Ground penetrating radar high signal-to-noise ratio image imaging method

Abstract

The application belongs to the technical field of tunnel lining detection, and specifically discloses a ground penetrating radar high signal-to-noise ratio image imaging method, which comprises the following steps: constructing a training data set of pairs of noisy ground penetrating radar B-scan images and clean images; building an improved U-Net denoising network model, embedding a CBAM attention module in the encoder and the decoder, and enhancing effective signal features and suppressing noise through channel attention and spatial attention; using the training data set to train the model, optimizing network parameters with clean images as expected output; inputting the image to be processed into the trained model, and outputting a high signal-to-noise ratio image. The application can effectively separate disease signals in a strong noise interference background under single disease scene and combined disease scene conditions, and does not appear signal distortion, amplitude attenuation and edge blur, and the denoised image is almost consistent with the original clean image, thereby providing a strong basis for subsequent intelligent disease recognition.

Description

Technical Field

[0001] This invention belongs to the field of tunnel lining detection technology, and specifically relates to a high signal-to-noise ratio image imaging method using ground penetrating radar. Background Technology

[0002] In the process of non-destructive target detection of tunnel lining, the uneven distribution of the medium and the complex electromagnetic environment in the field detection environment cause various forms of noise signal components to be superimposed on the effective reflected signals of defects in the radar profile images. Therefore, directly using the original B-scan images for intelligent defect identification is ineffective and seriously hinders the fine-grained localization of defect features. Noise reduction preprocessing of the radar profile images is also required. Existing image denoising methods include a denoising method based on Complete Set Empirical Mode Decomposition (CEEMD) combined with Permutation Entropy (PE) (CEEMD-PE) and a denoising method based on CEEMD combined with Wavelet Transform (WT) (CEEMD-WT). However, both of these methods have significant limitations in practical applications.

[0003] CEEMD-PE primarily relies on permutation entropy thresholds to separate effective signals from noise. It only reflects the sequence complexity and randomness of IMF components at each order, representing a single-dimensional linear statistical characteristic. In high-intensity noise environments, the permutation entropy values ​​of noise and weak signals highly overlap, easily misjudging IMF components containing weak effective signals as noise and suppressing them. This leads to loss of detail in the defective signal, reduced fidelity, and a significant amount of residual noise in the reconstructed image. Furthermore, determining the permutation entropy threshold is highly subjective, has poor scene adaptability, and is prone to over-denoising leading to signal distortion or excessive noise residue.

[0004] While the CEEMD-WT method outperforms CEEMD-PE in terms of denoising quantization metrics and image detail preservation, it suffers from high computational cost. Denoising a single defect model takes 2 to 3 minutes, consuming excessive time when dealing with large amounts of tunnel lining inspection data. Furthermore, the wavelet decomposition performance of this method heavily relies on the wavelet basis type, number of decomposition levels, wavelet threshold quantization method, and wavelet threshold function selection. Many parameters require extensive manual tuning beforehand, leading to low computational efficiency.

[0005] Furthermore, the CEEMD-PE and CEEMD-WT methods only demonstrate some effectiveness under low-noise interference conditions. When the noise intensity increases further, the noise energy density is much higher than the effective signal of the disease, resulting in blurred edges and poor continuity of the strong reflection phase axis of the disease, and complete loss of weak reflection details, leading to extremely low similarity to the original uncluttered image. For example, although CEEMD-WT noise reduction suppresses noise to some extent, a large amount of residual noise still exists, resulting in an unsatisfactory noise reduction effect. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a high signal-to-noise ratio image imaging method for ground penetrating radar.

[0007] This invention provides a high signal-to-noise ratio (SNR) image imaging method for ground-penetrating radar, comprising the following steps: S1: Construct a training dataset, which includes multiple sets of pairs of noisy ground-penetrating radar B-scan images and corresponding clean images; S2: Construct an improved U-Net denoising network model, which includes an encoder, a decoder, a skip connection connecting the encoder and the decoder, and CBAM attention modules embedded in the encoder and decoder. S3: The improved U-Net denoising network model is trained using the training dataset. The noisy ground-penetrating radar B-scan image is used as input, and the clean image is used as the desired output. The network parameters are optimized through backpropagation to obtain the trained denoising model. S4: Input the ground-penetrating radar B-scan image to be processed into the trained noise reduction model, and output a ground-penetrating radar image with a high signal-to-noise ratio.

[0008] A further scheme is that, in step S1, the noisy ground-penetrating radar B-scan image is generated by adding noise with a signal-to-noise ratio of -15dB to a clean image, or by directly acquiring the original ground-penetrating radar B-scan image from the tunnel site without preprocessing; the clean image is an image processed by background removal, digital filtering, and gain adjustment of the measured data; the training dataset includes four types of defects: voids, looseness, cracks, and voids.

[0009] A further approach is that, in step S1, the noisy ground-penetrating radar B-scan image and the corresponding clean image are uniformly cropped or scaled to a fixed resolution, and divided into a training set, a validation set, and a test set in an 8:1:1 ratio.

[0010] A further embodiment is that the encoder in step S2 consists of multiple cascaded downsampling modules, each downsampling module including two consecutive 3×3 convolutional layers, a ReLU activation function, and a 2×2 max pooling layer; the decoder consists of multiple cascaded upsampling modules, each upsampling module including a 2×2 transposed convolutional layer, a feature concatenation layer, and two consecutive 3×3 convolutional layers and a ReLU activation function.

[0011] A further approach is to use the skip connection to concatenate the feature map of the corresponding layer of the encoder after central cropping with the feature map of the decoder after upsampling along the channel dimension, thereby achieving the fusion of shallow detail features and deep semantic features.

[0012] A further embodiment is that the CBAM attention module in step S2 includes a cascaded channel attention unit and a spatial attention unit. The channel attention unit is configured to perform the following steps: The input feature map is subjected to global max pooling and global average pooling respectively, and the resulting two channel vectors are input into a shared multilayer perceptron. The two output vectors of the multilayer perceptron are added together and then normalized by the Sigmoid activation function to obtain the channel attention weight map; The channel attention weight map is multiplied with the input feature map channel by channel to achieve feature filtering in the channel dimension, and an intermediate feature map is obtained. The spatial attention unit is configured to perform the following steps: The feature maps are subjected to both channel-dimensional max pooling and average pooling. The two single-channel feature maps are concatenated along the channel dimension; The importance weights of each spatial location are calculated using 7×7 convolution; The spatial attention weight map is obtained by normalizing using the Sigmoid activation function; The spatial attention weight map is then multiplied element-wise with the intermediate feature map to obtain the final output feature map.

[0013] A further approach is to use the mean squared error loss function as the training loss in step S3, and calculate the loss value with a clean image as the desired output.

[0014] A further approach involves optimizing network parameters through backpropagation, including: calculating the gradient of the loss function with respect to the network parameters and updating the network parameters using an optimizer; and employing an early stopping mechanism during training to prevent overfitting.

[0015] A further approach is to use signal-to-noise ratio, peak signal-to-noise ratio, root mean square error, and structural similarity index as evaluation metrics to assess the difference or similarity between the denoised image and the clean image, thereby quantifying the denoising effect.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention further enhances the feature extraction capability of the model by introducing a CBAM module during the upsampling and downsampling process of the encoder and decoder in the U-net network. For simulation data with added -15dB high-intensity noise, the improved U-net model effectively separates the disease signal under strong noise interference in both single and combined disease scenarios, accurately recovers the phase characteristics of the disease hyperbola (black-white-black), and does not exhibit signal distortion, amplitude attenuation, or edge blurring. The denoised image is almost identical to the original clean image, providing a strong basis for subsequent intelligent disease identification.

[0017] (2) After denoising using the improved U-net model, the random noise and clutter in the background region of the B-scan image are completely suppressed, the background is clean and uniform, and the gray value is close to 0, which greatly improves the distinction between the disease target and the background. During the denoising process, the hyperbolic edge contour is highly preserved, the phase axis is continuous and smooth, and the phase characteristics of "black-white-black" are restored. There is no phenomenon of blurred disease edges or amplitude distortion common in traditional denoising methods.

[0018] (3) For measured data with complex backgrounds and dense reinforcement interference, the improved model can effectively suppress random scattering caused by non-uniform media and electromagnetic interference noise from the external environment. In particular, for concrete areas with dense reinforcement, the improved model can effectively suppress the ringing effect and multiple wave interference caused by reinforcement reflection, significantly improve target resolution, and has good generalization ability in practical engineering. Attached Figure Description

[0019] The following figures are for illustrative purposes only and are not intended to limit the scope of the invention, wherein: Figure 1 : Flowchart of the high signal-to-noise ratio image imaging method of ground penetrating radar of the present invention; Figure 2 Example image of a simulated noisy image dataset; Figure 3 Example image of the actual test dataset; Figure 4 Schematic diagram of the improved U-Net noise reduction network structure; Figure 5 Schematic diagram of the self-encoder structure; Figure 6 Network model training flowchart; Figure 7 Schematic diagram of the channel attention unit structure; Figure 8 Schematic diagram of spatial attention unit structure; Figure 9 : Schematic diagram of the overall structure of the CBAM attention module; Figure 10Comparison of tunnel lining defects images after improved U-Net noise reduction; Figure 11 The first set of measured data denoising results (a. Standard reference image, b. Original noisy image, c. Improved U-Net denoising results); Figure 12 The second set of measured data noise reduction results (a. standard reference image, b. original noisy image, c. improved U-Net noise reduction result). Detailed Implementation

[0020] To make the objectives, technical solutions, design methods, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0021] This invention provides a high signal-to-noise ratio image imaging method for ground penetrating radar, such as... Figure 1 As shown, it includes the following steps: Step S1: Construct a training dataset, including multiple pairs of noisy ground-penetrating radar B-scan images and their corresponding clean images. The clean images serve as the model's true labels, helping the model assess denoising effectiveness, while the noisy ground-penetrating radar B-scan images serve as the model's input labels. To further verify the denoising effect of the U-net model, -15dB high-intensity noise was added to the simulation data as noisy images. Simultaneously, these four disease types were randomly combined to construct diverse noisy combined disease simulation data, such as... Figure 2 As shown. Since the above data are all simulation data, in order to avoid the model overfitting to the simulation scenario and ensure that the model can also achieve good noise reduction performance in the actual field, and improve the generalization ability of the model in actual engineering applications, this embodiment collects multiple sets of field-measured noisy data as input noisy images. Since the U-net model requires a one-to-one correspondence between noisy images and clean images as input labels when using noise reduction, the measured noisy data also needs to be preprocessed accordingly to obtain the corresponding clean images. In this embodiment, the measured noisy ground-penetrating radar B-scan images are collected using the domestic LTD ground-penetrating radar developed by the Qingdao Branch of the China Institute of Radio Wave Propagation, obtaining ground-penetrating radar format data with the .ite suffix. The images after background removal, digital filtering, gain adjustment, etc., of the measured data are processed by IDSP7 software and used as clean images, such as... Figure 3The image shown is a measured noisy ground-penetrating radar B-scan image and a clean image after software preprocessing. In this embodiment, the U-net model input samples contain a total of 2371 pairs, including four types of defects: voids, looseness, cracks, and delamination, and their various combinations. This constructs a diverse dataset covering multiple types of defects and high-intensity noise. This dataset is divided into training, validation, and test sets in an 8:1:1 ratio. The training set is used for iterative optimization of model parameters, the validation set is used to monitor overfitting during training and adjust hyperparameters, and the test set is used to evaluate the final noise reduction performance of the model.

[0022] Step S2: Construct an improved U-Net denoising network model. This improved U-Net denoising network model includes an encoder, a decoder, and skip connections connecting the encoder and decoder. A CBAM module is introduced during the upsampling and downsampling process of the encoder and decoder in the U-Net network to improve the model's feature extraction capability. For example... Figure 4As shown, the left half of the network is the encoder structure, serving as the core unit for feature extraction and noise filtering. It consists of three cascaded downsampling modules and three 2×2 max-pooling layers. Each encoder module contains two consecutive 3×3 convolutional layers and a ReLU activation function. The spatial dimensions of the feature maps remain unchanged after convolution, ensuring feature alignment in subsequent skip connections while fully extracting local features. The ReLU activation function introduces non-linearity, ensuring the model learns the complex mapping relationship between noise and clean images. The entire feature extraction process exhibits a clear progression: the first encoding module receives a 3-channel input image, increases the number of channels to 64 through double convolution, and extracts basic detail features such as shallow edges and textures; after downsampling through a 2×2 max pooling layer, the feature map space size is halved while the number of channels remains unchanged and is input into the second encoding module, which increases the number of channels to 128, aggregates shallow details into intermediate local semantic features, and filters out some noise; after another max pooling downsampling, the third encoding module increases the number of channels to 256, extracts high-level core features related to the global structure of the image, and completely removes noise; finally, after a third max pooling, it is fed into the bottleneck layer, forming a downsampling feature extraction mode where the number of channels increases from 3 to 256 and the space size decreases by a factor of two, realizing feature aggregation and noise filtering layer by layer from details to semantics and then to core features. The right half of the network is the decoder, the core unit for feature upsampling and detail restoration. It is strictly symmetrically distributed with the encoder, consisting of three cascaded decoding modules and three 2×2 transposed convolutional layers (with a stride of 2). Each decoding module follows the standard process of "upsampling-feature alignment-channel concatenation-feature fusion": the transposed convolutional layer first doubles the spatial size of the feature map output from the bottleneck layer and halves the number of channels, successively 512→256, 256→128, 128→64; subsequently, to resolve the gap between the corresponding layer feature map of the encoder and the upsampled feature map of the decoder... To address potential minor size discrepancies, a center clipping operation is introduced to calibrate the spatial dimension of the encoder feature map, ensuring a perfect size match. Then, along the channel dimension where dim=1, the upsampled features are concatenated with the calibrated encoder features to achieve channel superposition; for example, 256-channel upsampled features and 256-channel encoder features are superimposed to form a 512-channel fused feature. The input of the concatenated fused feature includes two 3×3 convolutional layers and ReLU activation blocks. Nonlinear transformations are used to extract effective features and halve the number of channels while maintaining the same spatial dimension. The feature fusion process in the decoder has a clear hierarchical correspondence: (1) The top-level decoding module integrates the most abstract core features of the bottleneck layer with the high-level semantic features of the third layer of the encoder to achieve the initial restoration of the global structure of the image; (2) The intermediate layer decoding module integrates the decoding features of the previous layer with the semantic features of the middle layer of the encoder's second layer to complete the local structural details of the image; (3) The bottom-level decoding module integrates the decoding features of the previous layer with the shallow edge texture features of the first layer of the encoder, and finally restores the spatial size of the feature map to the original input size. Through skip connections, it achieves full-level fusion of deep core features, mid-level semantic features and shallow detail features, avoiding the image blurring problem caused by simple upsampling. Finally, the model output layer adopts a combination design of 1×1 convolutional layer and bilinear interpolation. The 1×1 convolutional layer maps the 64-channel feature map output by the decoder to a 3-channel image without changing the spatial dimension of the feature map. Bilinear interpolation further fine-tunes the spatial size of the output feature map to ensure that it is consistent with the dimension of the input image. Therefore, the U-net model adopted in this paper constructs a ground-penetrating radar B-scan image denoising architecture that takes into account both noise suppression and detail preservation capabilities through hierarchical layer feature extraction and layer-by-layer noise filtering of the encoder, precise upsampling and full-level feature fusion of the decoder. Compared with the traditional encoding and decoding structure, the U-net model constructs a full-level, high-precision skip connection between encoding and decoding. After each round of upsampling in the decoder, the corresponding feature map from the encoder, after being centrally cropped to ensure a perfect size match, is concatenated with the current decoder feature map. This design ensures that the decoder can directly obtain the shallow detail features from the encoder during the same period when restoring the size, forming a precise complement to its own high-level semantic features, thus solving the problem of semantics and detail disconnect. Therefore, this embodiment uses the U-net network as the basic model, combined with a channel attention module, to more comprehensively and accurately separate effective signals and noise, realizing deep learning-based ground-penetrating radar B-scan image denoising. Figure 5 This is the U-net decoding and encoding structure. The entire noise reduction process can be represented as: ; In the formula, This indicates the added noise; Represents the original clean image; This represents the noisy image after adding noise, which serves as the input to the denoising network; Indicates network parameters; Represents a denoised image; This represents the feature map output by the encoder; This represents the feature extraction function of the encoder; This represents the image reconstruction function of the decoder.

[0023] In the above, the CBAM (Convolution Block Attention Module) is a hybrid attention mechanism, employing a serial dual-branch structure that combines channel-dimensional filtering and spatial-dimensional focusing. It includes channel attention units and spatial attention units.

[0024] The channel attention unit focuses on the channel dimension dependency of the feature map, assuming that different channels correspond to different feature semantics. In the improved U-Net denoising network model, some channels extract effective information, while others are dominated by noise. Therefore, by learning the importance weight of each channel, channels that contribute significantly to the denoising task are identified and given high weights, thereby strengthening effective feature channels and suppressing redundant noise-dominated channels, thus improving the signal-to-noise ratio of the feature map. Figure 7 This describes the structure of the channel attention unit. Its processing flow includes the following steps: (1) Perform global max pooling and global average pooling operations on the input feature map to compress spatial information, resulting in two channels with the following numbers: vector and : (2) The above statistical feature vectors and The input is a shared multilayer perceptron (MLP), which learns channel weights through a compression-activation-expansion process; (3) After adding the two weight vectors, normalize them using the Sigmoid activation function to obtain the final channel attention weight map. .

[0025] ; In the formula, and These represent the weight matrices for global max pooling and global average pooling, respectively. This represents the Sigmoid activation function.

[0026] The spatial attention unit focuses on the spatial dimensional dependencies of feature maps, assuming that features at different spatial locations contribute differently to the image. In ground-penetrating radar B-scan image denoising, features at the location of the lesion are more important than noisy regions. Therefore, by learning pixel-level weights, it focuses on key spatial regions and weakens interfering regions, achieving refined spatial selection of features. Figure 8 This describes the structure of a spatial attention unit. Its processing flow includes: (1) Feature map of spatial attention unit after channel attention weighting Perform max pooling and average pooling at the channel dimension, 3D feature map transformed into 3D feature map and ; (2) The two single-channel feature maps are spliced ​​along the channel dimension to obtain Feature map; (3) In passing 7 7. Convolution calculates the importance weights of each spatial location; (4) The spatial attention weight map is obtained by normalization using the Sigmoid activation function. .

[0027] ; Figure 9 For the network structure of the CBAM module, input feature map After processing by the channel attention unit of the CBAM module, a one-dimensional channel attention feature map is first output. The feature map is obtained by performing channel-by-channel multiplication with the original feature map to achieve feature filtering along the channel dimension. Feature map After spatial attention processing, a two-dimensional spatial attention feature map is generated. Finally, combine it with The final output feature map is obtained through element-wise multiplication. The entire process is shown in the following formula: ; Step S3: Train the improved U-Net model using the training dataset. Input noisy ground-penetrating radar B-scan images into the network, and perform forward propagation to obtain the denoised predicted images; use the clean image as the expected output, and calculate the mean squared error loss between the predicted image and the clean image; calculate the gradient of the loss with respect to the parameters of each layer through backpropagation, update the network weights using the optimizer, and iterate until the model converges to obtain the trained denoised model. Specifically, the model training parameters are set as follows: The loss function is a quantitative metric in deep learning that measures the difference between the predicted result and the true label. Its output is a non-negative real number. The smaller the loss value, the closer the model's prediction is to the reality; conversely, the larger the loss value, the greater the difference. The trend of the loss function is a core basis for evaluating the training effect: during training, if the training loss continuously decreases and tends to stabilize, it indicates that the model has gradually learned the mapping rule between noise and clean images, reaching a convergent state; if both the training loss and validation loss are high and decrease slowly, it indicates that the model has not fully learned effective features and is in an underfitting state, requiring measures such as increasing the number of training rounds. If the training loss continuously decreases, but the validation loss first decreases and then increases, it indicates that the model overfits the noise in the training set, has poor generalization ability, and requires data augmentation. In this embodiment, mean squared error loss (MSE) is used, and the loss function expression is: ; In the formula, Represents the pixel value of the image after noise reduction; Represents the pixel values ​​of a clean image. This represents the total number of pixels in the image.

[0028] Batch size represents the number of samples input to the model during each training iteration. Its primary function is to balance memory usage while ensuring gradient stability; its selection directly impacts the feasibility of model training. Batch size is generally a multiple of 2, with common values ​​being 2, 4, 8, 16, and 32. A value that is too small will lead to excessive iterations, prolonged training time, and gradient oscillations; a value that is too large will cause memory overflow, preventing successful execution. Therefore, this parameter needs to be flexibly adjusted in conjunction with device hardware adjustments and dataset size.

[0029] The learning rate (LR) is a core hyperparameter in deep learning training that controls the magnitude of model parameter updates. It is positively correlated with batch size: the larger the batch size, the more stable the gradient, and the larger the step size it can withstand. A larger learning rate allows for larger parameter adjustments each time, resulting in faster model learning; a smaller learning rate allows for finer adjustments, but slower learning. In this embodiment, the U-net model uses a learning rate of 0.01, which avoids gradient oscillations and vanishing gradients, ensuring a gradual improvement in feature extraction capabilities.

[0030] The number of training epochs is a core metric for measuring the adequacy of training in deep learning and is crucial for controlling underfitting and overfitting. When there are too few epochs, the model only learns some noise patterns and cannot adequately fit the general mapping of the training data, resulting in poor denoising in noisy regions and high training and validation losses. When there are too many epochs, after fitting the mapping, the model begins to force-fit specific noise patterns in the training set, causing the training loss to continuously decrease while the validation loss initially decreases and then increases. Therefore, during deep learning training, the epoch value can be preset based on the training set size and noise intensity. The trends of training and validation losses can be observed throughout the training process, and the epoch value can be adjusted based on these trends to ensure that the model can adequately learn the noise patterns of the training set at the optimal number of training epochs. Finding the optimal epoch value requires multiple adjustments and observation of the corresponding training and validation loss trends. However, each training session consumes a significant amount of time. To reduce this time cost, this paper introduces an early stopping mechanism in the model, which can prevent overfitting even with a larger epoch value.

[0031] In deep learning training, the optimizer is the core tool for implementing the gradient descent algorithm. It receives the gradient signal calculated from the loss function through backpropagation, combines it with hyperparameters, dynamically adjusts the model's learnable parameters, and ultimately minimizes the loss function, allowing the model to gradually learn the mapping relationship between noisy and clean images. This embodiment uses the Adam optimizer, which achieves adaptive step size adjustment and gradient ripple smoothing by dynamically estimating the first and second momentum of the parameters. Compared to traditional optimizers, Adam balances convergence speed, stability, and generalization ability, making it suitable for the U-Net image denoising task.

[0032] Since the simulated B-scan image resolution is 2500×1875, while the measured B-scan image resolution is 1910×528, the image resolution needs to be unified before training the network model. To ensure the model can train normally and to ensure that the resolution of each image in the dataset is not too low, which would lead to low learning efficiency, the image resolution was adjusted several times, and finally, a resolution of 1280×960 was adopted for network model training. The relevant parameters of the network model are: Batch Size of 4, Epoch of 100, Early Stopping Threshold of 20, LR of 0.0001, and the Adam optimizer is used. Figure 6 This is a flowchart of the network model training process.

[0033] Step S4: Input the actual collected ground-penetrating radar B-scan image to be processed into the trained noise reduction model. The model outputs a ground-penetrating radar image with a high signal-to-noise ratio, in which background noise is significantly suppressed and the hyperbolic reflection features of the disease are clear and complete.

[0034] After adding -15dB noise to simulated ground-penetrating radar images, the noise energy is far higher than the effective reflected signal energy of the disease. The hyperbolic reflectance characteristics of the disease target are submerged by the noise, resulting in an extremely low image signal-to-noise ratio. Traditional CEEMD-PE and CEEMD-WT joint denoising methods, relying solely on time-frequency analysis and statistical filtering, are insufficient to extract subtle disease features from a strongly interfering background. To demonstrate the denoising performance of the U-net network in extreme noise environments, this embodiment selects simulated data and field-measured data for denoising processing, and evaluates the model performance using metrics such as SNR, PSNR, and RMSE. Figure 10As shown, the B-scan image after denoising using the improved U-net model achieves good visual quality. Random noise and clutter in the background region are completely suppressed, resulting in a clean and uniform background with grayscale values ​​close to 0. This significantly improves the distinction between the disease target and the background, effectively separating noise from the disease reflection signal. Simultaneously, the improved U-net model effectively extracts disease reflection features through an encoder-decoder structure, highly preserving the hyperbolic edge contour during denoising, maintaining a smooth and continuous phase axis, and restoring the "black-white-black" phase characteristics. It avoids the blurring of disease edges or amplitude distortion commonly seen in traditional denoising methods.

[0035] Because the reflected signals from tiny cracks are weak, and the reflected signals from loose, porous defects are chaotic and dense, traditional algorithms struggle to recover phase information when processing these two types of defects. The improved U-net model effectively separates the effective signal from the deep feature space and reconstructs the detailed features of the target hyperbola. For the fixed vertices and extended wing portions of the defect hyperbola, the improved model accurately extracts edge contours from a noisy background. Even in complex scenarios with overlapping defects, the improved model can still efficiently separate individual defect features, demonstrating that the network model can effectively learn the nonlinear structural characteristics of the effective signal from ground-penetrating radar defects during noise reduction, rather than simply using spectral filtering.

[0036] To further objectively evaluate the denoising performance of the improved U-net model, Table 1 shows the quantitative evaluation indicators for different types of diseases after denoising using the improved model under a -15dB Gaussian noise environment. Here, SNR is the signal-to-noise ratio, PSNR is the peak signal-to-noise ratio, RMSE is the root mean square error, and SSIM is the structural similarity index. Traditional denoising methods achieve a maximum SNR of only 12dB after adding 1dB of noise. However, the improved U-net model in this embodiment achieves high SNRs for all denoised images under a -15dB noise input. Specifically, the SNR for a single crack disease reaches 25.39dB, and even for complex, non-dense diseases, the SNR reaches 22.18dB. Therefore, the improved U-net model can effectively extract weak disease signals under strong noise interference. Traditional algorithms tend to destroy the phase information of signals, resulting in low similarity in brightness, contrast, and structure between denoised and clean images. However, after denoising using the improved model, the SSIM values ​​of different types of disease images all exceed 0.98, with single voids and cracks achieving SSIM values ​​of 0.9918 and 0.9919 respectively. This indicates that the denoised image is almost identical to the original clean image. This high-fidelity feature reconstruction provides a strong basis for subsequent intelligent disease identification. Simultaneously, the RMSE of different disease types is below 0.0212. The RMSE for a single crack is 0.0135, further validating the model's pixel-level signal recovery capability and maximizing the preservation of amplitude values ​​in the original clean image. The improved model also demonstrates effectiveness and robustness in complex disease combination scenarios. For three different disease combinations—voids, cracks, and looseness—the denoised image maintains an SNR of 22.42 dB and an SSIM of 0.9849. Therefore, the improved model does not experience performance degradation due to an increase in the number of disease targets or the complexity of the scenario, demonstrating strong generalization ability.

[0037] Table 1. Noise Reduction Evaluation Indicators for Simulated Single Diseases and Combined Diseases

[0038] To further verify the generalization ability and practicality of the improved U-net model in actual tunnel engineering environments, this embodiment selects two sets of representative field-measured ground-penetrating radar B-scan data for verification. The raw data was processed using the mature radar data processing software IDSP7, including DC removal, background elimination, and digital filtering, to produce a clean reference image. The unprocessed raw B-scan image acquired in the field was used as the noise map input, thus forming a pair of corresponding input labels to evaluate the noise reduction performance of the improved U-net model on the measured data. Figure 11 (a) Figure 11 (b) and Figure 11 (c) shows the standard reference image, the original noisy image, and the denoised image after processing with the improved U-net model for the first set of measured data.

[0039] Because of the presence of electrical equipment such as lighting systems, ventilation systems, and communication cables at the tunnel site, electromagnetic radiation is generated during operation. This radiation couples with the electromagnetic waves emitted by radar equipment, resulting in equipment interference noise in the raw data. Furthermore, since the tunnel lining concrete is a non-homogeneous medium, electromagnetic waves undergo random scattering as they propagate within it, causing the scattered waves to interfere with each other during propagation. Therefore, in Figure 11 In (b), the original B-scan profile image is filled with numerous random black and white spots resembling Gaussian noise, obscuring the edge contours of the effective disease reflection signal, resulting in a low signal-to-noise ratio. However, the improved U-net model demonstrates good feature extraction capabilities and anti-interference performance in the face of random scattering caused by non-uniform media and random black and white spot noise caused by external equipment interference in the measured data. For example... Figure 11 As shown in (c), the image after noise reduction filters out a large amount of background clutter, the background gray value is close to zero, the boundary contour of the detached disease signal reflection feature is clear after noise reduction, the phase feature is preserved, there is no phenomenon of signal amplitude attenuation, and the disease feature is significantly highlighted.

[0040] Figure 12 The image presents the standard reference image, the original noisy image, and the denoising results after processing the second set of measured data using radar software. The measurement line for the second set of measured data is located in a densely reinforced concrete lining area. Due to the high dielectric constant of the metal and the large difference in parameters between it and the surrounding medium, the steel mesh exhibits total internal reflection of radar electromagnetic waves. This results in continuous reflection signals from the steel mesh with extremely high energy in the shallow part of the original noisy image, accompanied by strong ringing effects and multiple wave interference below. These high-energy clutter waves cause strong interference superposition of the hyperbolic reflection curves of adjacent steel bars in both the horizontal and vertical directions, blurring the characteristics of individual steel bars and making waveform separation difficult using conventional filtering methods. The image after denoising using the U-net model is shown below. Figure 12 As shown in (c), the intertwined hyperbolic features of the reinforcing bars in the noisy image are clearly restored to independent, well-defined hyperbolic features. This successfully suppresses ringing and trailing noise around the reflected signals of the reinforcing bars, and converges the energy that was originally diverged and aliased due to diffraction. This significantly improves the resolution of densely reinforced targets and is highly consistent with the standard reference image after fine processing by radar software. Therefore, the improved model can not only effectively distinguish between target signals and clutter interference, but also focus on the signal characteristics of the target, effectively overcoming the distortion of target signals caused by excessive noise reduction.

[0041] To further verify the effectiveness of the improved U-net model in actual tunnel engineering scenarios, Table 2 shows the quantitative evaluation indicators of the two sets of measured data after denoising. The SNR of both sets of measured data after denoising is above 20dB. The SNR value of measured data one is slightly higher, mainly because the ringing noise caused by dense rebar in measured data two is high-energy coherent noise, which is more difficult to remove than the random noise generated by scattering. However, the signal-to-noise ratio after denoising can still reach 20.95dB, without significant energy attenuation. This indicates that the improved model can also effectively suppress high-energy ringing interference and maintain the energy ratio of the effective signal at a good level. PSNR is mainly used to measure the ratio of the highest signal power to the noise power. The PSNR values ​​of both sets of measured data exceed 36dB. Measured data two, even with extremely strong rebar shielding, still achieves a peak signal-to-noise ratio of 36.31dB, effectively suppressing the ringing effect while preserving the peak intensity information of the rebar reflected wave to the greatest extent. The root mean square error (RMSE) of both sets of measured data was controlled at a low level. The RMSE of measured data two was 0.0153, slightly higher than that of measured data one. This is mainly due to the difference in signal amplitude. Measured data three had a large amount of dense steel mesh along the survey line, resulting in a high amplitude reflected signal. Therefore, even small waveform fitting deviations were amplified in the RMSE calculation. The structural similarity (SSIM) values ​​of both sets of measured data were above 0.97, indicating that the model's ability to extract the overall structure and texture of the image was not affected by random noise and clutter interference in actual tunnel engineering, demonstrating the robustness of the improved model in complex real-world environments.

[0042] Table 2. Noise Reduction Evaluation Index Results of Two Sets of Measured Data

[0043] The evaluation results above show that, for simulation data with added -15dB high-intensity noise, the improved U-net model effectively separates the defect signal under strong noise interference in both single and combined defect scenarios, accurately recovering the phase characteristics of the defect hyperbola without signal distortion, amplitude attenuation, or edge blurring. The SNR values ​​of multiple sets of simulation data all exceed 21dB, RMSE is below 0.022, and SSIM values ​​exceed 0.98. For two sets of measured data with complex backgrounds and dense rebar interference, the improved model effectively suppresses random scattering caused by non-uniform media and external electromagnetic interference noise. The reconstructed SNR values ​​for the two sets of measured data are 22.16dB and 20.95dB, respectively. Particularly for concrete areas with dense rebar, the improved model effectively suppresses ringing effects and multiple wave interference caused by rebar reflection, significantly improving target resolution and demonstrating good generalization ability in practical engineering.

[0044] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for ground penetrating radar high signal-to-noise ratio image imaging, characterized in that, Includes the following steps: S1: Construct a training dataset, which includes multiple sets of pairs of noisy ground-penetrating radar B-scan images and corresponding clean images; S2: Construct an improved U-Net denoising network model, which includes an encoder, a decoder, a skip connection connecting the encoder and the decoder, and CBAM attention modules embedded in the encoder and decoder. S3: The improved U-Net denoising network model is trained using the training dataset. The noisy ground-penetrating radar B-scan image is used as input, and the clean image is used as the desired output. The network parameters are optimized through backpropagation to obtain the trained denoising model. S4: Input the ground-penetrating radar B-scan image to be processed into the trained noise reduction model, and output a ground-penetrating radar image with a high signal-to-noise ratio.

2. The ground penetrating radar high signal-to-noise image imaging method of claim 1, wherein, In step S1, the noisy ground-penetrating radar B-scan image is generated by adding noise with a signal-to-noise ratio of -15dB to a clean image, or by directly acquiring the original unprocessed ground-penetrating radar B-scan image from the tunnel site as the noisy ground-penetrating radar B-scan image; the clean image is an image processed by background removal, digital filtering, and gain adjustment of the measured data; the training dataset includes four types of defects: voids, looseness, cracks, and voids.

3. The high signal-to-noise ratio image imaging method for ground-penetrating radar according to claim 2, characterized in that, In step S1, the noisy ground-penetrating radar B-scan image and the corresponding clean image are uniformly cropped or scaled to a fixed resolution, and divided into training set, validation set and test set according to an 8:1:1 ratio.

4. The high signal-to-noise ratio image imaging method for ground-penetrating radar according to claim 3, characterized in that, The encoder in step S2 consists of multiple cascaded downsampling modules, each of which includes two consecutive 3×3 convolutional layers, a ReLU activation function, and a 2×2 max pooling layer; the decoder consists of multiple cascaded upsampling modules, each of which includes a 2×2 transposed convolutional layer, a feature concatenation layer, two consecutive 3×3 convolutional layers, and a ReLU activation function.

5. The high signal-to-noise ratio image imaging method for ground-penetrating radar according to claim 4, characterized in that, The skip connection concatenates the feature map of the corresponding layer of the encoder with the upsampled feature map of the decoder along the channel dimension after central cropping, thereby achieving the fusion of shallow detailed features and deep semantic features.

6. The high signal-to-noise ratio image imaging method for ground-penetrating radar according to claim 1, characterized in that, The CBAM attention module in step S2 includes a cascaded channel attention unit and a spatial attention unit. The channel attention unit is configured to perform the following steps: The input feature map is subjected to global max pooling and global average pooling respectively, and the resulting two channel vectors are input into a shared multilayer perceptron. The two output vectors of the multilayer perceptron are added together and then normalized by the Sigmoid activation function to obtain the channel attention weight map; The channel attention weight map is multiplied with the input feature map channel by channel to achieve feature filtering in the channel dimension, and an intermediate feature map is obtained. The spatial attention unit is configured to perform the following steps: The feature maps are subjected to both channel-dimensional max pooling and average pooling. The two single-channel feature maps are concatenated along the channel dimension; The importance weights of each spatial location are calculated using 7×7 convolution; The spatial attention weight map is obtained by normalizing using the Sigmoid activation function; The spatial attention weight map is then multiplied element-wise with the intermediate feature map to obtain the final output feature map.

7. The high signal-to-noise ratio image imaging method for ground-penetrating radar according to claim 1, characterized in that, In step S3, the mean squared error loss function is used as the training loss, and the loss value is calculated with a clean image as the desired output.

8. The high signal-to-noise ratio image imaging method for ground-penetrating radar according to claim 1, characterized in that, Optimizing network parameters through backpropagation includes: calculating the gradient of the loss function with respect to the network parameters and updating the network parameters using the optimizer; and employing an early stopping mechanism during training to prevent overfitting.

9. The high signal-to-noise ratio image imaging method for ground-penetrating radar according to claim 1, characterized in that, The method also includes using signal-to-noise ratio, peak signal-to-noise ratio, root mean square error, and structural similarity index as evaluation indicators to assess the difference or similarity between the denoised image and the clean image, so as to quantify the denoising effect.

Images