Suppression of Multiple Reflection Interference Signals in GPR Images Caused by Rebar Using VAE-GAN

Abstract

Due to the rebars layer’s shielding effect on Ground Penetrating Radar (GPR) waves, the hyperbolic clutter generated by the rebars interferes with the echoes from void beneath them. The overlapping waveforms of both signals result in attenuation and distortion of the void signals, making it difficult to identify void defects under the rebar. This study proposes an unsupervised generative network model based on Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs). Through a shared latent space, mapping is achieved between two image domains, effectively eliminating the multiple reflection interference signals caused by the rebar while accurately reconstructing the void defects, generating GPR B-Scan images without rebar clutter. Additionally, the channel and spatial attention module (CSA) is implemented into the model to help the network to better focus on the essential information in GPR images. The proposed model was validated through ablation and comparative experiments using synthetic data. Finally, real GPR data from the Husa Tunnel were used to verify the model’s effectiveness in practical engineering applications. The results showed that this model is highly effective; it improves the visibility of void defects signals, thereby enhancing the interpretability of GPR data for tunnel lining inspections.

1. Introduction

In the inspection of tunnel linings, the presence of voids represents a common and serious structural defect. These voids not only compromise the load-bearing capacity of the structure but also pose a risk of further structural deterioration, potentially leading to severe damage [1,2,3]. Therefore, accurately identifying and assessing voids in tunnel linings is of critical importance. Ground Penetrating Radar (GPR) is a widely used non-destructive testing (NDT) technology that utilizes high-frequency electromagnetic waves to detect subsurface structures. A typical GPR system consists of a transmitting and a receiving antenna. The transmitting antenna emits electromagnetic pulses into the structure, and when these waves encounter interfaces between materials with different dielectric properties—such as air voids, concrete, or rebars—part of the wave is reflected back to the receiving antenna, while the remainder continues to propagate. The received signals are then processed to generate radar images, which provide valuable insights into the internal structure of the tunnel lining and its surrounding environment [4]. However, GPR performance is significantly affected by environmental conditions. The penetration depth, resolution, and signal attenuation are influenced by the dielectric properties of the materials, with high moisture content and heterogeneous structures often degrading detection accuracy. A particularly significant challenge in GPR-based void detection is the strong electromagnetic reflections caused by rebars in tunnel linings. These reflections create prominent hyperbolic clutter in GPR images, which can obscure signals from voids beneath the rebars, making detection highly challenging [5,6]. To improve the accuracy of void detection and facilitate automated interpretation of GPR data, it is essential to develop effective techniques for suppressing rebar-induced interference. Enhancing interference mitigation methods will contribute significantly to improving the reliability of GPR-based inspections, making void detection in tunnel linings more precise and efficient [7,8,9].

The goal of rebar interference signal suppression is to minimize the interference of rebar signals on defect signals, allowing for clear identification of defect signals. Traditional identification methods for defective signals primarily rely on handcrafted filtering techniques or transform-based approaches, such as wavelet decomposition or Hyp-curvelet transform. Wu et al. [10] proposed a method for detecting and suppressing rebar signals. They used Hyp-curvelet transform to project the GPR time-space 2D signals into the scale space. By combining the initial phase of the defect signals under the rebar as well as the time-frequency characteristics, they eliminated the rebar signal components in the scale space, achieving rebar signal suppression by transforming the data back to the time-space domain. Terrasse et al. [11] introduced a method for clutter suppression that utilizes the curvelet transform. By combining coefficient distribution and prior information such as clutter direction, they used curvelet transform in the clutter, noise, and artifact removal process, reducing noise and clutter artifacts in the data and improving the readability of GPR data during the buried pipeline localization process. Ma et al. [12] proposed a noise reduction approach that combines complete ensemble empirical mode decomposition with wavelet decomposition. This approach can remove effective information within the intrinsic mode function (IMF) components during the de-noising process. Ge et al. [13] proposed a method based on iterative F-k migration and demigration, combined with a mask window and classic GPR data processing steps, to remove rebar clutter in concrete structures, validated by numerical simulations and field experiments. Zhang et al. [14] proposed a rebar clutter suppression method based on range migration compensation and low-rank sparse decomposition with total variation regularization for through-the-wall radar (TWR). The method locates rebars using Hough transform, transforms rebar echoes into a low-rank subspace via range migration compensation, and extracts target signals through inverse range migration. Numerical simulations and experiments demonstrate its superiority in a target-to-clutter ratio (TCR). However, these traditional approaches often suffer from inefficiencies in processing large-scale data and require expert interpretation, limiting their practical applicability.

With the advancement of deep learning technology, new opportunities have emerged for GPR-based rebar defect identification in underground structure monitoring, making it a promising trend for the future. Deep learning models enable the automatic recognition and classification of underground structural features, such as rebars and voids, significantly improving data processing speed, accuracy, and reliability. In particular, under strong rebar signal interference, these models can learn distinctive signal features from large datasets, effectively separating rebar and void signals. However, many existing deep learning-based methods rely on supervised learning, which requires a large amount of paired training data—data that are often difficult to obtain in real-world applications. Kang et al. [15] proposed a background filtering algorithm based on the basis pursuit to eliminate dominant linear features and extraneous pattern noise in B-scan images. They then trained a pre-trained AlexNet network model with a 2D grid image composed of several B and C scan images to detect and classify four different types of cavities. Kim et al. [16] considered the statistical distribution of GPR data and automatically determined the amplitude of feature enhancement to enhance the target signal’s features, improving the identification accuracy of B-scan images. Wang et al. [17] proposed a supervised approach for rebar clutter removal and defect echo enhancement, achieving both efficient clutter suppression and precise defect reconstruction. Using this approach, the accuracy of identifying defects under rebar increased from 0.208 to 0.850. Despite these advances, supervised methods remain constrained by the availability of high-quality labeled data, making unsupervised learning an attractive alternative.

Deep generative models have demonstrated remarkable capabilities in cross-domain feature learning, effectively modeling complex data distributions to achieve high-quality feature transformation and mapping between different visual domains. Their ability to generate realistic transformations makes them particularly suitable for unsupervised learning, where paired training data are often unavailable [18]. This enables effective suppression of rebar clutter while preserving defect signals in GPR images. Wang et al. [19] proposed an unsupervised rebar clutter elimination network (RCE-GAN) designed for detecting void in tunnel linings. The network comprises two sets of generators and discriminators, and by incorporating an attention module and a dilation center component, it enhances the effectiveness of rebar clutter removal. Wang et al. [20] proposed a rebar clutter suppression approach based on GANs to suppress rebar signals in different defect structures. By incorporating four contrasting encoders and two similarity encoders, they effectively suppressed rebar responses and precisely reconstructed defect signals. Dai et al. [21] proposed a VAE and RefineNet network algorithm based on a hybrid loss function for GPR clutter suppression and target imaging. The algorithm has a strong capability to capture detailed features and can significantly improve the quality of target imaging. Lei et al. [22] proposed a DR-GAN network for clutter suppression in GPR B-scan images. It extracts target and clutter features using dedicated encoders and employs a generator to suppress clutter. Unlike supervised methods, DR-GAN does not require paired data for training. Experiments show that it outperforms the RNMF method, improving IF by 17.85 dB. Wang et al. [23] introduced a residual channel attention network (RCAN) that leverages a channel-based attention mechanism to suppress rebar clutter, effectively revealing hidden defects. Ge et al. [24] proposed a Wavelet-GAN network that combines GAN and DWT to remove clutter and noise from GPR images. This method utilizes DWT for decomposition, CNN and GAN for signal reconstruction, and IDWT for generating high-quality GPR images. It demonstrates excellent performance in training on small datasets and processing real-world data. Ren et al. [25] proposed an end-to-end unsupervised deep learning-based network to suppress rebar signals and enhance void defect identification in GPR images. The network integrates feature attention and mix-up modules, optimizes image reconstruction through feature space comparison and cycle perceptual consistency loss, and improves Faster-RCNN for detecting void defects beneath rebars. The method’s effectiveness was validated using synthetic and real-world datasets.

In this study, we propose an unsupervised learning-based method for suppressing rebar interference in GPR images, enhancing void detection in tunnel linings by learning feature relationships between two sets of incompletely corresponding images. This achieves the transformation from GPR images containing rebar interference to those without it, reducing the interference of rebar clutter with void defect signals and enabling more accurate identification of void signals in the images. The proposed method is based on UNIT [26]. We design a channel and spatial attention (CSA) module for the network to help it better concentrate on critical information in GPR images, allowing for more accurate distinction between rebar clutter and void signals. Specifically, the inclusion of CSA allows the network to consider relationships between different channels and spatial positions simultaneously when processing images. In complex situations with overlapping features, this allows the network to adjust its focus on particular channels and spatial regions, depending on the distribution patterns of rebar interference and void signals. This enables the network to better distinguish between and process these two structures and specifically suppress rebar interference signals, thereby more accurately reconstructing complete void defect signals. Experimental results demonstrate that our method effectively suppresses rebar clutter while maintaining the integrity of defect signals, making it well suited for practical engineering applications.

2. Materials and Methods

2.1. Overall Network Structure

The overall network model consists of a generative network based on Variational Autoencoders (VAEs) and a discriminative network based on Generative Adversarial Networks (GANs). It comprises two domain-specific image encoders, $E_A$ and $E_B$, two domain-specific image generators, $G_A$ and $G_B$, and two domain-specific image adversarial discriminators, $D_A$ and $D_B$. The encoder–generator pair $\{E_A,G_B\}$ is used to learn the mapping between two GPR image domains, A to B. The encoder $E_A$ is responsible for mapping GPR images with rebar clutter into shared latent space z, and the generator $G_B$ maps z back to the output data space, generating images without rebar clutter. The discriminator $D_B$ ensures the generation of more realistic images without rebar clutter. The primary objective of the model is to generate images without rebar clutter from GPR images with rebar clutter interference. The overall network structure is illustrated in Figure 1.

The generative network includes two sets of encoders and generators, corresponding to two different image domains, domain A, which contains GPR images with rebar clutter, and domain B, which contains GPR images without rebar clutter. $E_A$ and $G_A$ are the encoder and generator for domain A, referred to as ${\mathrm{VAE}}_1$. Similarly, $E_B$ and $G_B$ are the encoder and generator for domain B, referred to as ${\mathrm{VAE}}_2$. The generative network achieves the mapping relationship between the two VAE by a shared latent space. Specifically, the encoders are tasked with extracting high-level feature representations from the input images of both domains and encoding them into a latent space of the same probability distribution. This latent space can be viewed as a "bridge" containing key information from both domains. Similarly, the generators decode from the latent space to reconstruct the feature representations of the input images. This design enables the two VAEs to jointly learn the shared feature representations between the two domains. In ${\mathrm{VAE}}_1$, the encoder $E_A$ can extract features of rebar signals and void signals from the input images, which are mapped into the shared latent space z. On the side of generator $G_A$, z is used to reconstruct the images. Similarly, in ${\mathrm{VAE}}_2$, the encoder $E_B$ can also learn the features of void signals from the input images, which are mapped into z, allowing the generator $G_B$ to reconstruct the images. This design allows the model to perform mapping translations between the two image domains A to B and B to A. Images from the two domains can be mapped into the shared latent space z through encoders $E_A$ and $E_B$, respectively, and then mapped back to their respective domain’s output data space through generators $G_A$ and $G_B$. Each domain’s generator can produce two types of images: reconstructed images within their domain and translated images from the other domain.

For a GPR image $x\in A$ with rebar clutter, the encoder $E_A$ first maps x into z. Subsequently, both sets of generators decode z. The primary objective of generator $G_A$ is to reconstruct image ${\tilde{x}}^{A\to A}$ with rebar clutter that closely resembles the original image x, emphasizing high consistency with GPR images from the original domain A. Generator $G_B$, on the other hand, aims to generate a clear image ${\tilde{x}}^{A\to B}$ without rebar clutter interference. This means that $G_B$ can extract features related to rebar clutter from z to eliminate clutter interference in the original image, producing a clear image containing only void defects.

The structure of the generator is depicted in Figure 2a. The encoder consists of three convolutional layers at the front end. These layers progressively decrease the dimensions of the image while expanding the number of channels to extract image features. Following this, there are four basic residual blocks in the backend, which aim to capture higher-level features and finer patterns in the image. The generator mirrors the encoder structure. It starts with four basic residual blocks at the front end, which receive the representation from the shared latent space and reconstruct features. These residual blocks are responsible for restoring details and textures of the image from the latent space. Subsequently, three transposed convolutional layers are used in the backend to progressively increase the dimensions of the feature maps while compressing the number of channels in order to restore the image to the input size and generate the final output image.

The discriminator network includes discriminators $D_A$ and $D_B$. In $D_B$, the discriminator $D_B$ output TRUE for real images is sampled from domain B, while its output FALSE for images is generated by $G_B$. $G_B$ generates two types of images: reconstructed images ${\tilde{y}}^{B\to B}$ and translated images ${\tilde{x}}^{A\to B}$. Since reconstructed images can be trained with supervision, only the translated images are used for adversarial training. A similar approach is applied to $D_A$: discriminator $D_A$ output TRUE for real images is sampled from domain A, while its output FALSE for images is generated by $G_A$.

The structure of the discriminator is depicted in Figure 2b. It consists of six convolutional layers, which progressively decrease the dimensions of the input image while increasing the depth and complexity of the features. Notably, the network processes the input image through different scales of downsampling operations, each handled by the same discriminator network, resulting in three different scales of feature outputs: 16 × 16, 8 × 8, and 4 × 4. Each specific scale’s output independently performs real vs. fake discrimination. This enables the network to assess the authenticity of the image from multiple dimensions, with each scale providing a different perspective and information, thereby enhancing the model’s discriminative ability.

2.2. CSA Module

Due to its metallic properties, rebar generates strong, high-amplitude reflective signals in GPR images, typically appearing as bright, high-contrast lines or points. In contrast, voids produce weaker reflective signals with lower amplitude, appearing as darker areas or lighter lines. To better suppress rebar clutter and achieve more effective reconstruction of void signals, the channel and spatial attention module (CSA) was designed in the encoder–generator pairs $\{E_A,G_A\}$, $\{E_A,G_B\}$, $\{E_B,G_A\}$, and $\{E_B,G_B\}$, as shown in Figure 2a. The CSA module is inspired by the Convolutional Block Attention Module (CBAM), which integrates both channel and spatial attention mechanisms. Similarly, CSA adopts this dual attention strategy to optimize feature extraction and processing from both the channel and spatial dimensions [27]. The module inputs features from two directions: high-level semantic information from the encoder that contains complex structures such as rebar and void, and the other includes features reflecting the gradual recovery process from the generator. This design is intended to simultaneously process and integrate features from both the encoder and the generator, helping the network more accurately identify and eliminate rebar clutter while preserving the integrity of critical structures like voids during reconstruction.

Channel attention primarily focuses on weighting feature channels. Some channels might be better at capturing rebar reflection signals while others may be more sensitive to voids or other structural features. This module learns the weights for each channel, identifying the most critical channel features for the current task. High-weight channels indicate that their features are particularly important for distinguishing rebar and non-rebar regions. By assigning higher weights to important channels, their feature representation is enhanced, allowing these crucial details to be better restored and utilized in subsequent decoding processes. In contrast, spatial attention focuses on the spatial positions, weighting these positions to highlight areas more important for the task. This means the model can distinguish between and focus on specific spatial regions containing rebar signals. By assessing the importance of each position in the image, it highlights regions containing important information such as voids. Spatially weighted regions receive more attention during feature extraction and image recovery stages, ensuring that the features in these areas are effectively preserved and enhanced in the final output image.

The structure of the CSA module is depicted in Figure 3a. The network begins with the channel attention module, as shown in Figure 3c. First, the input features $F_1$ and $F_2$ are subjected to average pooling and max pooling operations, respectively. Max pooling emphasizes key features in the image, while average pooling helps retain more comprehensive background information. The pooled features are then concatenated to integrate information from both inputs, generating two new feature vectors from the average and max pooled features. These vectors are then passed through a shared layer (Shared MLP), which first compresses the number of channels to 1/4 of the original size, applies a ReLU activation function, and then expands them back to the original channel size. The two resulting outputs use element-wise summation, and then weight $M_C$ for each channel is output via a Sigmoid activation function. These weights are then element-wise multiplied with $F_2$ to produce $F_2^{\prime }$, achieving weighted channel features for the input feature. The weighted features $F_2^{\prime }$ and the original generator input features $F_1$ are fed into the spatial attention module for further spatial-level weighting, as shown in Figure 3b. First, the dimensionality of the input features is reduced separately through convolutional blocks with a kernel size of 1 × 1. The features then use element-wise summation to integrate important feature information. The combined features are passed through a ReLU activation function, followed by another convolutional block, and then weight $M_S$ for each spatial position is output via a Sigmoid activation function. These weights are then element-wise multiplied with $F_2^{\prime }$ processed by the channel attention to produce $F_2^{″}$, achieving weighted spatial features for the input feature. By sequentially applying channel and spatial attention modules, the CSA module helps the network focus more on key feature information, enhancing the network’s expressive power.

2.3. Loss Function

The loss function of the model consists of the following components: VAE loss, adversarial loss, and cycle consistency loss.

The VAE loss includes both the reconstruction loss and the Kullback–Leibler (KL) divergence. The former aims to maintain content consistency between the generated images and the original input images by using a pixel-level loss function to measure the differences between them. The latter aims to minimize the difference between the posterior distribution of the generated latent variables and the standard normal distribution. This encourages the encoder to learn the latent space with good representation, enabling the generator to produce realistic images. The formulas for VAE loss are shown in Equations (1) and (2):

$$\begin{array}{cc}\hfill L_{{\mathrm{VAE}}_1}(E_A,G_A)& ={\lambda }_0\mathrm{KL}\left(q_1\left(z_A\right|x)\parallel p_{\eta }\left(z\right)\right)-{\lambda }_1{\mathbb{E}}_{z_A\sim q_1\left(z_A\right|x)}\left[log{p}_{G_A}\left(x\right|z_A)\right]\hfill \end{array}$$

(1)

$$\begin{array}{cc}\hfill L_{{\mathrm{VAE}}_2}(E_B,G_B)& ={\lambda }_0\mathrm{KL}\left(q_2\left(z_B\right|y)\parallel p_{\eta }\left(z\right)\right)-{\lambda }_1{\mathbb{E}}_{z_B\sim q_2\left(z_B\right|y)}\left[log{p}_{G_B}\left(y\right|z_B)\right]\hfill \end{array}$$

(2)

The adversarial loss forces the generator to learn the mapping relationship between the two image domains. During training, the discriminator optimizes its ability to classify real and generated images correctly by maximizing the probability of correctly identifying real images and generated ones. The generator optimizes its image generation ability by minimizing the discriminator’s probability of recognizing the generated images, making the generated images visually resemble the real images. The formulas for the adversarial loss are shown in Equations (3) and (4):

$$\begin{array}{cc}\hfill L_{{\mathrm{GAN}}_1}(E_B,G_A,D_A)& ={\lambda }_2{\mathbb{E}}_{x\sim p\left(x\right)}\left[log{D}_A\left(x\right)\right]+{\lambda }_2{\mathbb{E}}_{z_B\sim q_2\left(z_B\right|y)}\left[log(1-D_A\left(G_A\left(z_B\right)\right))\right]\hfill \end{array}$$

(3)

$$\begin{array}{cc}\hfill L_{{\mathrm{GAN}}_2}(E_A,G_B,D_B)& ={\lambda }_2{\mathbb{E}}_{y\sim p\left(y\right)}\left[log{D}_B\left(y\right)\right]+{\lambda }_2{\mathbb{E}}_{z_A\sim q_1\left(z_A\right|x)}\left[log(1-D_B\left(G_B\left(z_A\right)\right))\right]\hfill \end{array}$$

(4)

The cycle consistency loss ensures that during the mapping between image domains, the generated image can return to the original domain and should be as close as possible to the original image. This loss helps preserve important features and structures of the original image, enhancing the quality and stability of image translation. The formulas for the cycle consistency loss are shown in Equations (5) and (6):

$$\begin{array}{cc}\hfill L_{{\mathrm{cyc}}_1}(E_A,G_A,E_B,G_B)& ={\lambda }_3\mathrm{KL}\left(q_1\left(z_A\right|x)\parallel p_{\eta }\left(z\right)\right)+{\lambda }_3\mathrm{KL}\left(q_2\left(z_B\right|x^{A\to B})\parallel p_{\eta }\left(z\right)\right)\hfill \\ \hfill & -{\lambda }_4{\mathbb{E}}_{z_B\sim q_2\left(z_B\right|x^{A\to B})}\left[log{p}_{G_A}\left(x\right|z_B)\right]\hfill \end{array}$$

(5)

$$\begin{array}{cc}\hfill L_{{\mathrm{cyc}}_2}(E_B,G_B,E_A,G_A)& ={\lambda }_3\mathrm{KL}\left(q_2\left(z_B\right|y)\parallel p_{\eta }\left(z\right)\right)+{\lambda }_3\mathrm{KL}\left(q_1\left(z_A\right|y^{B\to A})\parallel p_{\eta }\left(z\right)\right)\hfill \\ \hfill & -{\lambda }_4{\mathbb{E}}_{z_A\sim q_1\left(z_A\right|y^{B\to A})}\left[log{p}_{G_B}\left(y\right|z_A)\right]\hfill \end{array}$$

(6)

The formula for the total loss is obtained, as shown in Equation (7). The weight values for each part are ${\lambda }_0$ = ${\lambda }_3$ = 0.01, ${\lambda }_1$ = ${\lambda }_4$ = 10, and ${\lambda }_2$ = 1.

$$\begin{array}{cc}\hfill L_{\mathrm{total}}& =L_{{\mathrm{VAE}}_1}(E_A,G_A)+L_{{\mathrm{VAE}}_2}(E_B,G_B)+L_{{\mathrm{GAN}}_1}(E_B,G_A,D_A)\hfill \\ \hfill & +L_{{\mathrm{GAN}}_2}(E_A,G_B,D_B)+L_{{\mathrm{cyc}}_1}(E_A,G_A,E_B,G_B)+L_{{\mathrm{cyc}}_2}(E_B,G_B,E_A,G_A)\hfill \end{array}$$

(7)

3. Validation on Synthetic GPR Data

3.1. Data Preparation and Network Training

Currently, the available real GPR image data are limited and inadequate for model training. This study used the Finite-Difference Time-Domain (FDTD) method for forward numerical simulation to generate simulated GPR data of void defects [28]. According to the different characteristics of various media, through the interaction between electromagnetic waves and different media, two sets of GPR simulated images were obtained: one with rebar clutter and the other without. It is important to note that these two sets of images are unpaired. The relative permittivity and conductivity of various media are shown in

Table 1. The relative permittivity and conductivity of various media.

Media	Relative Permittivity	Conductivity (S/m)
Rebar	300	$1.0\times {10}^8$
Concrete	8∼10	$1.0\times {10}^{-4}$
Water	81	$1.0\times {10}^{-3}$
Air	1	0

. The model grid size was set to 2.08 m (width) × 1.08 m (depth) with a grid spacing of 0.004 m. The central frequency of the transmit antenna was 800 MHz, with the receive antenna located 0.1 m from the transmit antenna. The time window was 20 ns. During the simulation process, the antenna moved in steps of 0.008 m, performing 240 scans per B-scan image. To reduce electromagnetic losses and improve numerical simulation accuracy, Perfectly Matched Layers (PMLs) were placed around the model boundaries for absorbing conditions [29].

To enhance the authenticity of the synthetic data and align with the actual tunnel construction lining standards, this study’s simulation model followed standard tunnel lining design specifications for rebar placement. By reasonably adjusting the spacing and positioning of the rebars, it aimed to match the common rebar arrangement patterns observed in real-world scenarios.

We placed rebars layer 0.1 m below the air layer, with a spacing of 0.2 m between the rebars, and randomly generated positions and shapes of defects beneath the rebars. The defects were filled with air or water, with concrete serving as the background fill. Direct waves were removed from the GPR images generated after simulation, as shown in Figure 4. Ultimately, we obtained a dataset of 600 pairs of non-paired GPR images, which included GPR images containing echoes of void defects and rebar clutter, as well as images containing only echoes of void defects. These images were split into training and testing datasets in an 8:2 ratio, with each image sized 256 × 256.

The model was trained on a Linux server equipped with an NVIDIA RTX 3090 (24 GB) GPU (NVIDIA Corporation, Santa Clara, CA, USA). The framework for implementing is Pytorch. The total number of iterations for model training was set to 10, 000, utilizing the Adam optimizer [30]. Due to the limited training dataset, we did not use randomly initialized parameters to train the network initially. In the first 3000 iterations, we pre-trained the network using the simulated dataset from Wang et al. [19]. Subsequently, we trained the network with our dataset based on this pre-training.

3.2. Comparative Experiments

To validate the effectiveness of the proposed model in suppress rebar clutter, three widely used unsupervised learning models were chosen as baselines (UNIT, CycleGAN [31], DualGAN [32]). UNIT combines the concepts of VAE and GAN to achieve unsupervised image-to-image translation across different styles and content by shared latent representations. CycleGAN and DualGAN achieve unsupervised image-to-image translation between two different domains through adversarial training and cycle consistency loss. We conducted comparative experiments between these three models and the our proposed model.

In the experiments, we selected GPR images of void defects with rebar clutter interference from the test dataset as input for the trained models. We compared the suppression effects of different models, with the outcomes displayed in Figure 5.

As shown in the figure, it is evident that our model performs best in suppressing rebar clutter as the rebar clutter in the original images is effectively suppressed while more features of void defects are preserved. UNIT and CycleGAN also effectively suppress rebar clutter signals, but they reconstruct void defects incompletely with less clarity and accuracy in details. CycleGAN even incorrectly reconstructs some non-void clutter. In contrast, DualGAN shows slightly poorer performance in clutter suppression as rebar clutter is not completely suppressed and more clutter remains at the overlap of void signals, leading to distortion of void defect signals. Overall, our model significantly outperforms other models by effectively suppressing rebar clutter and preserving clear and intact void defect signals. The void defect signals that were previously masked or interfered with by rebar clutter can now be more easily identified. This validates the superior performance of our model.

To quantitatively compare the effectiveness of various models in suppress rebar clutter, this study uses Root Mean Square Error (RMSE), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index Measure (SSIM) to evaluate the images before and after rebar suppression [33,34]. RMSE measures the root mean square error between the original and denoised image. The formula is shown as Equation (8):

$$\begin{array}{cc}\hfill \mathrm{RMSE}& =\sqrt{{\displaystyle \frac{1}{mn}}\sum _{i=1}^m\sum _{j=1}^n{\left[I(i,j)-K(i,j)\right]}^2}\hfill \end{array}$$

(8)

where $I(i,j)$ and $K(i,j)$ denote the pixel values of the original and denoised images, respectively, while m and n denote the width and height, respectively. PSNR measures the peak signal-to-noise ratio between the original and denoised image. The formula is shown in Equation (9):

$$\begin{array}{cc}\hfill \mathrm{PSNR}& =10{log}_{10}\left({\displaystyle \frac{{\mathit{MAX}}^2}{\mathit{MSE}}}\right)\hfill \end{array}$$

(9)

where $\mathit{MAX}$ is the maximum possible pixel value, $\mathit{MSE}$ is the mean squared error. SSIM measures the similarity between the original and denoised images. The formula is shown in Equation (10):

$$\begin{array}{cc}\hfill \mathrm{SSIM}(x,y)& ={\displaystyle \frac{(2{\mu }_x{\mu }_y+C_1)(2{\sigma }_{xy}+C_2)}{({\mu }_x^2+{\mu }_y^2+C_1)({\sigma }_x^2+{\sigma }_y^2+C_2)}}\hfill \end{array}$$

(10)

where ${\mu }_x$ and ${\mu }_y$ denote the mean pixel values of the images, while ${\sigma }_x^2$ and ${\sigma }_y^2$ denote the variances of the pixel values. ${\sigma }_{xy}$ is the covariance between the pixel values of the two images. Constants $C_1$ and $C_2$ are introduced to prevent a very small denominator.

By inputting the test set of GPR images with rebar clutter interference into the trained models, the average evaluation metrics were calculated for the results, as presented in

Table 2. Comparison of models evaluation metrics.

Models	RMSE	PSNR (dB)	SSIM
Ours	$5.0145$	$27.3702$	$0.8761$
UNIT	$5.2327$	$25.7002$	$0.8496$
CycleGAN	$5.1943$	$25.7281$	$0.8374$
DualGAN	$5.4676$	$23.0580$	$0.8259$

. As seen in the table, the three evaluation metrics of our proposed model perform the best. These results indicate that our model can effectively remove rebar clutter and retain the details and structural information of the void defects in GPR images with rebar clutter interference.

3.3. Ablation Experiments

We designed the CSA module for our model to validate the effectiveness of this module in reconstructing void defects. We modified our proposed model to create two variant models: one retaining only channel attention (removing spatial attention) and another retaining only spatial attention (removing channel attention). We conducted comparative experiments between these two variant models and our proposed model with the CSA module.

In the experiments, we selected GPR images of void defects with rebar clutter interference from the test dataset as input for the trained models. We compared the suppression effects of different models, with the outcomes displayed in Figure 6.

As shown in the figure, the model with the CSA module, which simultaneously applies channel and spatial attention, more comprehensively enhances useful features and suppresses noise. The suppression results show more layers and details of the void, with more prominent and continuous waveforms. When only the channel attention module is added, the lack of spatial attention may lead to insufficient capture and suppression of clutter features in the spatial distribution. This results in less accurate reconstruction of void details, with some non-void clutter being slightly mis-reconstructed. When only the spatial attention module is added, the lack of feature selection and weighting across channels may result in an inability to fully utilize information from different channels to suppress clutter, making the void signals appear slightly blurred. Overall, the model with the CSA module demonstrates the best performance. The other variant models show some deficiencies in reconstructing void defects, especially in correctly reconstructing the weak void waveforms under rebar clutter. This validates the superior performance of the CSA module in removing rebar clutter while maintaining void defect integrity.

The suppression effect of the model was evaluated using the same RMSE, PSNR, and SSIM metrics. The results are shown in

Table 3. Comparison of module evaluation metrics.

Models	RMSE	PSNR (dB)	SSIM
Ours	$5.0145$	$27.3702$	$0.8761$
The model with channel attention	$5.0776$	$26.5532$	$0.8647$
The model with spatial attention	$5.1816$	$25.7657$	$0.8528$

. It is evident that the model incorporating both channel attention and spatial attention modules outperforms the other two variant models, further validating the necessity of these attention mechanisms for improving the model’s performance in rebar clutter suppression. Additionally, all models outperform the scenario without any attention modules. This demonstrates the superiority of the CSA module in improving performance.

3.4. Feature Visualization Analysis

To visually demonstrate the feature extraction effectiveness of the CSA module, we compared the output feature maps of the last CSA module in the generation path with those from a model without the CSA module at the corresponding node. For the model without the CSA module, we used skip connections as a substitute. The final visualization results are shown in Figure 7.

In the figure, the model with the CSA module exhibits clearer feature details and more effectively focuses on the critical void defect feature locations, making the defects more prominent. Simultaneously, this model successfully ignores unrelated rebar clutter interference. In contrast, the model without the CSA module still shows noticeable rebar clutter at the top of the image, and the void defect features are relatively less distinct. This indicates that the inclusion of the CSA module enables the network to concentrate more effectively on relevant features, assigning higher weights to these features. Consequently, the network can reconstruct cavity defects more accurately.

3.5. Suppression Under Different Arrangement Conditions

To further evaluate the adaptability of the proposed model under different rebar arrangement conditions, additional simulated datasets with varying rebar spacings and defect depths were constructed based on the core dataset (rebar spacing of 0.2 m). Keeping other parameters unchanged, the same model training strategy was employed to conduct suppression experiments.

Figure 8 presents the suppression results under different rebar spacings (0.15 m, 0.2 m, 0.25 m and 0.3 m). It can be observed that rebar spacing significantly affects the distribution of multiple reflection signals. When the rebar spacing is 0.15 m, the rebar clutter is highly dense, leading to overlapping multiple reflection signals that obscure the defect region, making it difficult to identify clearly. As the spacing increases to 0.2 m and 0.25 m, the multiple reflection interference is reduced to some extent, and the visibility of defect contours improves. Notably, at a spacing of 0.25 m, compared to 0.15 m, the signal structure becomes more distinct, and the defect region appears clearer. When the spacing is further increased to 0.3 m, the intervals between rebar clutters expand, significantly reducing multiple reflection signals. Consequently, the defect boundaries become the most distinct, and the signal structure appears the most well defined. After rebar suppression processing, the proposed model demonstrates strong performance across different rebar spacing conditions, effectively mitigating multiple reflection interference while preserving defect information integrity. Particularly when the rebar spacing is small (0.15 m), although the background interference remains strong, the suppression results show a significant reduction in clutter, allowing defect contours to gradually emerge.

Figure 9 presents the suppression results for different defect depths. As seen in the figure, the proposed model continues to perform effectively in suppressing multiple reflection interference as defect depth increases. For shallow defects (rebar-defect distance within 0.3 m), the defect region exhibits strong reflection signals but is easily affected by background clutter, leading to blurred defect signals in the original images. After suppression processing, despite the shallowness of the defects, the model effectively attenuates multiple reflection interference, preserves the defect region’s distinct contours, and enhances signal clarity. For deeper defects (rebar-defect distance beyond 0.6 m), the defect reflection signals gradually weaken, while the influence of background clutter becomes relatively minor. Under these conditions, the model remains capable of accurately suppressing reinforcement clutter and multiple reflection signals, resulting in clearer defect boundaries and more defined signal structures.

Overall, the experimental results demonstrate that the proposed model consistently delivers stable and high-quality interference suppression under varying rebar spacing and defect depth conditions. This significantly enhances defect detectability, indicating the model’s strong generalization ability and robustness in complex scenarios.

4. Validation on Real GPR Data

In this section, we use real GPR images collected from the Husa Tunnel project to verify the effectiveness of our model, as shown in Figure 10. The Husa Tunnel is located in the western region of Yunnan Province, China, at an altitude between 1404∼2020 m. It is the longest tunnel on the Tengchong to Longchuan highway.

We were authorized to use the SIR3000 geological radar system, manufactured by Geophysical Survey Systems, Inc. (GSSI, Nashua, NH, USA), equipped with a 400 MHz central frequency antenna, to inspect the concrete lining structure of the tunnel segment YK81+310∼YK82+840. The system operates within a frequency range of 16–2200 MHz, with a sampling rate of 0–8000 ns and a depth penetration capability of up to 50 m in concrete, as specified by the manufacturer. In this study, each A-scan consisted of 512 samples, ensuring sufficient resolution for analyzing subsurface reflections. Similar applications of this system have been reported in previous studies [35,36], providing a basis for interpreting our results. When the vertical depth of the defect is less than 10 cm, the defect is classified as non-dense. When the vertical depth of the defect is greater than 10 cm, the defect is classified as a void [37]. Ultimately, five detection lines with void characteristics were identified. We randomly selected data from three of these lines for sliding-window-based data extraction while usingg data augmentation techniques such as geometric or color transformations to enhance the real GPR image dataset. This resulted in 100 real GPR images with rebar clutter, and we also selected 100 real GPR images with only void defects. The data from the remaining two lines were used for the final testing of the model.

Due to factors such as the site environment, real data are often affected by background noise and exhibit significant differences from simulated data. To enhance the model’s ability to adapt to different data distributions and environmental conditions in real-world scenarios, this study adopted a transfer learning and fine-tuning strategy. First, the model was pre-trained on simulated data at 800 MHz, which has higher resolution and clearer defect features. This allowed the model to better learn how to remove rebar clutter and retain defect information during the early stages of training. Then, the model was fine-tuned using real data collected at 400 MHz, enabling it to adapt to the GPR signal characteristics in actual engineering environments. This helped bridge the distribution gap between simulated and real data, further improving the model’s applicability in real tunnel tasks.

Figure 11a shows an GPR image of a void defect at the arch top of the right section entrance of the Husa Tunnel, within the range of YK81+546∼YK81+549. Before inputting the GPR image into the trained model for rebar clutter removal, data preprocessing was first performed to suppress noise and highlight target echoes. This included time-zero correction, DC offset removal, and one-dimensional frequency filtering. Figure 11b shows the processed GPR image. The effect after rebar suppression by our model is shown in Figure 11c. It is evident from the figures that, after rebar suppression, the rebar reflection signals in the image were significantly weakened, making the outline of the void defect more distinct and easier to identify. The results showed that our rebar suppression method effectively improves the visibility of void defects signals, thereby enhancing the interpretability of GPR data for tunnel lining inspections.

5. Conclusions

This study proposes an unsupervised generative network model based on VAEs and GANs to suppress multiple reflection interference signals caused by rebar in GPR data. The model transforms GPR images with rebar interference into clutter-free representations, effectively reducing rebar clutter and enhancing void defect signals. By incorporating a channel and spatial attention (CSA) module, the model significantly improves the accuracy and integrity of void defect reconstruction while suppressing rebar-induced noise. Comparative experiments with baseline models demonstrate that our approach achieves superior performance in reducing clutter and preserving void defect signals, as evaluated by RMSE, PSNR, and SSIM. Ablation studies further confirm the effectiveness of the CSA module, with its inclusion leading to a reduction in RMSE from 5.2327 to 5.0145, an increase in PSNR from 25.7002 to 27.3702, and an improvement in SSIM from 0.8496 to 0.8761, highlighting its role in enhancing reconstruction quality. Additionally, experiments under different rebar arrangement conditions validate the robustness of our method. Finally, validation using real GPR data from the Husa tunnel demonstrates that our approach effectively enhances void defect visibility and improves GPR data interpretability in practical applications.

Although this study employs real GPR data from the Husa tunnel, the proposed method is designed to address the widely existing issue of rebar interference in GPR imaging. Since rebar clutter is a common challenge in the GPR inspection of reinforced concrete structures, our approach is not limited to a specific tunnel but has broader applicability. Future research will focus on improving the model’s adaptability to different GPR devices, antenna configurations, and structural conditions, ensuring its robustness across diverse engineering scenarios. Expanding the training dataset with varied GPR system frequencies, structural compositions, and environmental factors will further enhance its generalization ability. Additionally, integrating domain adaptation techniques and adaptive learning mechanisms will improve transferability to new datasets and enable real-time applications. These advancements will strengthen the model’s effectiveness, making it applicable to a wider range of infrastructure inspections, including tunnels, bridges, and other reinforced concrete structures.