Ground penetrating radar data multi-frequency fusion method, device, equipment, medium and product

Abstract

The invention discloses a ground penetrating radar data multi-frequency fusion method and device, equipment, a medium and a product, and relates to the field of ground penetrating radar data processing, and the method comprises the steps: obtaining multi-frequency ground penetrating radar data; the multi-frequency ground penetrating radar data comprises at least two pieces of ground penetrating radar data with different frequency characteristics; preprocessing the multi-frequency ground penetrating radar data to obtain preprocessed multi-frequency data; performing cross-frequency mapping on the preprocessed multi-frequency data by using a Cycle-GAN network to obtain pseudo high-frequency ground penetrating radar data; and performing adaptive fusion on the pseudo high-frequency ground penetrating radar data and the actual high-frequency data by using a Transform fusion module to obtain a weighted fusion result. According to the invention, the quality of multi-frequency data fusion can be improved.

Description

Technical Field

[0001] This application relates to the field of ground penetrating radar data processing, and in particular to a method, apparatus, equipment, medium and product for multi-frequency fusion of ground penetrating radar data. Background Technology

[0002] The performance of Ground Penetrating Radar (GPR) systems is inherently constrained by the relationship between antenna frequency, detection depth, and spatial resolution: high-frequency antennas (such as 2 GHz) can achieve millimeter-level shallow resolution, but suffer from severe signal attenuation, making it difficult to penetrate deep media; while low-frequency antennas (such as 50 MHz), although capable of detection depths of tens of meters, cannot meet the requirements for fine target identification. Multi-frequency data fusion technology was developed to address this contradiction, but in the field of GPR multi-frequency data fusion, the problem of low data fusion quality still exists. Summary of the Invention

[0003] The purpose of this application is to provide a method, apparatus, equipment, medium, and product for multi-frequency fusion of ground-penetrating radar data, which can improve the quality of multi-frequency data fusion.

[0004] To achieve the above objectives, this application provides the following solution.

[0005] In a first aspect, this application provides a multi-frequency fusion method for ground-penetrating radar (GPR) data, comprising: acquiring multi-frequency GPR data; the multi-frequency GPR data including GPR data with at least two different frequency characteristics; preprocessing the multi-frequency GPR data to obtain preprocessed multi-frequency data; performing cross-frequency mapping on the preprocessed multi-frequency data using a Cycle-GAN network to obtain pseudo-high-frequency GPR data; and adaptively fusing the pseudo-high-frequency GPR data and actual high-frequency data using a Transformer fusion module to obtain a weighted fusion result.

[0006] In one embodiment, the preprocessing includes zero-time correction, direct wave removal, time-varying gain processing, and amplitude normalization.

[0007] In one embodiment, the loss function of the Cycle-GAN network is a weighted sum of bidirectional adversarial loss, cycle consistency loss, and multi-scale structural similarity loss.

[0008] In one embodiment, the loss function of the Transformer fusion module includes weighted pixel-level error and structural similarity loss.

[0009] In one embodiment, the calculation process of the loss function of the Transformer fusion module specifically includes: determining the prior weights of high-frequency components and spatial gradients based on pseudo-high-frequency ground-penetrating radar data and actual high-frequency data; determining the joint weights based on the prior weights of high-frequency components and spatial gradients; and determining the weighted pixel-level error and structural similarity loss based on the joint weights.

[0010] In one embodiment, the Transformer fusion module includes a multi-band feature embedding module, a multi-head attention extraction module, and a dynamic weight prediction module connected in sequence.

[0011] Secondly, this application provides a multi-frequency fusion device for ground-penetrating radar (GPR) data, comprising: an acquisition module for acquiring multi-frequency GPR data, wherein the multi-frequency GPR data includes GPR data with at least two different frequency characteristics; a preprocessing module for preprocessing the multi-frequency GPR data to obtain preprocessed multi-frequency data; a cross-frequency mapping module for performing cross-frequency mapping on the preprocessed multi-frequency data using a Cycle-GAN network to obtain pseudo-high-frequency GPR data; and an adaptive fusion module for adaptively fusing the pseudo-high-frequency GPR data and actual high-frequency data using a Transformer fusion module to obtain a weighted fusion result.

[0012] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned ground-penetrating radar data multi-frequency fusion method.

[0013] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned ground-penetrating radar data multi-frequency fusion method.

[0014] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned ground-penetrating radar data multi-frequency fusion method.

[0015] Based on the specific embodiments provided in this application, the following technical effects are disclosed.

[0016] This application provides a method, apparatus, device, medium, and product for multi-frequency fusion of ground-penetrating radar (GPR) data. The preprocessed multi-frequency data is cross-frequency mapped using a Cycle-GAN network to obtain pseudo-high-frequency GPR data. The pseudo-high-frequency GPR data and actual high-frequency data are then adaptively fused using a Transformer fusion module to obtain a weighted fusion result. Combining the Cycle-GAN network and the Transformer fusion module achieves effective low-frequency to high-frequency conversion and adaptive high-frequency fusion, preserving shallow texture and edge information while enhancing the expressive power of deep structures, thereby improving the quality of data fusion. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, 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 based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of a multi-frequency fusion method for ground-penetrating radar data.

[0019] Figure 2 This is a schematic diagram of the preprocessing of simulated data for the expected gprMax.

[0020] Figure 3 This is a graph showing the preprocessing results of the measured data.

[0021] Figure 4 This is a structural diagram of the Cycle-GAN network.

[0022] Figure 5 This is a structural diagram of the Transformer fusion module.

[0023] Figure 6 The image shows the results of a simulated experiment.

[0024] Figure 7 This is a graph showing the actual measured data results.

[0025] Figure 8 This is a flowchart of a multi-frequency fusion method for ground-penetrating radar data.

[0026] Figure 9 This is a schematic diagram of the functional modules of a multi-frequency fusion device for ground-penetrating radar data.

[0027] Figure 10 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0028] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and 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.

[0029] In the field of GPR multi-frequency data fusion, traditional methods mainly fall into three categories: time-domain fusion, frequency-domain fusion, and time-frequency fusion. Time-domain fusion methods are computationally intuitive and efficient, but it is difficult to accurately quantify the contribution of different frequency components in the time domain. Frequency-domain and time-frequency-domain methods rely on accurate modeling of antenna characteristic parameters and medium property information. Inaccurate parameter estimation can easily introduce fusion errors, and most methods lack in-depth modeling of the multi-scale spatial structure and high-level semantic information of the image, making it difficult to effectively preserve shallow detail features and enhance the representation of deep targets.

[0030] In recent years, deep learning technology has demonstrated its potential in this field, but existing methods still face challenges. Some methods, such as Cycle-GAN, lack guidance from shallow, real high-frequency data, making them prone to artifacts or loss of high-frequency details when reconstructing shallow targets. Other methods, such as Transformer, while able to preserve shallow high-frequency information, struggle to improve the spatial resolution of the fused data, failing to meet the needs of advanced applications. Therefore, a new GPR multi-frequency data fusion method is urgently needed to address the shortcomings of existing technologies.

[0031] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0032] In one exemplary embodiment, such as Figure 8 As shown, a multi-frequency fusion method for ground-penetrating radar data is provided, including the following steps.

[0033] Step 801: Acquire multi-frequency ground-penetrating radar data; the multi-frequency ground-penetrating radar data includes ground-penetrating radar data with at least two different frequency characteristics.

[0034] Step 802: Preprocess the multi-frequency ground-penetrating radar data to obtain preprocessed multi-frequency data.

[0035] Step 803: The preprocessed multi-frequency data is cross-frequency mapped using a Cycle-GAN network to obtain pseudo-high-frequency ground-penetrating radar data.

[0036] Step 804: Adaptively fuse the pseudo-high-frequency ground-penetrating radar data and the actual high-frequency data using the Transformer fusion module to obtain a weighted fusion result. In practical applications, the actual high-frequency data is the real high-frequency image.

[0037] By combining the Cycle-GAN network with the Transformer fusion module, effective conversion from low frequency to high frequency and adaptive fusion of high frequency are achieved. This preserves shallow texture and edge information while enhancing the expressive power of deep structures, thereby improving the quality of data fusion.

[0038] In one exemplary embodiment, the preprocessing includes zero-time correction, direct wave removal, time-varying gain processing, and amplitude normalization.

[0039] In an exemplary embodiment, the loss function of the Cycle-GAN network is a weighted sum of bidirectional adversarial loss, cycle consistency loss, and multi-scale structural similarity loss.

[0040] In one exemplary embodiment, the loss function of the Transformer fusion module includes weighted pixel-level error and structural similarity loss.

[0041] In practical applications, the calculation process of the loss function of the Transformer fusion module specifically includes: determining the prior weights of high-frequency components and spatial gradients based on pseudo-high-frequency ground-penetrating radar data and actual high-frequency data; determining the joint weights based on the prior weights of high-frequency components and spatial gradients; and determining the weighted pixel-level error and structural similarity loss based on the joint weights.

[0042] In practical applications, the Transformer fusion module includes a multi-band feature embedding module, a multi-head attention extraction module, and a dynamic weight prediction module connected in sequence.

[0043] like Figure 1 As shown, its core idea can be summarized as: how to synergistically utilize the deep penetration advantage of low-frequency data and the shallow high-resolution capability of high-frequency data to achieve high-quality data fusion across the entire depth domain.

[0044] This application discloses a multi-frequency fusion method based on multi-scenario ground-penetrating radar data.

[0045] In one exemplary embodiment, such as Figure 1 As shown, a multi-frequency fusion method for ground-penetrating radar (GPR) data is provided to improve GPR image quality and information fidelity, meeting the detection needs in complex real-world environments. The method includes the following steps.

[0046] Step 1: Acquire GPR multi-frequency data; the multi-frequency data must contain ground penetrating radar data with at least two different frequency characteristics.

[0047] Step 2: Multi-frequency ground-penetrating radar data preprocessing, and construct a multi-frequency fusion dataset based on the preprocessed ground-penetrating radar data.

[0048] Step 3: Implement cross-frequency mapping from low-frequency images to the high-frequency domain based on Cycle-GAN. During training, the input is the preprocessed high- and low-frequency ground-penetrating radar (GPR) images obtained in Step 2, and the output is a pseudo-high-frequency GPR image. In practical applications, the low-frequency GPR image is used to output a pseudo-high-frequency GPR image.

[0049] Step 4: Using the Transformer fusion module, the pseudo-high-frequency image generated by Cycle-GAN is adaptively fused with the real high-frequency image to output a weighted fusion result of the measured data and the pseudo-high-frequency data. The "real high-frequency image" referred to in this application originates from the high-frequency components in ground-penetrating radar (GPR) multi-frequency data. The original radar signal is stored in matrix form, representing the electromagnetic properties of the subsurface medium. Through digital signal processing (such as gain, filtering, and transfer) and visualization techniques, this data matrix can be converted into a grayscale image or pseudo-color image for training deep learning networks. For ease of description, this application refers to the original matrix without visualization processing as "data," and the visual representation generated after processing as "image."

[0050] Multi-frequency ground-penetrating radar (GPR) data includes: simulated GPR data obtained using gprMax simulation software and field-measured data using professional GPR equipment; at least one type of data is required. Field-measured data acquired using the IDS RIS-K2 is also required. Frequency ranges include 400MHz, 900MHz, and 1600MHz. High-frequency and low-frequency GPR data can be paired or unpaired, and acquisition methods can include gprMax software simulation and field measurements.

[0051] Preprocessing includes zero-time correction, direct wave removal, time-varying gain processing, and amplitude normalization of the raw data; for simulated data, Gaussian white noise is added before data gain to simulate noise interference in real-world GPR data; for measured data, time zero-point correction, automatic gain control, and amplitude normalization are performed to ensure the comparability and consistency of each dataset in subsequent fusion analysis.

[0052] The Cycle-GAN network contains two generator-discriminator pairs, where the generator... and This achieves bidirectional conversion from the low-frequency domain (A) to the high-frequency domain (B). The converted data is then processed by a discriminator. and By comparing the generated image with real data, the realism of the generated image is improved. Then, by inverse mapping back to the original domain, the cycle consistency loss is calculated with the initial data to ensure the physical interpretability of the generated result. The generator adopts a U-net architecture, consisting of an encoder, a converter, and a decoder.

[0053] The encoder extracts features from the input image. It consists of three convolutional layers that progressively compress the image size and increase the number of channels to extract higher-level features. This first layer has 64 filters, each 7×7 in size, with a stride of 1, and uses the ReLU activation function. It performs initial feature extraction on the input image, typically capturing basic features such as edges and textures. Using a larger 7×7 convolutional kernel captures a wider range of contextual information and basic texture features in the initial stage. A stride of 1 means that the spatial size of the image (K×K) is not changed; it primarily maps the input image from 3 channels to 64 feature channels for initial feature extraction. The ReLU activation function introduces non-linearity, enabling the network to learn more complex patterns. The second layer has 128 filters, each 3×3 in size, with a stride of 2, and uses the ReLU activation function. Through convolutions with a stride of 2, the image size is compressed to half its original size, which helps to expand the receptive field of subsequent layers, allowing the network to focus on more macroscopic structures. This step effectively reduces computation while extracting features. The third layer has 256 filters, each 3x3 in size with a stride of 2, using the ReLU activation function. The image size is further compressed to one-quarter of its original size, while the number of channels increases to 256 to extract more abstract and complex features. At this point, the encoder has compressed the original image into a small but highly dimensional tensor, which contains the core content and structural information of the original low-frequency image.

[0054] The transformer is the core of the generator, responsible for transforming the features extracted by the encoder into features of the target domain. There are 8 residual blocks: each block contains two 3×3 convolutional layers (256 kernels, stride s=1), and establishes a "skip connection" between the input and output of the block. The ReLU activation function is used. These blocks directly add the input to the output through skip connections, effectively mitigating the vanishing gradient problem and allowing the network to be trained deeper, thus learning more complex mapping relationships. They are responsible for performing deep transformations on the features extracted by the encoder, converting features from low-frequency images to features from high-frequency images. It finely adjusts and reorganizes the features without changing the feature map size, achieving "style" transfer.

[0055] The decoder functions as the encoder; it receives the high-dimensional features processed by the transformer and gradually restores the spatial resolution and details of the image through a series of deconvolution (or transposed convolution) operations, ultimately generating the target high-frequency image. The first layer is a deconvolution layer with 128 filters, each 3×3 in size with a stride of 2, using the ReLU activation function. It restores the image size to half its original size, reducing the number of channels to 128. The second layer is also a deconvolution layer with 64 filters, each 3×3 in size with a stride of 2, using the ReLU activation function. It restores the image size to its original size. The third and final layer has 3 filters, each 7×7 in size with a stride of 1, using the ReLU activation function. It restores the number of channels to 3 (typically corresponding to an RGB image), ultimately outputting the synthesized high-frequency image. Using a large 7×7 convolutional kernel helps integrate a wider range of features, generating a smooth and natural image.

[0056] The discriminator is designed to determine whether an input image is "real" (from a real dataset) or "fake" (synthesized by a generator). This architecture uses the PatchGAN discriminator. Traditional discriminators output a single "real / fake" probability value for the entire image. The PatchGAN discriminator, however, outputs an N×N feature map, where each element (or "pixel") corresponds to a local region (patch) in the input image, and determines the authenticity of that region. The advantages of this design are: it can more effectively determine the realism of local details and textures in an image, which is crucial for texture-rich scenes like GPR images; it also has relatively fewer parameters and runs faster.

[0057] Input Layer: Receives an image, which could be a real high-frequency GPR image or a fake high-frequency image generated by the generator G_AB. The first convolutional layer has 64 4×4 kernels with a stride of s=2, using the LeakyReLU activation function. It performs initial feature extraction and downsampling on the input image. The combination of 4×4 kernels and a stride of 2 is common in discriminators. LeakyReLU is an improved ReLU that allows a small non-zero slope when the input is negative, which helps prevent neuron "death," allows gradients to flow better during training, and enhances the stability of the discriminator training. The second convolutional layer has 128 4×4 kernels with a stride of s=2, using LeakyReLU. It further extracts deeper features and halves the feature map size. The number of channels is increased to capture more complex patterns. The third convolutional layer has 256 4×4 kernels with a stride of s=2, using LeakyReLU. The feature extraction process is further refined, with the feature map size halved again. The fourth convolutional layer has 512 4×4 kernels with a stride of s=2, using LeakyReLU. This is the deepest feature extraction layer, possessing the most feature channels (512), capable of capturing very abstract and complex discriminative features. The fifth layer (output layer): has one 4×4 kernel with a stride of s=1. This is the final discriminative layer. It compresses the 512 high-dimensional feature maps extracted by the previous layer into a single-channel feature map (i.e., N×N×1). Each value in this output map represents the discriminator's "realism" score for a local region (patch) corresponding to the input image. During training, the network aims to make the values ​​of real images approach 1 on this output map, while the values ​​of fake images approach 0. This output map is called the criterion-based realism judgment.

[0058] The loss function of this module consists of adversarial loss, cycle consistency loss, and multi-scale structural similarity loss. The final total loss function is a weighted sum of these three. Hyperparameters are determined through heuristic search to balance model complexity and fitting performance. The quality and physical interpretability of the generated images are optimized using a specific loss function.

[0059] In one embodiment, Cycle-GAN is used to perform cross-frequency mapping from low-frequency images to the high-frequency domain, with the model's input size uniformly set to 256×256 pixels. The loss function of this module consists of three parts: adversarial loss, cycle consistency loss, and multi-scale structural similarity loss. The adversarial loss is jointly optimized by the generator and discriminator. The generator aims to minimize the probability of the generated image being detected by the discriminator, while the discriminator aims to maximize its ability to distinguish between real and generated images. The bidirectional adversarial loss (…) The definition is as follows: .

[0060] Here, the formula represents the bidirectional adversarial loss, which consists of the generator loss and discriminator loss from the low-frequency domain to the high-frequency domain, and the generator loss and discriminator loss from the high-frequency domain to the low-frequency domain, respectively. a and b represent the high-frequency and low-frequency dataset samples, respectively. and These represent the generation processes from the low-frequency domain to the high-frequency domain and from the high-frequency domain to the low-frequency domain, respectively. and The discriminator is shown as the corresponding domain, where a~A and b~B represent sampling from the low-frequency domain and the high-frequency domain, respectively. This is an adversarial loss function.

[0061] Cyclic consistency loss This is the core mechanism of the Cycle-GAN network, and its role is to ensure the generalization ability of the generator and reduce its dependence on paired data. The specific formula is as follows: .

[0062] Multiscale structural similarity loss Constraining the structure of generated data by examining its detail, contrast, and brightness to suppress artifacts has been proven to be effective in deblurring. The specific calculation formula is as follows: .

[0063] .

[0064] Let x and y represent the mean, variance, and covariance of the images, respectively, and a and b represent high-frequency and low-frequency dataset samples, respectively. and These represent the generation process from the low-frequency domain to the high-frequency domain and from the high-frequency domain to the low-frequency domain, respectively. Indicates calculation and b's multi-scale structural loss, Indicates calculation The weighted multi-scale structural loss is calculated using 'a' and 'a'. Specifically, MS-SSIM employs five different scales of SSIM computation, with weights... Attenuation by depth: .

[0065] Finally, the total loss function The weighted sum of the above three parts is: .

[0066] The hyperparameters λ and γ were determined heuristically to balance model complexity and fitting performance. Experiments show that when λ=0.1 and γ=0.001, the model achieves the optimal balance between training stability and objective function convergence. During optimization, all experiments used an initial learning rate of 0.002, which was kept constant for the first 100 epochs and then linearly decayed to zero for the subsequent 100 epochs to accelerate convergence and avoid overfitting.

[0067] Step 4 of the Transformer fusion module includes a multi-band feature embedding module, a multi-head attention extraction module, and a dynamic weight prediction module. The multi-band feature embedding module inputs the original high-frequency image and the pseudo-high-frequency image into their respective embedding layers, maps them to a unified semantic space, generates feature vector representations, and concatenates them to obtain a joint feature representation. The multi-head attention extraction module uses a multi-head self-attention mechanism to model the long-range dependencies between different time steps, captures response patterns, and obtains a comprehensive feature representation containing multi-dimensional and multi-level contextual information. The dynamic weight prediction module generates adaptive fusion weights through feedforward layers and linear layers.

[0068] Step 4 is used to design the time-frequency optimization loss function, including high-frequency component prior weight extraction, spatial gradient prior weight extraction, joint weight calculation, and construction of the final loss function. A short-time Fourier transform is performed on the input signal to obtain its time-frequency representation. The high-frequency prior weight is defined as the energy within a preset high-frequency band. A second-order difference operator is used to detect drastic changes in the spatial gradient, generating spatial gradient prior weights. The high-frequency prior and spatial gradient weights are linearly combined to generate the spatial-frequency joint weights. The final loss function consists of weighted pixel-level error and structural similarity loss, guiding the model to focus on regions requiring accurate reconstruction, improving the fidelity of the generated image in high-frequency details and structural boundaries, and accelerating convergence.

[0069] In one embodiment, the Transformer fusion module in step 4 above consists of three key components: 1) a multi-band feature embedding module: which embeds the original high-frequency image... Pseudo-high frequency images after Cycle-GAN transformation Each input embedding layer is mapped to a unified semantic space to generate a feature vector representation.

[0070] .

[0071] .

[0072] in , , and The learnable weight matrix is ​​concatenated to obtain the joint feature representation. The original GPR data was transformed into a feature sequence that contains both data content information and spatial location information.

[0073] 2) Multi-head attention extraction module: A multi-head self-attention mechanism is employed to model long-range dependencies between different time steps, capturing response patterns from both measured and generated data across multiple time scales. This adaptively identifies regions where true signals need to be preserved and regions suitable for enhancing high-frequency details. The attention mechanism is based on Scaled Dot-Product Attention. To compute self-attention, the feature sequence E is processed through three independent learnable linear projection matrices. , , Generate query (Q), key (K), and value (V) matrices respectively: .

[0074] .

[0075] .

[0076] Based on this, the multi-head attention mechanism computes attention in parallel by performing h independent linear projections on Q, K, and V, as follows.

[0077] .

[0078] .

[0079] MultiHead(Q, K, V) .

[0080] Among them, Q (query), K (key), and V (value) all come from the feature sequence output in the previous stage. It is the learnable projection matrix of the i-th head. Let represent the output of the i-th attention head, and h represent the number of attention heads. The outputs of all heads are concatenated using the Concat operation, and then projected through the final projection matrix. The fusion process yields a comprehensive feature representation, MultiHead(Q, K, V), denoted as H, which contains multi-dimensional and multi-level contextual information. Each element in H encodes the fusion decision basis for the original high-frequency signal and the pseudo-high-frequency signal under global correlation, laying a solid foundation for subsequent adaptive fusion weight prediction.

[0081] 3) Dynamic weight prediction module: Generates adaptive fusion weights through feedforward and linear layers. .

[0082] .

[0083] in The hidden states represent the features of both real high-frequency data and pseudo-high-frequency data. For residuals, The final output of the Sigmoid function is... Weighted fusion of measured data and pseudo-high-frequency data: .

[0084] Here, ⊙ represents the Hadamard product, which ensures consistency in time and magnitude during the fusion process through element-wise multiplication. Represents the original high-frequency image. Characterize pseudo-high frequency images.

[0085] In one embodiment, the loss function of the fusion module in step 4 above specifically includes the following parts.

[0086] 1) Prior weights of high-frequency components Extraction: High-frequency components in ground-penetrating radar images typically correspond to the scattering signals of subtle bedding, textures, and small anomalies, and are crucial for improving image resolution. A short-time Fourier transform (STFT) is performed on the input signal (A-scan) to obtain its time-frequency representation. .

[0087] .

[0088] in, It is the i-th A-scan. This is a window function; here, the Hamming window is chosen to suppress spectral leakage. Then, high-frequency prior weights are defined. Preset high frequency band Energy within. This is the lower limit of the effective frequency band. The upper limit of the effective frequency band, This indicates the actual frequency of the data. For example, using 900MHz data... and Using values ​​of 800 and 1000 respectively allows the model to focus on spatiotemporal locations that should physically contain rich high-frequency information.

[0089] .

[0090] 2) Spatial gradient prior weights Extraction: Reflections from targets such as formation interfaces, pipelines, or root systems in GPR images often exhibit abrupt signal changes. To highlight these important structural boundaries, a second-order difference operator (i.e., a one-dimensional Laplacian operator) is used to detect drastic changes in spatial gradients. This operator generates prior weights for the spatial gradient through convolution with the image. .

[0091] .

[0092] 3) Joint weight calculation and final loss function: The high-frequency priors and spatial gradient weights are linearly combined to generate spatial-frequency joint weights. .

[0093] .

[0094] Final loss function From weighted pixel-level error and structural similarity loss constitute.

[0095] .

[0096] .

[0097] .

[0098] .

[0099] For structural similarity loss, ⊙ denotes element-wise multiplication. This design intelligently guides the model's attention to the regions most in need of accurate reconstruction. This targeted optimization strategy based on frequency and gradient priors not only significantly improves the fidelity of the generated images in high-frequency details and structural boundaries but also provides the model with a clearer optimization direction, thereby effectively accelerating convergence.

[0100] This application acquires ground-penetrating radar (GPR) data, which needs to include high- and low-frequency data of different frequencies. The acquisition method can be simulated using gprMax software or collected in the field by a professional GPR system. The acquired data is preprocessed, and a multi-frequency GPR dataset is constructed based on the preprocessed data. The multi-frequency GPR data fusion features are dominated by high-frequency information in shallow layers and by low-frequency information in deeper layers. Multiple training datasets from different frequencies and scenarios are input into a Cycle-GAN-based cross-frequency mapping module for frequency domain transformation and time domain alignment. A Transformer network is used to fuse the synthesized pseudo-high-frequency data with the real high-frequency data to obtain the final fusion result. The trained model is then applied to the fusion of GPR data in different scenarios. By combining Cycle-GAN with Transformer, we achieve efficient conversion of low-frequency images to the high-frequency domain and adaptive fusion of high- and low-frequency data. This preserves shallow texture and edge information while enhancing the expressive power of deep structures. The designed time-frequency optimized loss function improves the fidelity of the generated images in high-frequency details and structural boundaries. Validated on multi-scene datasets, it exhibits good robustness and universality, providing an effective solution for multi-frequency GPR image fusion in complex media environments. It can be widely applied in fields such as geological exploration and engineering inspection.

[0101] In one embodiment, such as Figure 2 As shown, where, Figure 2 (a) Figure 2 (b) and Figure 2 (c) in the figure represents the original simulation data at 400MHz, 900MHz, and 1600MHz, respectively. Figure 2 (d) Figure 2 (e) and Figure 2 (f) represents the preprocessed data at 400MHz, 900MHz, and 1600MHz, respectively. Figure 2 (g) in Figure 2 (h) and Figure 2 In the diagram (i), the results are shown after adding Gaussian noise that increases with depth. The first row shows the data without preprocessing. Due to the inherent attenuation of electromagnetic waves propagating in the medium, the reflected signal energy of deep targets is very weak and almost invisible in the image. The second row shows the results after data preprocessing. It can be seen that the direct wave is effectively suppressed, and through gain compensation, the hyperbolic reflection characteristics of deep targets are significantly enhanced, and the signal-to-noise ratio is greatly improved. The third row shows the results after adding Gaussian white noise. This processing method significantly amplifies the background noise in the deep layer (below 15 ns) by the gain function, resulting in a significantly lower signal-to-noise ratio (SNR) in the deep layers of the image compared to the shallow layers. This is highly consistent with the actual distribution characteristics of the data collected on-site.

[0102] In one embodiment, such as Figure 3 As shown, Figure 3 (a) and Figure 3 (b) shows the 1600MHz and 900MHz data collected in the Mu Us Desert. After data preprocessing, the 1600MHz high-frequency data displayed extremely high resolution, capable of finely depicting the subtle layering structure of the near-surface (approximately 0-10 ns) soil medium. However, this came at the cost of severely insufficient penetration depth, with almost no effective signal at deeper levels. Figure 3 Although the 900 MHz data in (b) is not as good as 1600 MHz in terms of shallow detail resolution, its detection depth is significantly increased, clearly revealing the stratigraphic structure at a deeper level (about 10-20 ns).

[0103] In one embodiment, such as Figure 4 The diagram illustrates the basic structure of the Cycle-GAN cross-frequency mapping module. It includes a generator and a discriminator, optimizing the quality and physical interpretability of the generated images through a specific loss function. The input is high- and low-frequency ground-penetrating radar images, and the output is the corresponding pseudo-high-frequency image.

[0104] In one embodiment, such as Figure 5 As shown, the basic structure of Transformer high-frequency feature fusion is illustrated. It mainly consists of three parts: a multi-band feature embedding module, a multi-head attention extraction module, and a dynamic weight prediction module.

[0105] In one embodiment, such as Figure 6 The diagram illustrates the effect of the simulated data after steps 3 and 4, as well as the results of using steps 3 and 4 independently. The specific task is to fuse low-frequency (400MHz and 900MHz) simulated data with high-frequency (1600MHz) data. Figure 6 (a) and Figure 6 (b) shows the fusion results of 400MHz, 900MHz, and 1600MHz obtained using the module in step 3, respectively. Figure 6 (c) and Figure 6 (d) represents the fusion results of 400MHz, 900MHz, and 1600MHz obtained using module 4, respectively. Figure 6 (e) and Figure 6 (f) represents the fusion results of 400MHz, 900MHz and 1600MHz obtained by using modules in steps 3 and 4, respectively. Figure 6 (a) and Figure 6 (b) shows the results of fusing high- and low-frequency data using the step 4 module alone. Figure 6 (c) and Figure 6 (d) shows the results using only the step 3 module and the high and low frequency data. Figure 6 (e) and Figure 6 Figure (f) shows the result of using steps 3 and 4 together. Using only step 4, the model fails to learn the mapping relationship well, resulting in the frequency transition band between shallow and deep layers being almost completely submerged in noise, while the middle soil layer remains blurred. Although the result improves as the frequency difference narrows, significant blocky noise remains in the background, and the overall quality of the fused image is not high. Using only step 3, while the hyperbolic contour of the underground target can be roughly reconstructed, the shallow image suffers from severe resolution deficiency, and soil background clutter interference is significant. The image fused using steps 3 and 4 shows a significant reduction in artifacts in the shallow region, and the visual effect is closer to the ideal high-frequency data.

[0106] In one embodiment, such as Figure 7 The results show the effect of the measured data after passing through steps 3 and 4, as well as the results after passing through steps 3 and 4 alone. The specific task is to fuse low-frequency (900MHz) and high-frequency (1600MHz) data. Figure 7 (a) shows the result of fusing 900MHz and 1600MHz using the module in step 3. Figure 7 (b) shows the result of fusing 900MHz and 1600MHz using the module in step 4. Figure 7 Image (c) shows the fusion result of 900MHz and 1600MHz obtained using modules 3 and 4. Using only the result from step 3 (a) enhances the detail representation of deep soil layers, but its reconstruction effect in shallow areas is not ideal. Using only the result from step 4 (b), while preserving the high-resolution fine roots of shallow layers, does not improve the resolution of deep soil layers. The fusion result combining steps 3 and 4 shows high visual consistency with real high-frequency data, especially in key soil stratification areas, where high-frequency texture recovery is significant, while also well preserving important fine root details in the shallow layer (0~5ns). This indicates that the multi-scene ground-penetrating radar multi-frequency fusion method not only improves resolution but also successfully leverages the penetration depth advantage of low-frequency signals to generate a comprehensive image containing richer geological structure information and reducing information loss.

[0107] Based on the same inventive concept, this application also provides a ground-penetrating radar data multi-frequency fusion device for implementing the above-mentioned ground-penetrating radar data multi-frequency fusion method. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more ground-penetrating radar data multi-frequency fusion device embodiments provided below can be found in the limitations of the ground-penetrating radar data multi-frequency fusion method above, and will not be repeated here.

[0108] In one exemplary embodiment, such as Figure 9 As shown, a ground-penetrating radar data multi-frequency fusion device is provided, comprising: The acquisition module is used to acquire multi-frequency ground-penetrating radar data; the multi-frequency ground-penetrating radar data includes ground-penetrating radar data with at least two different frequency characteristics.

[0109] The preprocessing module is used to preprocess multi-frequency ground-penetrating radar data to obtain preprocessed multi-frequency data.

[0110] The cross-frequency mapping module is used to perform cross-frequency mapping on the preprocessed multi-frequency data using a Cycle-GAN network to obtain pseudo-high-frequency ground-penetrating radar data.

[0111] The adaptive fusion module is used to adaptively fuse the pseudo-high-frequency ground-penetrating radar data and the actual high-frequency data using the Transformer fusion module to obtain a weighted fusion result.

[0112] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 10 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores multi-frequency fusion data from ground-penetrating radar (GPR). The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a multi-frequency fusion method for GPR data.

[0113] Those skilled in the art will understand that Figure 10The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method embodiments.

[0114] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the above-described method embodiments.

[0115] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the above-described method embodiments.

[0116] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0117] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0118] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0119] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0120] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for multi-frequency fusion of ground-penetrating radar data, characterized in that, The ground-penetrating radar data multi-frequency fusion method includes: Acquire multi-frequency ground-penetrating radar data; the multi-frequency ground-penetrating radar data includes ground-penetrating radar data with at least two different frequency characteristics; The multi-frequency ground-penetrating radar data is preprocessed to obtain preprocessed multi-frequency data. The preprocessed multi-frequency data is cross-frequency mapped using a Cycle-GAN network to obtain pseudo-high-frequency ground-penetrating radar data. The pseudo-high-frequency ground-penetrating radar data and the actual high-frequency data are adaptively fused using the Transformer fusion module to obtain a weighted fusion result.

2. The multi-frequency fusion method for ground-penetrating radar data according to claim 1, characterized in that, The preprocessing includes zero-time correction, removal of direct waves, time-varying gain processing, and amplitude normalization.

3. The multi-frequency fusion method for ground-penetrating radar data according to claim 1, characterized in that, The loss function of the Cycle-GAN network is a weighted sum of bidirectional adversarial loss, cycle consistency loss, and multi-scale structural similarity loss.

4. The multi-frequency fusion method for ground-penetrating radar data according to claim 1, characterized in that, The loss function of the Transformer fusion module includes weighted pixel-level error and structural similarity loss.

5. The multi-frequency fusion method for ground-penetrating radar data according to claim 4, characterized in that, The calculation process of the loss function of the Transformer fusion module specifically includes: The prior weights of high-frequency components and spatial gradients are determined based on pseudo-high-frequency ground-penetrating radar data and actual high-frequency data. The joint weights are determined based on the prior weights of high-frequency components and the prior weights of spatial gradients. The weighted pixel-level error and structural similarity loss are determined based on the joint weights.

6. The multi-frequency fusion method for ground-penetrating radar data according to claim 1, characterized in that, The Transformer fusion module includes a multi-band feature embedding module, a multi-head attention extraction module, and a dynamic weight prediction module connected in sequence.

7. A multi-frequency fusion device for ground-penetrating radar data, characterized in that, The ground-penetrating radar data multi-frequency fusion device includes: An acquisition module is used to acquire multi-frequency ground-penetrating radar data; the multi-frequency ground-penetrating radar data includes ground-penetrating radar data with at least two different frequency characteristics; The preprocessing module is used to preprocess multi-frequency ground-penetrating radar data to obtain preprocessed multi-frequency data. The cross-frequency mapping module is used to perform cross-frequency mapping on the preprocessed multi-frequency data using a Cycle-GAN network to obtain pseudo-high-frequency ground-penetrating radar data. The adaptive fusion module is used to adaptively fuse the pseudo-high-frequency ground-penetrating radar data and the actual high-frequency data using the Transformer fusion module to obtain a weighted fusion result.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the ground-penetrating radar data multi-frequency fusion method according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the multi-frequency fusion method for ground-penetrating radar data as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the multi-frequency fusion method for ground-penetrating radar data as described in any one of claims 1-6.

Images