advantages of Transformer with the local feature extraction characteristics of CNN, initialize the network model through transfer learning, import the training set into the model, and train the model; the specific process includes:
[0044] S201. The input image is segmented into overlapping block regions. Then, shallow features are extracted through convolutional layers, and an encoder-decoder structure is used to gradually recover deep semantic information. Finally, in the feature reconstruction stage, channel dimension compression is achieved through cross-layer feature fusion and 1×1 convolution operations, while preserving key detail information. This segmentation of the input image into overlapping block regions effectively alleviates the common boundary artifact problem when Transformer processes images. The shallow features are extracted through convolutional layers, and an encoder-decoder structure is used to gradually recover deep semantic information. Finally, in the feature reconstruction stage, channel dimension compression is achieved through cross-layer feature fusion and 1×1 convolution operations, while preserving key detail information.
[0045] S202. Input the data in the training set into the generator, and the decoder in the generator converts the input data into simulated image data. Then, the discriminator judges and scores the generated simulated image data: that is, after random data is input into the generator, the decoder in the generator converts the random data into image format data, and the discriminator judges the generated data (determines whether it is real data) and obtains a score.
[0046] S203. First, fix the generator parameters and train the discriminator parameters. Then, fix the discriminator parameters and train the generator parameters again. Repeat this alternating process until a generator and discriminator parameter combination that meets the scoring requirements is obtained (i.e., alternately fix and train the generator and discriminator parameters, iteratively optimize, until the optimal generator and discriminator parameter combination is obtained. For example, first fix the generator parameters (such as the weights and biases of each network layer), train the discriminator parameters separately, until the optimal discriminator parameters under the fixed generator parameters are obtained; then fix the obtained discriminator parameters, and train the generator parameters in reverse, until the optimal generator parameters under the fixed discriminator parameters are obtained. Alternately fix and train the generator and discriminator parameters in this way, iteratively optimize, until the optimal generator and discriminator parameter combination is obtained).
[0047] Specifically, the network model in S201 includes a generator network and a discriminator network. The generator network integrates a Transformer-CNN coupled module. The input noisy image is first processed through overlapping blocks, and then initial features are extracted through convolutional layers. The encoder employs a four-level downsampling structure, with each level consisting of alternating Transformer-CNN coupled modules and convolutional layers, progressively generating feature maps containing multi-scale dependencies. The decoder converts low-resolution latent features into high-resolution semantic features through four upsampling operations. To compensate for information loss during downsampling, the network uses a feature pyramid fusion strategy, achieving information complementarity through cross-layer feature concatenation, followed by channel reconstruction via 1×1 convolution. The final output stage uses a residual connection mechanism to perform difference fusion between the generated image and the original noisy image, ensuring that key detail features are preserved. The discriminator network consists of four convolutional layers and one fully connected layer. The generator converts random data into virtual images, the discriminator judges the virtual images based on real image data, and finally optimizes the network parameters of the generator G and discriminator D based on the judgment results.
[0048] Specifically, the network model in S202 uses a composite loss function to train the generator network, achieving optimization of generation quality by integrating pixel-level reconstruction errors and adversarial constraints. Specifically, the generator's total loss function consists of three parts: mean squared error, absolute error, and adversarial loss.
[0049] Specifically, in the Transformer-CNN module of the generative network, during the initial feature transformation stage, the convolutional branch and the attention branch perform three parallel 1×1 convolution operations to generate intermediate features. The convolutional branch linearly projects the input features using local kernel weights, achieving spatial feature reorganization. The attention branch aggregates contextual information through convolution operations, generating feature maps corresponding to the query, key, and value projections in the self-attention mechanism, and expanding the feature dimension to a preset depth. In the feature fusion stage, the convolutional branch uses a fully connected transformation along the channel dimension to adjust the number of feature channels, matching it to the output dimension of the previous stage. Then, it integrates spatial information from different directions through feature shift operations to generate a shifted feature map. Finally, it uses depthwise convolution to align the output channel dimension of the self-attention path, completing the feature output of the convolutional branch. Simultaneously, the attention branch groups the intermediate feature maps, learns the attention weights of each group in parallel, and generates transposed attention features with linear complexity. The attention weights are fused with the original features through a residual connection mechanism, enhancing feature representation while preventing model overfitting.
[0050] Specifically, the discriminant network's role is to distinguish between real and fake generated denoised images, and to update the parameters of both the generator and discriminant networks based on the discrimination results. The discriminant network consists of four convolutional layers and one fully connected layer (FC). Similar to the generator network, a batch normalization (BN) layer is used after the convolutional layers, but the activation functions used are Leaky ReLU and Sigmoid. Furthermore, a dropout layer is introduced to prevent overfitting in the discriminant network.
[0051] Specifically, the discriminant network has 5 layers, and the specific implementation steps are as follows: (1) Convolutional layer 1 (Conv_1): The convolutional kernel size is 5×5, the number of convolutional kernels is 64, the convolutional stride is 2, the image size is 256×256×3, and the output size after convolution is 128×128×64. After convolution, the Leaky ReLU function is used for activation, and a dropout layer is added to output to convolutional layer 2. (2) Convolutional layer 2 (Conv_2): The convolutional kernel size is 5×5, the number of convolutional kernels is 128, the convolutional stride is 2, and the output image size after convolution is 65×65×128. After convolution, batch normalization is performed, and then Leaky ReLU and dropout are applied to output to convolutional layer 3. (3) Convolutional layer 3 (Conv_3): The kernel size is 5×5, the number of kernels is 256, and the stride is 2. After convolution and zero-padding (ZP), the output size is 33×33×256. Then, batch normalization is performed, and Leaky ReLU is activated and a dropout layer is added before outputting to convolutional layer 4. (4) Convolutional layer 4 (Conv_4): The kernel size is 5×5, the number of kernels is 512, and the stride is 1. After convolution, the output image size is 33×33×512. After convolution, batch normalization is performed, and Leaky ReLU is activated and a dropout layer is added before outputting to the fully connected layer. (5) Fully connected layer (FC): The input size is 33×33×512. After activation by the Sigmoid function, the output is 0 or 1, indicating whether the judgment result is false or true.
[0052] Specifically, the generator's total loss function consists of three parts: mean squared error. Absolute error and combat losses See the formula :
[0053] .
[0054] In the formula, and Both methods are used to quantify the pixel-level differences between the generated image and the real label. The former enhances the influence of outliers through square operations and is more sensitive to high-frequency details; the latter uses a linear error metric, which alleviates overfitting while preserving the overall structure. The synergistic effect of the two can significantly improve the pixel-level fidelity of the image. Among them:
[0055] .
[0056] The discriminator network is trained using the binary cross-entropy loss function, which is shown in the equation. :
[0057] .
[0058] In the formula, and These are the scores given by the discriminator network to the generated data and the labeled data, respectively. It is 0.005; It is 0.1.
[0059] S3. Continuously test and optimize the trained model using a test set until the simulated denoised image data generated by the generator can be discriminated by the discriminator. This yields a denoising model for denoising ground-penetrating radar data. Model testing and optimization are performed, and parameters are adjusted based on the training effect and performance of the network model. The generator G and discriminator D are continuously tuned. The specific process includes:
[0060] S301. Adjust the model hyperparameters and continuously test and optimize the trained model using a test set. Generate a batch of simulated denoised image data multiple times per iteration: for example, generate a batch of denoised radar image data every 10 iterations to observe the model training progress and effect.
[0061] S302. Compare the generated simulated denoised image data with the original real data, comprehensively consider the quality of the generated data and the model convergence speed, verify the network model performance until the global optimum is reached, and save the model. That is, use the trained model to generate denoised radar image data, compare the generated data with the original real data, comprehensively consider the quality of the generated data and the model convergence speed, and verify the network performance.
[0062] Specifically, the initial learning rate of both the generator and discriminator in the network model is set to 0.0002, the number of iterations is set to 200, the batch size for each training iteration is 16, and the model parameters are saved every 10 iterations. The optimization algorithm uses the Adam optimizer, and the hyperparameters adjusted are the initial learning rate, the number of iterations, and the batch size. The main basis for adjusting the model hyperparameters is the model loss function with different parameter settings, the model convergence time, and the hyperparameter settings corresponding to the optimal model performance are saved. The essence of the generative adversarial network training process is to continuously adjust and optimize the parameters of the generator G and the discriminator D. The optimization method is to minimize the generator and maximize the discriminator. The global optimum is that the denoised image generated by the generator can fool the discriminator, ultimately achieving a realistic effect. Specifically, the quality of the generated data and the algorithm running time should be statistically analyzed to determine whether the data denoising requirements are met, in order to ensure the practical application capability of this invention. If the expected requirements are not met, errors should be statistically analyzed and summarized, and the model should be retrained and optimized until all indicators meet the expected requirements.
[0063] This invention establishes an FDTD-simulated noisy-noise image pair dataset and generates random samples for practical detection scenarios, resulting in higher quality generated radar data. The simulation dataset enables precise matching of noisy-noise image pairs, providing ideal conditions for the network to learn noise distribution characteristics. The parameter randomization generation mechanism ensures data diversity, contributing to improved model generalization ability. The strict division between the test and training sets follows machine learning standards, laying the foundation for objective evaluation of model performance. Furthermore, this invention utilizes a Transformer-CNN hybrid module to optimize the generator network and build the network model. Through a dual-path feature fusion mechanism, it efficiently captures fundamental spatial features using depthwise separable convolutions and parses the semantic relationships between feature channels using a channel attention mechanism, while mitigating the memory consumption issues associated with traditional global attention computation. While ensuring high-quality data denoising, it reduces the time cost of data generation, which is beneficial for improving the intelligence level, detection capabilities, and efficiency of the engineering detection industry.
[0064] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A ground penetrating radar deep learning data denoising method based on a generative adversarial network, characterized in that: The method includes the following steps: S1. The raw geological data are preprocessed using the finite-difference time-domain (FDTD) method to obtain the training set and the test set; S2. Construct a network model containing Transformer and CNN modules based on Generative Adversarial Network (GAN), and then import the training set into the network model for training to obtain the trained model; S3. The training model is continuously tested and optimized using a test set until the simulated denoised image data generated by the generator can be discriminated by the discriminator. This yields a denoising model for denoising ground-penetrating radar data.
2. The ground penetrating radar deep learning data denoising method based on a generative adversarial network according to claim 1, characterized in that: Step S1 specifically includes: S101. By dynamically adjusting at least one variable among the background medium electrical parameters, stratigraphic spatial distribution, target geometric features and antenna center frequency, a geological data sample library based on random parameters is constructed. S102. Use the FDTD algorithm to perform forward modeling on each sample data in the above geological data sample library to obtain a preprocessed dataset consisting of noisy and noiseless B-scan images; randomly allocate the preprocessed dataset to obtain a training set and a test set. S103. To improve the generalization ability of the model, random spatial cropping preprocessing is performed on the original GPR image to uniformly adjust it to the standard size of the same pixel.
3. The ground penetrating radar deep learning data denoising method based on a generative adversarial network according to claim 2, characterized in that: In step S101, the geometric features of the target body include at least one of the following: the number of target bodies, the size of the target body, and the burial depth of the target body; In step S102, the number of noisy and noiseless B-scan images in the preprocessed dataset is no less than 1000 sets; the ratio of the number of noisy and noiseless B-scan image sets in the training set and the test set is 9~8:1~2.
4. The ground penetrating radar deep learning data denoising method based on a generative adversarial network according to any one of claims 1-3, characterized in that: Step S2 specifically includes: S201. The input training set image is segmented into overlapping block regions. Then, shallow features are extracted through convolutional layers, and deep semantic information is gradually recovered using an encoder-decoder structure. Finally, in the feature reconstruction stage, channel dimension compression is achieved through cross-layer feature fusion and 1×1 convolution operation, while preserving key detail information. S202. Input the data in the training set into the generator, and the decoder in the generator converts the input data into simulated image data. Then, the discriminator judges and scores the generated simulated image data. S203. First, fix the parameters of the generator and train the parameters of the discriminator. Then fix the parameters of the discriminator and train the parameters of the generator again. Repeat this process until a combination of generator and discriminator parameters that meets the scoring requirements is obtained.
5. The ground penetrating radar deep learning data denoising method based on a generative adversarial network according to claim 4, characterized in that: In step S202, a composite loss function is used to train the generator network; the composite loss function consists of mean squared error, absolute error and adversarial loss.
6. The ground penetrating radar deep learning data denoising method based on a generative adversarial network according to any one of claims 1-3, characterized in that: Step S3 specifically includes: S301. Adjust the model hyperparameters and continuously test and optimize the trained model using the test set. Generate a batch of simulated denoised image data every 5 to 15 iterations. S302. Compare the generated simulated denoised image data with the original real data, and verify the network model performance by comprehensively considering the quality of the generated data and the speed of model convergence.
7. The ground penetrating radar deep learning data denoising method based on a generative adversarial network according to claim 6, characterized in that: In step S301, the model hyperparameters include the initial learning rate, the number of iterations, and the number of images; the optimization is performed using the Adam optimizer to minimize the generator and maximize the discriminator.
8. The ground penetrating radar deep learning data denoising method based on a generative adversarial network according to claim 7, characterized in that: The initial learning rates for both the generator and discriminator in the training model are set to 0.0001~0.0002, the number of iterations is set to 100~200, the number of images trained each time is 8~16, and the model parameters are saved once after every 5~15 iterations.
9. The ground penetrating radar deep learning data denoising method based on a generative adversarial network according to claim 6, characterized in that: In step S301, the main basis for adjusting the model hyperparameters is the model loss function and the model convergence time when different parameters are set, and the hyperparameter settings corresponding to the optimal model performance are saved.
10. The ground penetrating radar deep learning data denoising method based on a generative adversarial network according to claim 1, characterized in that: The requirement that the simulated denoised image data generated by the generator can be identified by the discriminator means that the pass rate of the simulated denoised image data generated by the generator being identified as real denoised image data by the discriminator is not less than 90%.