Intelligent Recognition of Road Internal Void Using Ground-Penetrating Radar

Abstract

Internal road voids can lead to decreased load-bearing capacity, which may result in sudden road collapse, posing threats to traffic safety. Three-dimensional ground-penetrating radar (3D GPR) detects internal road structures by transmitting high-frequency electromagnetic waves into the ground and receiving reflected waves. However, due to noise interference during detection, accurately identifying void areas based on GPR-collected images remains a significant challenge. Therefore, in order to more accurately detect and identify the void areas inside the road, this study proposes an intelligent recognition method for internal road voids based on 3D GPR. First, extensive data on internal road voids was collected using 3D GPR, and the GPR echo characteristics of void areas were analyzed. To address the issue of poor image quality in GPR images, a GPR image enhancement model integrating multi-frequency information was proposed by combining the Unet model, Multi-Head Cross Attention mechanism, and diffusion model. Finally, the intelligent recognition model and enhanced GPR images were used to achieve intelligent and accurate recognition of internal road voids, followed by engineering validation. The research results demonstrate that the proposed road internal void image enhancement model achieves significant improvements in both visual effects and quantitative evaluation metrics, while providing more effective void features for intelligent recognition models. This study offers technical support for precise decision making in road maintenance and ensuring safe road operations.

1. Introduction

The formation of internal road voids is primarily influenced by natural geological conditions, engineering factors, and environmental effects, specifically including groundwater loss, construction quality, drainage systems, vehicle loads, underground pipeline leakage, and environmental changes. These factors collectively lead to the loss of subgrade materials and structural damage, significantly affecting road structure stability, reducing load-bearing capacity, and potentially causing road collapse under the stress of heavy vehicles. Additionally, if internal road voids are not repaired in a timely manner, they will continue to develop and expand under the influence of vehicle loads, increasing the cost of road maintenance and affecting traffic efficiency during the repair process [1,2,3]. To ensure the health of road operations and traffic safety, scholars have conducted research on detection methods for concealed void-related diseases within roads. By accurately understanding the distribution and development of voids, repair strategies can be formulated to minimize their impact on traffic. At the same time, numerous intelligent recognition algorithms have been proposed, enhancing the efficiency of internal road void detection [4,5]. However, due to factors such as noise interference and the complexity of the medium materials, it is challenging to maintain high-quality internal road image data, affecting the accuracy of void area discrimination [6,7,8]. Therefore, researching image enhancement methods for low-quality internal road void image data can improve the accuracy of void identification.

Compared to methods for detecting apparent road diseases, the internal detection cannot be directly assessed through visual observation. The detection includes invasive methods such as slot cutting [9,10,11], as well as non-invasive techniques such as acoustic testing and electromagnetic wave detection [12,13,14]. Compared to destructive testing, non-destructive testing does not compromise the integrity of the road structure, allowing for a comprehensive assessment of the internal structure of the road and ensuring the continuity of operations. Among non-destructive methods, GPR is widely used. This method analyzes the internal state of roads by interpreting electromagnetic wave signals. Compared to other methods, GPR features strong penetration, high resolution, and fast data collection. Moreover, the detection process does not require contact with the surface of the object, enabling the rapid, comprehensive, and safe detection of internal road voids. Therefore, scholars often use GPR to conduct research on road defects.

GPR signal processing significantly impacts data accuracy and interpretation speed. Researchers have achieved road internal structure assessment through Neyman–Pearson spatial correlation analysis [15,16] and clarified radar signal attenuation characteristics and penetration depth [17]. To improve signal quality, methods including two-dimensional digital filter clutter removal, fast independent component analysis with multifractal spectrum denoising [18,19], and parameter-optimized generalized S-transform [20] were proposed. These studies enhanced adaptability in complex environments by reducing noise and strengthening signal features. Research has shown that void areas appear as bright features in radar images, with high-frequency components closely related to details and edge variations [21,22,23]. Analysis of high-frequency image information significantly improves void area interpretation accuracy, providing reliable evidence for internal road anomaly detection.

Researchers have conducted intelligent void detection studies using YOLO series models [24,25,26,27], knowledge distillation techniques [28], and 3D convolutional neural networks [29], but have focused mainly on algorithm improvements while neglecting image quality enhancement. Although image quality enhancement research has been conducted in other fields, such as remote sensing image enhancement based on relation cross-attention modules [30], hybrid CNN for fruit classification [31], exposure difference network-based enhancement [32], DEANet for low-light image enhancement [33], underwater image enhancement [34], and wavelet-based MRI enhancement [35], these techniques are difficult to directly apply to GPR void image quality enhancement due to the unique characteristics of GPR images, necessitating targeted research.

In recent years, research on GPR image enhancement has gradually developed. Researchers have proposed image enhancement methods based on non-linear technology, using LIP model and CDF-HSD functions to improve target signal contrast while suppressing background noise [36]; data enhancement methods based on AutoAugment have improved the performance of intelligent recognition models [37]; deep learning methods based on diffusion models have achieved simultaneous optimization of image resolution enhancement and clutter removal, improving the quality of deep GPR images [38]. However, although existing enhancement methods have achieved good results in single tasks, they lack in-depth exploration and effective modeling of the feature correspondence relationship between high-quality and low-quality images. This limitation makes it difficult to accurately capture and maintain key feature information during the enhancement process, affecting the reliability of enhancement results.

Therefore, this study aims to conduct research on the intelligent recognition of road internal voids using 3D GPR. Initially, internal road void data were collected using 3D GPR. For the collected images, an image quality enhancement model was proposed specifically for void images, which integrated the advantages of the Unet model, Multi-Head Self-Attention (MHSA) mechanism, and Multi-Head Cross-Attention (MHCA) mechanism. Finally, intelligent recognition models of road voids were constructed using images before and after enhancement to clearly determine the impact of image enhancement on model performance. This research aims to provide an accurate method for recognizing internal road voids, rapidly assessing internal void areas, and providing technical support for comprehensive road maintenance.

Therefore, this study aims to conduct research on the intelligent recognition of road internal voids using 3D GPR. By proposing a quality enhancement model specifically designed for GPR images of internal road voids, this study aims to reduce the impact of noise interference and complex media materials during GPR image acquisition on image quality, thereby improving the accuracy of intelligent recognition models for internal road voids. The research assumes that the proposed image enhancement model can effectively improve the feature representation of void areas in GPR images and enhance image quality. Additionally, the enhanced images are expected to provide clearer void features for the intelligent recognition model, thus increasing the detection accuracy. Initially, internal road void data were collected using 3D GPR. For the collected images, an image quality enhancement model was proposed specifically for void images, which integrated the advantages of the Unet model, Multi-Head Self-Attention (MHSA) mechanism, and Multi-Head Cross-Attention (MHCA) mechanism. Finally, intelligent recognition models of road voids were constructed using images before and after enhancement to clearly determine the impact of image enhancement on model performance. This research aims to provide an accurate method for recognizing internal road voids, rapidly assessing internal void areas, and providing technical support for comprehensive road maintenance.

2. Materials and Methods

2.1. Data Acquisition

This study employed 3D GPR technology to detect voids within road structures. The 3D GPR system features a radar control module, along with an antenna array, Doppler range-finding technology, and an RTK system. The center frequency of antenna used by the radar is 200 MHz. The system is equipped with 15 antenna sets, which improves the efficiency of data acquisition. The effective width of radar scanning is 180 cm. In order to achieve coverage of the entire lane, the single lane is detected using two GPR scans.

In this study, 3D GPR was utilized to survey major transportation routes in Shenzhen. The survey covered hundreds of roads across multiple administrative districts in Shenzhen, including Bao’an, Longhua, Guangming, Luohu, Futian, and Yantian Districts. The detection mileage exceeded 1000 km, and more than 1700 void images were obtained. LabelImg v1.8.1 software was used to label the empty areas in the image, and the image was segmented into training, validation, and test sets for model training.

The on-site inspection image is shown in Figure 1.

2.2. Echo Characteristics of Internal Voids in Roads

The echo characteristics of voids in roads affect the accuracy of identifying void areas. Therefore, this section will conduct research on the echo characteristics of internal voids in roads. An example of an internal road void appears in Figure 2.

When electromagnetic waves propagate through the interior of a road and encounter a void area, the significant difference in dielectric properties between air and road materials results in a pronounced enhancement of the reflected electromagnetic signal. This is displayed in the image as a prominent bright area. Additionally, when electromagnetic waves encounter these voids, their propagation speed changes, which affects the continuity of the co-axial lines in the image. As the GPR is mobile during detection, the path of reflection is at its minimum when positioned directly over a void. As the radar distances itself from the void area, the path of reflection progressively extends, appearing in the radar image as a hyperbolic shape, creating unique wing-shaped patterns flanking the hyperbolic shape within the image.

The GPR image characteristics vary with different void formation causes and void dimensions. Three void images with different causes and dimensions were selected for comparison, as shown in Figure 3. For Type 1, the void has dimensions of length 1.4 m, width 1.1 m, burial depth 0.27 m, and net height 0.31 m. For Type 2, the void dimensions are length 1.3 m, width 1.2 m, burial depth 0.51 m, and net height 0.28 m. For Type 3, the void dimensions are length 2.8 m, width 2.3 m, burial depth 0.65 m, and net height 0.51 m. Type 1 and 2 voids were caused by pipeline damage, while Type 3 voids were caused by nearby construction activities.

Comparison of these three void images reveals that larger-sized voids produce wider reflection ranges with more pronounced extensibility in the images. Greater net height generates stronger reflection signals, while increased burial depth weakens signal strength, resulting in relatively blurred reflection features. Additionally, voids with different causes exhibit distinct characteristics. Pipeline damage-induced voids show localized strong reflections with clear hyperbolic features, well-defined boundaries, and relatively concentrated spatial distribution. In contrast, construction-induced voids display broader reflection ranges, more expanded hyperbolic features, more complex internal structures, and more dispersed boundary reflections.

During the data annotation process, based on clearly defined void image characteristics, we implemented a three-level expert annotation system. This included initial annotation by technicians, secondary review by technical staff, and final verification by inspection personnel. For cases with disputes, team discussions were mandatory to reach consensus, ensuring consistency and standardization in the annotation process. Additionally, we established a problem case database to record and analyze challenging annotation cases, providing references and accumulated experience for improving annotation quality.

2.3. Image Quality Evaluation Metrics

In this study, four metrics—PSNR, SSIM, LPIPS, and NIQE—were employed to evaluate the quality of enhanced images. These metrics are commonly used in various fields such as medical imaging, security surveillance, and remote sensing for assessing image compression and denoising effects, as well as evaluating image processing quality [39,40].

(1) 

PSNR

PSNR is a measure used to assess the quality of reconstructed images. A higher PSNR indicates a better quality of image reconstruction, suggesting closer approximation to the original image.

PSNR is calculated as shown in Equation (1).

$$PSNR=10\times {\log }_{10}\left({\displaystyle \frac{MA{X}_I^2}{MSE}}\right)$$

(1)

where ${MAX}_I$ denotes the highest pixel value. $MSE$ is the mean square error.

(2) 

SSIM

SSIM is utilized to assess the visual resemblance between two images and is suitable for evaluating the effects of image quality enhancement. SSIM takes into account the structural information of images, reflecting the perceptual differences in image quality as seen by the human eye. The calculation method for SSIM is shown in Equation (2).

$$SSIM\left(x,y\right)={\displaystyle \frac{\left(2{\mu }_x{\mu }_y+c_1\right)\left(2{\sigma }_{xy}+c_2\right)}{\left({\mu }_x^2+{\mu }_y^2+c_1\right)\left({\sigma }_x^2+{\sigma }_y^2+c_2\right)}}$$

(2)

where $x$ and $y$ are the window regions of the two images; ${\mu }_x$ and ${\mu }_y$ are the mean values; ${\sigma }_x^2$ and ${\sigma }_y^2$ are the variances; ${\sigma }_{xy}$ is the covariance; $c_1$ and $c_2$ are constants used to maintain stability.

(3) 

LPIPS

LPIPS is a metric for assessing image similarity that considers the characteristics of human visual perception, using deep learning models to evaluate visual differences between two images. The lower the LPIPS value, the smaller the perceptual differences between the two images, indicating that the corresponding algorithm maintains higher visual quality and similarity.

Unlike metrics based on simple pixel comparison, LPIPS utilizes the advanced features of images in calculating differences between them, providing better performance and robustness when dealing with complex features. The calculation method is detailed in Equation (3).

$$LPIPS\left(x,y\right)={\displaystyle \sum _{k=1}^N\left(\frac{1}{H_n{W}_n}{\displaystyle \sum _{i=1}^{H_n}{\displaystyle \sum _{j=1}^{W_n}{‖{\omega }_n\odot \left(F^n\left(x\right)-F^n\left(y\right)\right)‖}_2}}\right)}$$

(3)

where $LPIPS\left(x,y\right)$ represents the perceptual similarity measurement between images x and y; $N$ represents the count of layers; $F$ is the pretrained network; $H_n$ and $W_n$ are the height and width of the feature map; $F^n\left(x\right)$ and $F^n\left(y\right)$ are the feature representations; ${\omega }_n$ is the weight parameter at layer $n$; $\odot$ denotes element-wise multiplication; ${‖\cdot ‖}_2$ is the L2 norm.

(4) 

NIQE

NIQE is a no-reference image quality evaluation technique that operates without the need for an original or an ideal reference image. Based on a natural scene statistics model, NIQE learns statistical characteristics from high-quality natural images and constructs a probabilistic model describing these statistical features. NIQE assesses image quality by comparing its features with those of the learned model, determining deviations from this norm. A lower NIQE value indicates that the image quality is closer to that of high-quality natural images. The computational method for NIQE is shown in Equation (4).

$$NIQE=\sqrt{{\left(x-\mu \right)}^T{\Sigma }^{-1}\left(x-\mu \right)}$$

(4)

In the equation, $x$ represents the feature vector; $\mu$ and $\sum$ are the mean and covariance of the pretrained model. ${\left(x-\mu \right)}^T{\Sigma }^{-1}\left(x-\mu \right)$ represents the Mahalanobis distance, which quantifies the degree of deviation between the image features and the model features.

3. Results

3.1. Research on Preprocessing Techniques for Internal Void Images in Road

3.1.1. Image Size Optimization

Due to the diverse causes and random distribution of void areas within road interiors, influenced by the coverage of 3D GPR and the propagation characteristics of electromagnetic waves, some void regions are located near the edges of the radar images. In such images, it is particularly challenging to fully capture neighborhood pixel information at the edges.

In response to the aforementioned challenges, this section proposes an approach to enhance the image dimensions as a means to mitigate the low recognition accuracy stemming from void areas located near the image edges.

Consequently, padding is added symmetrically along the width of the image, based on its length, thereby increasing the overall image dimensions. To maintain proportional balance and prevent distortion of the targets due to uneven expansion, the expanded image is formatted into a square shape. This standardization facilitates more uniform processing within the algorithmic models and minimizes variability caused by image size differences. The effectiveness of this padding technique is demonstrated in Figure 4.

3.1.2. Image Quality Degradation and Enhancement

In practical scenarios, images of voids inside roads captured using GPR are affected by factors such as signal attenuation, noise interference, equipment resolution, and the complexity of the underground medium. These factors lead to instability in the quality and clarity of the images. Therefore, research into image quality enhancement techniques is necessary.

Typically, supervised learning processes require paired datasets, which comprise input images of low quality and target images of high quality. However, in the field, high-quality images corresponding to the low-quality images of voids inside roads are lacking, which hampers the learning process and renders supervised learning methods inapplicable. Unsupervised learning, on the other hand, mainly relies on analyzing the intrinsic structure and patterns within the data itself, independent of paired label data. However, it suffers from a lack of intuitiveness and a clear optimization goal, leading to unstable improvements in image quality and an inability to guarantee achieving the desired quality level.

Based on these challenges, this study proposes a method for constructing void image data aimed at supervised learning. This method primarily generates low-quality and high-quality images by degrading and enhancing the image quality, respectively, based on images collected in the field.

When identifying void areas within road interior radar images, the information at the image edges is particularly important as it can be used to determine the precise location, shape, size, and distribution of the void areas. In the image processing phase, the high-frequency information in an image can reflect details such as image feature characteristics, target area edges, and spatial changes. Therefore, during the process of image quality degradation, emphasis will be placed on processing the high-frequency information characteristics of the image. Fourier transform low-pass filtering is a technique used in the frequency domain to process images. It can be used to diminish and remove high-frequency elements in an image while retaining the low-frequency elements, thus achieving the goal of image quality degradation.

When collecting images of road interiors using 3D GPR, various types of noise are encountered. Electromagnetic interference noise mainly originates from electromagnetic waves in the surrounding environment, which can disrupt the reception of radar signals. Random noise is typically generated by the electronic components of the radar equipment itself, influenced by the quality of equipment manufacturing and external environmental factors. System noise includes noises produced by reflections and scatterings within the radar system, originating from radar wave reflections at the interfaces between different mediums. Echo blurring occurs due to the multi-directional scattering of radar waves during transmission, leading to the overlapping and blurring of echo signals, making it difficult to discern image details. The presence of these noises reduces the quality of the images. Gaussian filtering is effective at decreasing image noise through image smoothing while preserving edge information without damaging the edges.

The effects of image quality degradation and image quality enhancement are illustrated in Figure 5.

From Figure 5a, it can be observed that the void area displays alternating black and white highlighted regions. This is due to the dielectric constant at the void being lower than the surrounding materials, causing strong reflections of electromagnetic waves in this area. The image colors in non-void areas are relatively uniform, with some areas appearing wavy, which is caused by the inhomogeneity of the road medium and the differences in the propagation of electromagnetic waves through it. As shown in Figure 5b, after applying Fourier low-pass filtering to the image, the clear edges and details in the original image become relatively blurred, and the image contrast is reduced, resulting in a softer and smoother image. From Figure 5c, it is evident that after processing the image with Gaussian filtering, the edge information in the void area is well preserved, and the pixel value changes in non-void areas are relatively gentle. Gaussian filtering effectively smooths such areas and reduces noise.

3.2. Research on a Road Internal Void Image Enhancement Method Based on Improved Unet Model

In this section, while conducting research on the network model for road internal void imagery, the Unet neural network architecture was referenced [41]. Based on MHSA modules and MHCA modules, an image quality enhancement model aimed at internal road void imagery was designed [42,43].

3.2.1. Unet Neural Network Model

The Unet model, characterized by its symmetric encoder and decoder architecture, along with effective skip connections, is capable of deeply extracting and utilizing multi-level features, significantly enhancing its ability to process complex images. This makes the Unet model particularly well-suited for processing GPR images of road internals. Through end-to-end learning, the Unet can adaptively recover clear and precise structures from damaged images, effectively enhancing image quality and detail representation.

The Unet network employs an encoder–decoder structure, which includes a contracting path and a symmetric expansive path, connected through a bottleneck layer. The network’s architectural diagram is depicted in Figure 6.

The primary function of the encoder is to capture the image’s feature information and decrease its spatial dimensions. Simultaneously, it boosts the number of feature channels to augment the network’s ability to represent features. Within the encoder, the input image undergoes four stages of downsampling, each containing two 3 × 3 convolution layers, which are then succeeded by a 2 × 2 max pooling layer, effectively extracting multi-scale features and capturing a broader range and more complex features.

The decoder primarily focuses on gradually upsampling the encoded feature maps back to the original image size and conducts feature learning through convolution layers, thereby enhancing feature extraction capabilities. The decoder architecture includes four upsampling stages. Each upsampling module contains two convolution layers and one transpose convolution layer, effectively increasing the spatial dimensions of the feature maps and restoring their spatial resolution. After the upsampling module, a convolution layer is used to achieve two-channel feature maps, preserving the spatial dimensions of the feature layers while altering the channel count, mapping high-dimensional features to the required output dimensions for classifying and locating features and background in the target areas. Skip connections are a crucial component of the Unet structure. These connections directly tie feature maps from the encoder to matching feature maps in the decoder, helping to retain more contextual and positional information in the output and enhancing the restoration of image details through feature fusion.

3.2.2. MHSA Module and MHCA Module

• (1) MHSA Module

The MHSA mechanism, prevalent in deep learning models, allows the model to concurrently absorb information from various representational subspaces, thus boosting its ability to handle complex data. The mechanism primarily operates by dividing the “attention” operation into multiple heads, each independently learning different aspects of the input data, which are then combined to achieve a more comprehensive understanding.

The Query (Q), Key (K), and Value (V) vectors are the core components of the attention mechanism. Here, the query vector is used to locate areas of interest, the key vector identifies all potential points of focus, and the value vector stores the corresponding feature content.

In the MHSA mechanism, each head performs calculations independently, allowing for the parallel processing of various features and parts of the image. Ultimately, the outputs from all heads are merged together, forming a comprehensive feature representation that aids in improving the accuracy of image feature extraction.

• (2) Self-Attention Mechanism (SAM)

When utilizing the SAM, the procedure starts by converting the model input into Query, Key, and Value matrices. Subsequently, attention scores are calculated by conducting a dot product between Query and each Key. The specific computational process is illustrated in Figure 7.

Where X symbolizes the feature representation. The width, height, and channel count of the feature map are denoted by w, h, d, respectively.

$W^Q$, $W^K$, and $W^V$ are the learnable linear transformation matrices corresponding to the $Q$, $K$, and $V$.

• (3) MHSA mechanism

The layout of the MHSA mechanism is illustrated in Figure 8.

For each head h, different weight matrices are used to transform the input, and the attention scores for each head are calculated. The outputs of all heads are concatenated together and passed through a final linear transformation.

• (4) MHCA Module

The MHCA is an extension of the SAM that enables the model to effectively utilize the information from one input while processing another, enhancing the model’s recognition and understanding capabilities. As indicated in the red box in Figure 9, the feature map S, after being processed by the module CBSU, is used as the value V, and the features of the feature map Y are used as the query Q and key K. This utilizes the information from S to improve the model’s capability for feature extraction from Y.

The process for calculating MHCA is conducted as illustrated in Figure 9.

In the schematic, feature map A undergoes processing through the CBSU module to further extract and refine the information contained within feature map A and to expand the feature dimensions. Feature map S, after being processed by the CBR module, undergoes a dot product operation with the result from feature map A after CBSU module processing. This operation further emphasizes the important features within feature map S in the output’s weighting. Meanwhile, feature map Y, after processing through the UC module, is concatenated with the output from the aforementioned dot product operation. This enhances the richness of the model output, providing a deep insight into the image content and enhancing the accuracy and robustness of target area recognition.

3.2.3. Image Enhancement Model Design Based on an Improved Unet Model—MHUnet

3D GPR captures road interior void images that often suffer from inconsistent quality. Moreover, void areas feature diverse scales and complex, variable backgrounds. Consequently, an image enhancement network model for void images needs robust feature extraction capabilities. This model must accurately identify and analyze both global and detailed features of void areas across different scales of receptive fields, while also balancing the extraction of local features with the integration of global information, adapting to complex and variable image background conditions.

Although the Unet network model attempts to expand the receptive field through its multi-layer structure, each layer’s receptive field is limited, making it challenging to cover global information throughout the image. The structure’s skip connections help maintain local feature information, but they struggle with processing complex void images over large areas. Furthermore, the Unet model inadequately addresses image quality variations caused by noise and signal attenuation in diverse underground media, leading to inconsistent image reconstruction quality.

Therefore, this section introduces MHSA and MHCA mechanisms into the Unet network model structure. MHSA, by interacting with other units, can acquire rich global information. Simultaneously, by learning long-range dependencies across the entire input image, it effectively expands the model’s receptive field, enhancing the handling of large-scale, multiple voids, and complex images. MHCA effectively merges information between different network layers and modules, focusing on different feature combinations to improve the feature extraction performance of void areas and improve image quality enhancement effects.

The structure of the designed void image enhancement model is illustrated in Figure 10.

The initial part of the model is the StemConv module, which comprises four convolutional layers (Conv), batch normalization layers (BN), and GeLU activation functions. This module primarily serves to extract preliminary features from the image and enhance non-linear processing capabilities. The DownConv module reduces the spatial dimensions of the feature map through maximum pooling and deepens feature extraction with two convolutional passes to capture abstract features. The MHSA module enhances the capture of global dependency features, enhancing the model’s capability to identify key features across the entire image. The MHCA module integrates features from different levels to enhance feature expressiveness. The UpConv module is used to gradually reconstruct the spatial resolution of the features, providing a foundation for high-quality image reconstruction. Concat helps recover more detailed information in void areas. TConv uses two different sizes of convolutional kernels to further process and optimize the feature map, maintaining contextual information while adjusting the number of channels, proving helpful in modifying the dimensions and depth of the feature map.

3.2.4. Analysis of Image Enhancement Effects Based on the MHUnet Model

The MHUnet model was utilized to enhance the quality of low-quality void maps within road interiors, with corresponding results shown in Figure 11.

As evident from the figure, compared to the low-quality image, the enhanced image exhibits a deepening of black areas and a brightness increase in white and gray areas, significantly improving the image contrast. The enhanced contrast makes the distinction between void areas and the surrounding background more pronounced. The increase in brightness aids in better recognition of details within the image. In the enhanced image, the edges of the internal structure areas of the road are sharper and clearer, which helps in further distinguishing between void and non-void areas. In summary, the enhanced image has a higher degree of detail recognition, which helps to improve the accuracy and reliability of the interpretation of the void map.

The MHUnet model was used to enhance the quality of collected GPR images, with the comparison between pre- and post-enhancement shown in Figure 12.

As can be seen from the figure, compared to the original collected images, the enhanced images show improvements in multiple aspects. The void area boundaries are sharper, detail features are more distinguishable, and overall image blur is reduced; the contrast between void areas and background is more pronounced with richer grayscale levels; background noise is effectively suppressed while main features are more prominent; additionally, the hyperbolic features of void areas are more distinct, structural boundaries are clearer, and target areas show better differentiation from surrounding environments. These improvements contribute to higher accuracy in subsequent intelligent recognition.

To avoid having two different feature distributions (original and enhanced) in the dataset and reduce the negative impacts of data distribution differences on model training, the image enhancement method was applied to all 1700 images in the dataset. Additionally, this approach significantly simplifies the processing workflow by eliminating the need for image quality judgment and classification.

To demonstrate the advancement of our proposed method, comparison with existing image enhancement models is necessary. Among existing image enhancement models, NAFNet and Uformer are widely applied. NAFNet has extensive applications in image processing, primarily in low-light image enhancement, underwater image enhancement, and remote sensing image denoising. This method shows excellent performance in motion blur removal, Gaussian noise removal, and image quality improvement, with advantages in computational efficiency and small model size [44,45,46]. The Uformer method is mainly applied in low-light image enhancement, particularly excelling in ultra-high-definition image processing. This method not only improves image quality but also serves as a preprocessing step for downstream visual tasks such as face detection [47,48,49].

The image enhancement effect was quantitatively analyzed using four indicators: PSNR, SSIM, LPIPS, and NIQE. Among them, the quality-degraded image and the MHUnet method enhanced image corresponding indicators were calculated using the image preprocessed image as the reference image, as shown in

Table 1. Model performance metrics for different data types.

Image Type	PSNR (dB)	SSIM	LPIPS	NIQE
Not enhanced	28.21	0.9137	0.0253	12.1723
Unet	30.12	0.9309	0.0211	10.8357
NAFNet	32.38	0.9488	0.0197	10.2675
Uformer	34.06	0.9364	0.0145	9.3738
Enhanced by MHUnet	34.65	0.9695	0.0165	9.6543

.

The table reveals that the quality-degraded images have lower PSNR and SSIM values, and higher LPIPS and NIQE scores, indicating a certain similarity in image structure between the low-quality images and reference images, but still a significant difference in visual quality and image structure. After applying the MHUnet method, the PSNR and SSIM values increased from 28.21 and 0.9137 to 34.65 and 0.9695, respectively. This increase indicates effective suppression of image noise and overall improvement in image quality, bringing the enhanced image nearer to the high-quality reference in aspects like brightness, contrast, and structural details. Additionally, the preservation of local features and textural elements in the image is well-maintained.

The LPIPS and NIQE scores decreased from 0.0253 and 12.1723 to 0.0165 and 9.6543, respectively. This reduction shows that the image enhancement method decreases the perceptual differences between the quality-degraded images and the reference images, which helps to improve the accuracy of identifying voids in road interiors where high resolution is crucial. Notably, LPIPS is a metric calculated using deep learning methods that can recognize the deep features of void images and is particularly sensitive to complex textures and shapes. Therefore, the enhanced image quality in this method aids in the accurate identification of void areas.

On the other hand, compared to the enhancement results of three models—Unet, NAFNet, and Uformer—the MHUnet model demonstrates superior performance across all four metrics. Although the NAFNet model performs excellently in conventional image processing, it is not entirely suitable for enhancing GPR images of road internal voids. This is mainly because GPR images contain specialized geological structural information, which differs significantly from the features of natural images. Similarly, the Uformer model is not fully applicable to enhancing GPR images of road internal voids. This is primarily because GPR images are formed based on electromagnetic wave reflection signals, which are fundamentally different from natural lighting scenes.

This further validates the effectiveness and advancement of the method proposed in this research. This is attributed to the proposed MHUnet model, which references the Unet model architecture capable of reconstructing images at various scales, enhancing image brightness, clarity, and contrast. The inclusion of MHSA and MHCA in the model adjusts and optimizes perceptual image features, improving visual quality and consistency in image perception. Through comprehensive feature learning and multi-level information integration, the model can produce results that are statistically closer to high-quality reference images, thereby enhancing the accuracy and reliability of intelligent void area recognition.

3.3. Comparative Analysis of Void Intelligent Recognition Performance

Using the YOLOv8 model, the performance between original images and enhanced images was compared. When constructing the dataset, the images were divided into a training set, validation set, and test set in a ratio of 6:2:2. The model performance was analyzed using precision, recall, F1, and mAP. Furthermore, to demonstrate the effectiveness of enhanced images in improving intelligent void recognition performance, statistical analysis was conducted using t-test. Specifically, the test set was divided into 5 groups, and the mean and standard deviation of corresponding metrics for each group were calculated, along with the t-value in the t-test. The model performance metrics and statistical indicators are shown in

Table 2. Comparison of YOLOv8 model performance metrics and statistical indicators.

Image Type	Precision			Recall
	AV (%)	SD (%)	t-Value	AV (%)	SD (%)	t-Value
Original image	86.55	0.553	10.43	80.31	0.512	16.86
MHUnet Enhancement	87.94	0.601		81.99	0.486
Image Type	F1(%)			mAP(%)
	AV(%)	SD(%)	t-Value	AV(%)	SD(%)	t-Value
Original image	83.31	0.526	17.24	86.74	0.412	13.79
MHUnet Enhancement	84.86	0.415		87.62	0.387

, where AV represents average value and SD represents standard deviation.

As shown in Table 2, the YOLOv8 model trained with MHUnet-enhanced images demonstrates overall superior performance compared to the model trained with original images, indicating that enhanced images contribute to improving the performance of road internal void recognition models. In the t-test analysis, the calculated t-values for precision, recall, F1, and mAP were 10.43, 16.86, 17.24, and 13.79, respectively. Based on the degrees of freedom and significance level, referring to the t-value table, the critical value t is 2.306 under the conditions of 8 degrees of freedom and a significance level of 0.05. Since the t-values for all four metrics exceed 2.306, this indicates that the void recognition model trained with enhanced images significantly outperforms the model trained with original images.

To further demonstrate the universality of the MHUnet model-enhanced images in improving recognition model performance, we also compared the training effects of different images on YOLOv7, YOLOv9, and Faster-rcnn model. The performance indicators corresponding to the three hollow intelligent recognition models are calculated and shown in

Table 3. Comparison of detection performance between original and enhanced images.

Model Type	Image Type	Precision (%)	Recall (%)	F1 (%)	mAP (%)
YOLOv7	Original image	85.61	78.43	81.68	85.85
	MHUnet Enhancement	86.99	79.75	82.63	86.73
YOLOv8	Original image	86.55	80.31	83.31	86.74
	MHUnet Enhancement	87.94	81.99	84.86	87.62
YOLOv9	Original image	86.28	80.54	83.38	85.99
	MHUnet Enhancement	87.38	81.84	84.95	87.35
Faster-rcnn	Original image	88.61	81.57	84.69	87.85
	MHUnet Enhancement	90.38	83.31	86.15	88.42

.

According to the table, compared with the original data, all four models trained on the enhanced data based on the MHUnet model have higher accuracy, recall, F1, and mAP than the models corresponding to the original data. The enhancement of image quality effectively enhances the learning and generalization abilities of intelligent recognition models for road internal voids, demonstrating higher accuracy and reliability.

3.4. Engineering Validation

The image enhancement technology corresponding to the MHUnet model and the intelligent recognition model were applied and validated in the detection of road interior defects on Menghai Avenue and Qianhai Avenue in the Nanshan District of Shenzhen. Specific site images and validation results are shown in Figure 13 and

Table 4. Verification of model detection accuracy.

Type	Number of Voids	Accuracy (%)
Model detection	10	90
Accurate verification	9

.

The intelligent recognition model identified a total of 10 void areas on Menghai Avenue and Qianhai Avenue, with 9 of these void areas being accurately validated. This demonstrates that the image enhancement proposed in this study can be effectively applied in the detection and recognition of road interior void areas.

4. Discussion

This study utilized 3D GPR to detect internal road voids across multiple administrative regions in Shenzhen, establishing a dataset of 1700 void images. In existing related research, scholars have employed two approaches to build void image datasets: numerical simulation and on-site GPR collection. While numerical simulation is cost-effective, simulated images show characteristic differences from real images [50]. On-site collection, though costly, results in smaller datasets typically not exceeding 1000 images [51,52,53,54,55]. Therefore, compared to existing research, our dataset provides a larger-scale data foundation for internal road void image enhancement and intelligent recognition.

A supervised learning-oriented void image processing method was proposed based on Fourier transform low-frequency filtering and Gaussian filtering. This method increases pixel differences between degraded and enhanced images by decreasing and increasing grayscale values in void regions, thereby expanding the model’s learning space. In road-related research, image enhancement technology primarily focuses on image quality improvement under complex conditions, dynamic scene correction, and target detection enhancement [56,57,58]. Image enhancement includes traditional and intelligent methods. Traditional methods comprise histogram processing, spatial domain processing, and morphological processing. Compared to traditional methods, intelligent enhancement methods offer advantages such as scene-adaptive parameter adjustment, stronger robustness and generalization ability, end-to-end processing without manual feature design and rules, and efficient processing after training through learning optimal enhancement strategies from large datasets. Among intelligent enhancement algorithms, the Unet model and its improved structures are widely applied, demonstrating advantages in medical imaging, remote sensing imaging, and natural scene imaging, including good image structure preservation, strong detail recovery capability, and fast processing speed [59,60,61]. However, research on internal road void image enhancement technology is still in its early stages. Our study combines the advantages of MHSA and MHCA in global modeling and cross-domain feature fusion [62,63], and based on the Unet model structure, introduces MHSA and MHCA to construct an improved Unet model (MHUnet) for enhancing internal road void images. Compared to original images, this model significantly improves enhanced image quality in both visual effects and quantitative indicators. The model improves the brightness, clarity, and contrast of internal road void images, enhancing visual quality and image perception consistency. The enhanced internal road void images can provide more effective features for intelligent recognition models, effectively enhancing their learning capability and improving model accuracy and sensitivity.

5. Conclusions

This study proposes an intelligent recognition method for road internal voids based on improved image enhancement technology. Three-dimensional GPR detection was conducted in Shenzhen to establish a void image dataset. The proposed MHUnet model incorporates MHSA and MHCA mechanisms, achieving significant improvements in image brightness, clarity, and contrast, enhancing perceptual consistency, and matching statistical distributions with high-quality reference images. For the enhanced images, the quantitative indicators improved significantly: PSNR increased by 6.44 dB, SSIM by 0.0558, LPIPS by 0.0088, and NIQE by 2.518. Experimental validation demonstrates that the enhanced images provide higher-quality feature information, improving the model’s accuracy and sensitivity while exhibiting superior recognition capabilities and generalization performance. Compared to the original images, the improved model showed increased performance metrics: precision by 1.39%, recall by 1.68%, F1 score by 1.55%, and mAP by 0.88%. The research findings provide a new technical solution for the intelligent detection of road internal voids, with significant practical implications for improving the intelligence level of road maintenance.

The next steps will focus on further optimizing the intelligent recognition model for internal road voids and conducting larger-scale engineering practice verification. Additionally, we will conduct quantitative assessment research based on void images to evaluate void size, depth, deterioration degree, void grade, and impact on road damage. By establishing a comprehensive void assessment system, we aim to better guide road maintenance work prioritization, improve maintenance efficiency, and provide important technical support for road lifecycle management, thereby enhancing the engineering application value of the research.