Pavement structure disease identification model and training method for interlayer poor adhesion

Abstract

The application provides a pavement structure disease identification model and a training method for interlayer bonding defects, and belongs to the technical field of image data processing. The method comprises the following steps: S101, defects of a pavement structure are detected and multi-step signal processing is performed on the pavement structure by using a ground penetrating radar to obtain a measured disease image of the pavement structure; S102, the measured disease image is expanded by using a BL-GAN model to generate a simulated disease image; S103, the measured disease image and the simulated disease image are input into a pavement disease identification model for training to obtain a trained pavement disease identification model; the BL-GAN model is a generative adversarial network composed of a generator and a discriminator; the generator adopts a structure of a 5-layer deconvolution network; the discriminator adopts a structure of a 4-layer convolution network; the generator and the discriminator are both embedded with a self-attention mechanism after a specific feature layer; and the pavement disease identification model is constructed by embedding a lightweight visual Transformer module in a neck network of YOLOv5s. The method can balance the efficiency and accuracy of the model.

Description

Technical Field

[0001] This application relates to the field of image data processing technology, and in particular to a pavement structure defect identification model and training method for poor interlayer bonding. Background Technology

[0002] Ground-penetrating radar (GPR) can provide key information about poor interlayer bonding, and its detection is one of the effective means of identifying pavement structural defects caused by poor interlayer bonding.

[0003] Detecting reflected waves from poorly bonded interlayer defects faces technical challenges such as signal complexity, noise interference, and weak features. Some existing studies employ convolutional neural networks (including deep neural networks) to extract features from defect images obtained by ground-penetrating radar (GPR), thereby achieving automatic defect identification. However, convolutional neural networks (including deep neural networks) are limited by their limited local receptive fields and hierarchical structure, hindering their ability to capture global features and resulting in insufficient feature extraction capabilities for defects in complex backgrounds, leading to low model identification accuracy.

[0004] In recent years, the Transformer model, with its powerful global feature modeling capabilities, has demonstrated significant advantages in identifying pavement defects in complex backgrounds and small-target defects. Applying Transformer to defect detection tasks provides a new approach to improving detection accuracy and adaptability. However, compared to traditional convolutional neural networks or deep neural networks, Transformer networks have significantly increased complexity and computational cost. Their training requires large-scale labeled data, and model convergence is challenging. Existing research primarily focuses on using Transformer to optimize complex defect identification, failing to effectively address the need for large-scale labeled data or consider deployment in resource-constrained environments. Specifically: firstly, ground-penetrating radar (GPR) data on pavement defects is scarce and labeling is costly; existing Transformer models are prone to overfitting when data is insufficient, affecting their practical engineering applications. Secondly, Transformer models typically have a large number of parameters and high computational complexity, making them unsuitable for deployment in resource-constrained environments and limiting their widespread adoption in practical engineering. These issues significantly restrict the efficient application of Transformer technology in pavement defect detection and the full realization of its engineering value.

[0005] Therefore, how to effectively utilize Transformer technology to accurately identify pavement defects in complex backgrounds, while taking into account both computational efficiency and practicality requirements, so as to fully realize the engineering value of defect identification models, is one of the problems that the industry urgently needs to solve. Summary of the Invention

[0006] The purpose of this disclosure is to provide a pavement structure defect identification model and training method for poor interlayer bonding, in order to solve the problems of complex reflected wave characteristics of poor interlayer bonding defects, difficulty in accurately identifying multi-scale defects, scarcity of training data, noise interference, and insufficient lightweighting due to large number of model parameters, thereby significantly improving detection accuracy, robustness and adaptability to actual deployment.

[0007] This disclosure provides a pavement structural defect identification model and training method for poor interlayer bonding. The technical solution is as follows:

[0008] Step S101: Use ground-penetrating radar to detect defects in the road surface structure and perform multi-step signal processing to obtain measured images of road surface defects;

[0009] Step S102: Expand the measured disease image using the BL-GAN model to generate a simulated disease image;

[0010] Step S103: Input the measured defect image and the simulated defect image into the road surface defect recognition model for training, and obtain the trained road surface defect recognition model;

[0011] The BL-GAN model is an adversarial network consisting of a generator and a discriminator.

[0012] The generator uses a 5-layer deconvolutional network to progressively increase the resolution of the initial feature map, and embeds self-attention mechanisms after the second and third layers to obtain high-resolution images.

[0013] The discriminator uses a 4-layer convolutional network to downsample the high-resolution image layer by layer, and embeds a self-attention mechanism after the second and third layers to obtain low-dimensional feature maps.

[0014] The road surface defect identification model is constructed by embedding a lightweight Vision Transformer (ViT) module into the neck network of YOLOv5s, based on the YOLOv5s framework.

[0015] Preferably, the training dataset of the BL-GAN model includes the measured disease images and simulated images, wherein the simulated images are generated using the finite difference time-domain FDTD method.

[0016] Preferably, the BL-GAN model adopts a modular design, with the generator and the discriminator written into different function files.

[0017] Preferably, the BL-GAN model further includes a data loading and preprocessing module.

[0018] The data loading and preprocessing module is defined as an independent function file, used to read the measured disease images and the simulated images from the memory, and convert them into the input format required by the generator and the discriminator.

[0019] Preferably, during the training process of the BL-GAN model, a set of sample images is generated every N iterations, and t-SNE clustering analysis is performed on the sample images and the real images.

[0020] Preferably, the lightweight visual Transformer module is specifically a MobileViTv2 module.

[0021] Correspondingly, a lightweight visual Transformer module is embedded in the neck network of YOLOv5s, specifically as follows:

[0022] Replace the C3 module with 512 channels in the neck network of YOLOv5 with the MobileViTv2 module.

[0023] Preferably, the pavement defect identification model further includes:

[0024] An efficient channel attention (EAC) module is embedded after the feature layer with 1024 channels in the YOLOv5 backbone network.

[0025] Preferably, the pavement defect identification model further includes:

[0026] The YOLOv5 detection head was adjusted to the following scales: 512×512×2, 256×256×2, and 128×128×2. Multi-layer convolution operations were performed on these scales to obtain detailed features of small-target pavement defects.

[0027] Preferably, the model training in step S103 adopts a phased optimization strategy, including a first phase and a second phase;

[0028] The first phase of training is as follows:

[0029] The Adam optimizer is used, and a cosine annealing scheduling strategy is employed to gradually reduce the learning rate in order to train the entire road surface defect identification model until the model converges.

[0030] The second phase of training is as follows:

[0031] The parameters of the feature extraction layer of the backbone network of the road surface defect identification model are frozen. The SGD optimizer is used, and the learning rate is gradually reduced by a polynomial learning rate scheduling strategy. The neck network and the detection head are fine-tuned until the model converges.

[0032] Preferably, in step S101, the multi-step signal processing includes the following steps:

[0033] The original images acquired by the ground-penetrating radar were processed sequentially using a Gaussian filter and a median filter to obtain the filtered images.

[0034] An adaptive background removal algorithm is applied to eliminate the background signal in the filtered image to obtain a background-removed image, which is an image that retains only the abnormal waveform regions related to road surface defects.

[0035] The grayscale distribution of the image after background removal is optimized by using a histogram equalization algorithm and then normalized to obtain the measured disease image.

[0036] The technical solution of this embodiment has the following beneficial effects:

[0037] In this embodiment, a comprehensive solution is provided to address the challenges of low recognition accuracy in pavement structural defects with poor interlayer bonding, including extremely scarce training data, high noise interference from reflected waves, and complex reflected wave characteristics. The solution utilizes a constructed BL-GAN model to achieve high-quality augmentation of reflected wave data. By embedding a lightweight Visual Transformer (VIT) module into the neck network of YOLOv5s, and leveraging Transformer technology to accurately extract pavement defect features in complex backgrounds and multi-scale scenarios, a good balance is achieved between defect recognition accuracy and computational efficiency. This not only reduces the hardware requirements for detection equipment, making the model applicable to more resource-constrained edge devices, but also allows for model training with limited hardware resources and a small number of training samples in scenarios with small-scale datasets of pavement structural defects with poor interlayer bonding. This solves the practical application bottleneck problem in complex defect reflected wave detection. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the identification model and training method for pavement structural defects with poor interlayer bonding provided according to some embodiments of this application.

[0039] Figure 2 This is a technical logic diagram of a pavement structure defect identification model and training method for poor interlayer bonding provided according to some embodiments of this application.

[0040] Figure 3 This is a schematic diagram of a BL-GAN network structure provided according to some embodiments of this application.

[0041] Figure 4This is an exemplary image of a poorly bonded interlayer reflected wave generated by a BL-GAN network according to some embodiments of this application.

[0042] Figure 5 This is a schematic diagram of the network structure of a road surface defect identification model provided according to some embodiments of this application. Detailed Implementation

[0043] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0044] It should be noted that some existing solutions also mention using lightweight visual Transformer modules (such as MobileViT) to improve the YOLO model. However, existing improvements mainly adopt the method of backbone network modification, that is, directly replacing the backbone network of the YOLO model with the MobileViT model. However, this improvement method is not suitable for the scenario of identifying pavement structural defects with poor interlayer bonding. The reason is that replacing the backbone network with the MobileViT model requires a large number of training samples to retrain the backbone network. Since there are a large number of public datasets in general domains, there is no problem of lack of training data for retraining the backbone network. However, in the scenario of identifying pavement structural defects with poor interlayer bonding, the training data uses reflected wave data obtained by ground penetrating radar. This type of data does not have public datasets to support it, is difficult to collect, and has a limited scale of labeled datasets. If the backbone network is modified using MobileViT, the insufficient scale of training data will make it difficult for the backbone network training process to converge.

[0045] In view of this, this embodiment provides a pavement structural defect identification model and training method for poor interlayer bonding. Based on the YOLOv5s framework, a lightweight visual Transformer module is embedded in the neck network of YOLOv5s. Through the improvement of the embedding position, the training efficiency of the backbone network is significantly improved. At the same time, by embedding only the key feature extraction module from the original MobileViTv2 network, instead of replacing the entire MobileViTv2 network as the YOLO backbone feature network, the complexity and parameter scale of the model are greatly reduced, the training difficulty is reduced, the model's requirement for a large training dataset is reduced, and the number of times the lightweight Transformer module is called is reduced, significantly improving the model's training and running efficiency. This allows the constructed detection network (i.e., the pavement defect identification model) to run smoothly even on edge devices with weak performance, expanding the applicable scenarios of the pavement defect identification model and enabling it to be efficiently applied in engineering and exert greater engineering value.

[0046] It should also be noted that although the improvements to the above model reduce the requirement for a large training dataset, the extremely scarce reflected wave data obtained by ground-penetrating radar means that training the model solely with measured damage images still cannot achieve rapid fitting during the training process. To improve the fitting of the YOLOv5s-based pavement damage identification model during training, this embodiment uses a BL-GAN model to augment the reflected wave data obtained by ground-penetrating radar. The augmented simulated damage images and measured damage images are then used to retrain the pavement damage identification model, achieving the goal of rapid training of the pavement damage identification model.

[0047] To address the issue of data augmentation, existing technologies offer several solutions. For example, in computer vision, common augmentation methods include image flipping, cropping, scaling, and translation. These methods all involve geometric transformations of the image. However, the core feature of ground-penetrating radar (GPR) acquired reflection wave data is the reflection wave signal, not just the geometric features of the image. Therefore, the aforementioned methods are not applicable to the scenario described in this embodiment. The BL-GAN model provided in this embodiment is an improvement upon adversarial networks. Compared to traditional adversarial networks (such as SA-GAN networks), the BL-GAN model reduces the use of self-attention mechanisms and moderately increases network depth, making it less dependent on the amount of training data and enabling it to generate high-quality road surface damage images on small datasets.

[0048] Specifically, SA-GAN is designed to improve the model's ability to capture complex features, especially in applications with large-scale datasets, generating more diverse samples such as portraits and natural images. To achieve this, SA-GAN typically relies on a large amount of training data and introduces a self-attention mechanism into each layer's feature map to generate richer sample diversity and detail. However, this method performs poorly on small-scale ground-penetrating radar datasets with poor interlayer bonding because SA-GAN is highly dependent on the amount of data, while images of poorly bonded pavement structures typically have very limited data. When the data volume is insufficient, the quality and diversity of the generated images often fail to meet practical requirements, and may even lead to overfitting.

[0049] In contrast, BL-GAN is designed to generate high-quality disease images under scarce data conditions, primarily for disease detection tasks on small datasets. BL-GAN reduces the amount of self-attention used by embedding a self-attention mechanism in two layers (256 and 128 channels in the generator and discriminator, respectively) and moderately increases network depth, balancing the global structure and detail representation of the generated images. This is particularly effective in modeling disease features in complex geological formations, generating images with high-quality local details. This design allows BL-GAN to maintain training efficiency and improve the effectiveness of training data even with scarce data, avoiding the low training efficiency and overfitting problems encountered by SA-GAN when data is insufficient.

[0050] The embodiments of this application will now be described with reference to the accompanying drawings.

[0051] This embodiment provides a pavement structural defect identification model and training method for poor interlayer bonding, such as... Figure 1 As shown, the method includes:

[0052] Step S101: Ground penetrating radar is used to detect defects in the road surface structure and perform multi-step signal processing to obtain measured images of road surface defects.

[0053] It should be noted that during the data collection phase, the sampled road surface was designed with an asphalt surface layer and a cement-stabilized crushed stone subbase layer underneath. This multi-layered structure resulted in complex and diverse reflection characteristics of the defects.

[0054] In the technical solution provided in this embodiment, during the data acquisition stage, based on radar technology, the LTD-2600 ground penetrating radar is used as a sensor to perform high-resolution defect detection on the road surface.

[0055] Specifically, the antenna frequencies used by ground-penetrating radar include 270MHz, 400MHz, and 1500MHz. By employing multi-frequency antennas, high-resolution reflection images can be provided for the reflection characteristics of defects at different depths.

[0056] In this embodiment, the LTD-2600 ground penetrating radar is mounted on a vehicle platform (road inspection vehicle). By setting key acquisition parameters, such as 16-bit data depth, 128cm / s acquisition speed, 0.542cm trajectory spacing, and multiple time windows, high-resolution and high-fidelity reflected wave signals are ensured to be acquired.

[0057] The multiple time windows can be set to 80, 45, and 12 nanoseconds respectively. By setting multiple time windows, signal characteristics at different depths can be captured more comprehensively, enhancing the detection capability for different underground layers and structures.

[0058] Furthermore, the detection vehicle travels at a constant speed of 20 km / h. During the journey, the ground-penetrating radar system can collect data in real time, ensuring the stability and consistency of the data collection and providing a reliable data foundation for subsequent modeling and analysis.

[0059] In this embodiment, the raw data collected by the ground penetrating radar is a B-Scan image. This type of image is obtained by horizontally moving the radar antenna and recording the changes in the reflected signal at each location, which can represent the changes in the reflected wave signal at different locations over time.

[0060] In this embodiment, the B-Scan image is stored in a high-resolution 16-bit data format to fully record the spatiotemporal characteristics of the disease reflection waveform.

[0061] Considering that the raw data acquired by ground-penetrating radar typically contains noise, artifacts, and other unwanted interference, this embodiment includes a data preprocessing step after acquiring the raw data. By employing a multi-step signal processing procedure in the data preprocessing stage to process the raw data acquired by ground-penetrating radar, the quality and usability of the acquired data are improved, thereby effectively enhancing the accuracy and reliability of subsequent analysis.

[0062] Specifically, in step S101, the multi-step signal processing can be implemented according to the following steps:

[0063] The original images acquired by the ground-penetrating radar were processed sequentially using a Gaussian filter and a median filter to obtain the filtered images.

[0064] An adaptive background removal algorithm is applied to eliminate the background signal in the filtered image, resulting in a background-removed image that retains only the abnormal waveform regions related to road surface defects.

[0065] Histogram equalization algorithm was used to optimize the gray value distribution of the image after background removal and normalization was performed to obtain the measured disease image.

[0066] In the aforementioned multi-step signal processing, the Gaussian filter removes high-frequency noise (such as noise caused by radar wave reflection, weather, etc.) from the image while preserving low-frequency information. The sliding window midpoint filter removes extreme noise from the image. The combination of the two effectively removes noise and smooths the signal. The adaptive background removal algorithm eliminates background signals in the filtered image. By removing irrelevant background information, it enhances the focus on target features (such as road surface defects, cracks, etc.) and retains only abnormal waveform regions to enhance feature focusing. Histogram equalization optimizes the grayscale distribution of the image, improving the visibility and contrast of defect features. Finally, the processed image is normalized, linearly mapping grayscale values ​​to the [0,1] interval, providing a unified high-quality input for subsequent deep learning models.

[0067] Furthermore, the multi-step signal processing is carried out as follows:

[0068] (1) Set the kernel size and standard deviation of the Gaussian filter and apply it to the image matrix.

[0069] (2) Median filtering is achieved by using the sliding window method to replace outliers pixel by pixel, ensuring that the image is smooth and retains the main signal features.

[0070] (3) Apply an adaptive background removal algorithm to eliminate unwanted background signals in the image. Specific operations include:

[0071] (4) Calculate the mean value of each column of pixels and generate a background template;

[0072] (5) Subtract the background template from the original image, keeping only the abnormal waveform areas related to the disease;

[0073] (6) The image gray value distribution is optimized by using a histogram equalization algorithm. The gray value of each pixel is mapped to a new distribution range, making low-contrast details more prominent and enhancing the visibility of disease features in the image.

[0074] (7) Normalize the processed data, linearly mapping the grayscale range of all pixel values ​​to the [0,1] interval. Specific steps include:

[0075] Traverse the image matrix and calculate the global minimum and maximum values;

[0076] Each pixel value is transformed according to a standardized formula to ensure that the output image data has a uniform numerical range.

[0077] Step S102: Expand the measured disease images using the BL-GAN model to generate simulated disease images.

[0078] In this embodiment, the BL-GAN model was proposed to address the problem of scarce actual road damage data. By using the BL-GAN model to expand the measured road damage images, a large number of diverse and highly realistic simulated road damage images can be generated, thereby effectively increasing the size of the training dataset and helping the road damage identification model to better adapt to various road damage detection scenarios, improving its generalization ability and accuracy.

[0079] Here, the extreme scarcity of ground-penetrating radar (GPR) reflection data refers to the fact that relying solely on GPR reflection data is insufficient to meet the training requirements of either the Transformer architecture or traditional GAN ​​networks. For example, in the experimental scenario of this embodiment, only about 400 GPR reflection images could be collected through various means such as field measurements, literature, and publicly available data. This amount of data is far from meeting the data requirements for Transformer network training, and even falls short of the data requirements for traditional GAN ​​network training. Therefore, in the scenario of identifying pavement structural defects with poor interlayer bonding, if we want to apply Transformer in the network architecture to improve the ability to extract complex features, we must first expand the amount of data in the ground-penetrating radar reflection wave images. However, the training process of traditional GAN ​​networks itself requires the support of large-scale datasets, making it difficult to directly apply to data expansion in scenarios where ground-penetrating radar reflection wave data is extremely scarce. Based on this, we must improve traditional GAN ​​networks to meet the needs of high-quality expansion of ground-penetrating radar reflection wave images. In other words, the improvement of the BL-GAN model over the traditional GAN ​​is to generate high-quality simulated defect images when pavement structural defects with poor interlayer bonding are extremely scarce. The BL-GAN model was proposed to ensure that after the Transformer architecture is embedded in the pavement defect identification model, it can quickly converge during the training process and thus accurately identify the characteristics of pavement structural defects with poor interlayer bonding. The two are technically related and form a unified whole.

[0080] Specifically, the BL-GAN model is an adversarial network consisting of a generator and a discriminator. The generator uses a 5-layer deconvolutional network to progressively increase the resolution of the initial feature map, and embeds self-attention mechanisms after the second and third layers to obtain high-resolution images. The discriminator uses a 4-layer convolutional network to progressively downsample the high-resolution image, and embeds self-attention mechanisms after the second and third layers to obtain low-dimensional feature maps.

[0081] The following reference Figure 3The BL-GAN model structure provided in this embodiment will be described in detail.

[0082] like Figure 3 As shown, the BL-GAN model includes a generator and a discriminator. The generator consists of a 5-layer deconvolutional network, which progressively amplifies the resolution of the feature maps through multiple deconvolutional operations to obtain a high-resolution image. The discriminator consists of a 4-layer convolutional network, which progressively downsamples the input high-resolution image to a low-dimensional representation through multiple convolutional operations.

[0083] For example, in the generator's 5-layer deconvolutional network, the feature map sizes of each layer are as follows: 32×32×512, 64×64×256, 128×128×128, 256×256×64, and 512×512×32, that is, the number of channels in each layer are 512, 256, 128, 64, and 32, respectively. Through the generator's 5-layer deconvolutional network, the resolution of the initial feature map is progressively increased, ultimately resulting in a high-resolution image with a size of 512×512×32.

[0084] In the discriminator's four-layer convolutional network, the feature map sizes of each layer are 256×256×64, 128×128×128, 64×64×256, and 32×32×512, respectively, and the number of channels in each layer are 64, 128, 256, and 512, respectively. The high-resolution image (e.g., 512×512×32) is downsampled layer by layer through the four-layer convolutional network to obtain a low-dimensional feature map (32×32×512).

[0085] In this embodiment, self-attention mechanisms are set after the second and third layers of the generator and discriminator.

[0086] Reference Figure 3 The generator's second layer refers to a feature layer with a size of 64×64×256, and the third layer refers to a feature layer with a size of 128×128×128. The discriminator's second layer refers to a feature layer with a size of 128×128×128, and the third layer refers to a feature layer with a size of 64×64×256. By inserting a self-attention mechanism after each of the generator's second and third feature layers, the global feature association and detail representation in the generated image are strengthened. Similarly, by inserting a self-attention mechanism after each of the discriminator's second and third feature layers, the correlation between local details and global features is enhanced.

[0087] This embodiment reduces the complexity of the BL-GAN network and the amount of training data required by using the self-attention mechanism at appropriate locations in the network, instead of using it on all feature layers, thus ensuring training efficiency. Compared with traditional adversarial networks that use the self-attention mechanism, the improved BL-GAN model balances the quality of generated images and the amount of training data required by reducing the amount of self-attention mechanism used and appropriately increasing the network depth. This allows it to expand high-quality training data even when ground-penetrating radar reflection wave data is extremely scarce, thus meeting the training requirements of the road surface defect recognition model with embedded lightweight visual Transformer module.

[0088] To improve the quality of images generated by the BL-GAN model, in some embodiments, the training dataset for the BL-GAN model includes both measured disease images and simulated images, with the simulated images generated using the finite difference time-domain (FDTD) method.

[0089] Specifically, the simulated image is 100-dimensional Gaussian random noise generated by the Finite Difference Time-Domain (FDTD) method, with a Gaussian distribution having a mean of 0 and a standard deviation of 1. Therefore, by inputting the simulated image generated by the FDTD method into the BL-GAN model, the characteristics of the reflection waveforms of diseases under different scenarios can be simulated, filling the feature gaps in the actual data.

[0090] The simulated image is used as the input image of the generator. It is mapped to an initial feature map of 32×32×512 through a fully connected layer. The resolution of the feature map is then increased layer by layer through multiple deconvolution operations, and finally a high-resolution image of 512×512 is generated.

[0091] The input image of the discriminator consists of two parts: a high-resolution image generated by the generator and a measured image of the disease. The discriminator performs multi-layer convolution operations on the input image to gradually downsample it to a low-dimensional representation.

[0092] The BL-GAN model adopts a modular design, with the generator and discriminator written into different function files. Optionally, the BL-GAN model also includes a data loading and preprocessing module, which is defined as an independent function file. This module is used to read measured disease images and simulated images from memory and convert them into the input format required by the generator and discriminator.

[0093] By adopting a modular design and writing the generator, discriminator, and data loading and preprocessing modules into different functional files, the maintainability of the code is improved, making the code clearer and easier to manage. At the same time, the flexibility and scalability of the model are enhanced, making it easier to replace, optimize, and upgrade modules.

[0094] Specifically, the data loading and preprocessing module includes two sub-modules: a data preprocessor and a loader. The data preprocessor adjusts the image size to 512×512 and normalizes pixel values ​​to the range of [-1, 1] to ensure compliance with the input requirements of the generator and discriminator. The loader dynamically segments the preprocessed image to convert it into batch input for the model.

[0095] For example, the batch size can be set to 16. It is understood that the batch size can also be set to other values ​​as needed for the research, and this embodiment does not limit this.

[0096] Supporting dynamic batch segmentation through loaders facilitates efficient data loading and processing by deep learning frameworks.

[0097] In some embodiments, during the training of the BL-GAN model, a set of sample images is generated every N iterations, and t-SNE (t-Distributed Stochastic Neighbor Embedding) clustering analysis is performed on the sample images and the real images (measured disease images).

[0098] In this embodiment, measured disease images and simulated images are used as training datasets. The BL-GAN model is trained by integrating independent files with generator, discriminator and data loading functions. The network parameters are updated in an alternating optimization manner to enhance the model's ability to learn rare disease features.

[0099] As can be seen from the foregoing, the measured disease images are B-Scan images after multi-step signal processing, and these images are marked with information on small cracks and delamination; the simulation images are generated by the FDTD method.

[0100] During training, the generator and discriminator are updated alternately in an adversarial manner. The hyperparameters are set as follows:

[0101] The generator's learning rate is 0.0001;

[0102] The discriminator's learning rate is 0.0004;

[0103] Batch size is 16;

[0104] The total number of training iterations was 500.

[0105] The loss function chosen is binary cross-entropy (BCE) to balance the distribution differences between the real and generated images.

[0106] By updating network parameters in an alternating optimization manner during training, the model's ability to learn rare disease features is enhanced.

[0107] During training, a set of 3×3 sample images is generated every N iterations (e.g., N=10). Figure 4 As shown, this allows for real-time observation of the generator's performance, ensuring the effectiveness of the generator's training at each stage and providing a basis for further model adjustments.

[0108] By performing t-SNE clustering analysis on the generated sample images and real images, we can evaluate whether the generated sample images are close to the real images in the feature space, and verify their authenticity and feature preservation ability.

[0109] Step S103: Input the measured damage images and simulated damage images into the road damage recognition model for training to obtain a trained road damage recognition model.

[0110] The road surface defect identification model is constructed by embedding a lightweight visual Transformer (VIT) module into the neck network of YOLOv5s, based on the YOLOv5s framework.

[0111] In this embodiment, the pavement defect identification model is also called the detection network. Its function is to analyze the input image to be identified during the detection phase of the model, and identify and locate the pavement structural defect targets.

[0112] During the model training phase, the pavement distress identification model is trained based on measured distress images and simulated distress images. The measured distress images are real pavement distress data collected by ground-penetrating radar, including actual pavement structure distresses such as cracks and delamination. The simulated distress images are generated by a BL-GAN network to simulate the reflected waves from distresses of different types and morphologies of pavement structures under different environmental conditions.

[0113] The purpose of training the pavement defect recognition model by inputting measured and simulated defect images into the pavement defect recognition model is to enable the model to learn the characteristics of different forms and types of pavement structural defects so that it can accurately identify defect information in the image to be identified.

[0114] For example, during the training process of the road surface defect recognition model, the acquisition of measured and simulated defect images can be performed according to the following steps:

[0115] (1) Collect images of diseases such as small cracks and delamination from field collection and public datasets, and use them as real images to participate in model training. The resolution is uniformly adjusted to 512×512.

[0116] (2) Use the BL-GAN model to generate simulated disease images to expand the data scale, and filter the generated simulated disease images to remove poor quality samples to ensure the authenticity and diversity of the dataset.

[0117] (3) The collected and expanded dataset was divided into a training set and a validation set in a 7:3 ratio. At the same time, 10% of the real images were separately extracted as a test set to ensure the independence of the test data. Samples were randomly selected during the division to avoid the collection order or generation method causing deviations in the data distribution.

[0118] (4) Use professional annotation software LabelImg to manually annotate all images, marking the location and shape information of the reflected waves of the disease to ensure the accuracy and consistency of the annotation.

[0119] (5) The annotation file is saved in COCO format, including the image file name, size, coordinates of the annotation box (xmin, ymin, width, height) and class label, which is convenient for direct loading during subsequent training.

[0120] (6) According to the COCO dataset standard, the organized data is stored in three folders: training set, validation set and test set. Each folder contains an images folder to store image data and an annotations folder to store the corresponding annotation files.

[0121] (7) In the annotations folder, create three JSON files: instance_train.json, instance_val.json and instance_test.json, which correspond to the annotation information of the training set, validation set and test set respectively.

[0122] In this embodiment, the detection network (road surface defect identification model) is built based on the YOLOv5s framework. It should be noted that YOLOv5 is an object detection model based on a convolutional neural network (CNN), suitable for efficient object detection. YOLOv5s is a lightweight version of YOLOv5, with fewer parameters and lower computational complexity, making it suitable for running on devices with limited hardware resources.

[0123] As can be seen from the background section, convolutional neural networks are limited by their local receptive field and hierarchical structure, resulting in insufficient ability to extract features of road surface defects in complex backgrounds. Therefore, this embodiment improves upon this by embedding a lightweight visual Transformer module into the YOLOv5s framework to enhance its ability to extract road surface defect features in complex backgrounds.

[0124] To further reduce network complexity and ease model training difficulty, this embodiment specifically adopts an improved approach by embedding a lightweight visual Transformer module into the neck network of YOLOv5s. Compared with existing YOLO model improvement schemes, changing the embedding location from the backbone network to the neck network achieves the goal of improving model training and runtime efficiency.

[0125] It should be noted that, in this embodiment, the lightweight vision Transformer module can be MobileViT, EfficientViT, or T2T-ViT, etc. This embodiment does not limit the specific selection of the lightweight vision Transformer module.

[0126] As a preferred embodiment, the lightweight visual Transformer module is specifically the MobileViTv2 module. Correspondingly, the lightweight visual Transformer module is embedded in the neck network of YOLOv5, specifically by replacing the C3 module with 512 channels in the neck network of YOLOv5 with the MobileViTv2 module.

[0127] Compared to other lightweight visual Transformers, the MobileViTv2 module replaces the 512-channel C3 module in the neck network of YOLOv5. The advantages are as follows: MobileViTv2 combines a convolutional network (CNN) with a visual transformer (VIT). The convolutional layers extract local features, while the Transformer module captures global features. This allows the road surface defect recognition model to combine the computational efficiency of CNNs with the global modeling capabilities of Transformers, fully utilizing the lightweight Transformer architecture to achieve global encoding of feature maps. This is beneficial in situations with extremely scarce data and enhances its feature modeling capabilities in complex backgrounds and multi-scale defect scenarios. Furthermore, the combination of CNN and VIT reduces computational overhead, enabling deployment and use on resource-constrained edge devices and improving overall model performance. In contrast, models like EfficientViT and T2T-ViT rely more on Transformer layers to process the feature information of the entire image, potentially leading to higher computational overhead in road surface defect recognition and hindering deployment and application on edge devices.

[0128] Specifically, the MobileViTv2 module uses a lightweight Transformer architecture to globally encode feature maps. The input feature map first extracts local features through a 3×3 convolution, then divides the feature map into 4×4 windows using a slice operation, performing self-attention computation on each window to capture global information. The updated feature map then fuses global and local information through a 1×1 convolution, finally restoring the number of channels to match the input requirements of subsequent detection heads. This module enhances the modeling capability of multi-scale feature maps, particularly excelling in segmenting and locating smaller targets in complex backgrounds.

[0129] As a preferred embodiment, the road surface defect identification model further includes: embedding an efficient channel attention (EAC) module after the feature layer with 1024 channels in the YOLOv5 backbone network, for dynamically adjusting the channel feature weights.

[0130] In this embodiment, after embedding the EAC module into the feature layer with 1024 channels, the correlation between local channels is calculated through 1D convolution, the key channel representation of the deep feature map is optimized, the deep feature map's ability to represent small targets is enhanced, and interference from useless information is suppressed.

[0131] Specifically, in the feature extraction process of YOLOv5, the earlier network layers are mainly responsible for extracting low-level features (such as edges and textures). As the network depth increases, the feature maps become more abstract and complex. Embedding the EAC module into a layer with 1024 channels helps the network to adaptively select more meaningful features, thereby more effectively improving detection accuracy in complex backgrounds and small target scenarios. In other words, in the feature representation of 1024 layers, the network has already captured relatively rich high-level feature information. By weighting the channels after reaching 1024 channels, the attention to important channels can be significantly enhanced, helping the model better identify task-related important features and suppress unimportant features, thus improving the ability to identify diseases in complex backgrounds. Simultaneously, in deeper layers with larger feature map dimensions (such as layers with 1024 channels), directly applying the attention mechanism can reduce the computational burden on unimportant channels, significantly improving the expressive power of the feature maps, thereby optimizing the model's performance and efficiency.

[0132] In some embodiments, the pavement distress identification model further includes: adjusting the YOLOv5 detection head to the following scales: 512×512×2, 256×256×2, 128×128×2, and using multi-layer convolution operations at the scale to obtain detailed features of small target pavement distresses.

[0133] By adjusting the detection head to a scale suitable for small target detection and combining multi-layer convolution operations to extract detailed features, the model can enhance its ability to identify complex and subtle defects such as poor interlayer adhesion.

[0134] The detection head was modified to use scales of 512×512×2, 256×256×2, and 128×128×2 for detecting small targets. At these scales, multiple convolutional operations were added to extract detailed features of small targets. The optimized detection head retains information from low-level feature maps during feature fusion, which is used for the localization and classification of small targets.

[0135] The detection head includes multiple convolutional operations, with the convolutional kernels alternating between 1×1 and 3×3, a stride of 1, and the activation function SiLU. The final output is the class probability, confidence score, and bounding box position for each candidate box. A multi-scale feature fusion strategy optimizes and combines feature maps from different levels to ensure detection accuracy for small and medium-sized targets.

[0136] In summary, the pavement defect identification model provided in this embodiment is built on the YOLOv5s framework. By optimizing the neck network, backbone network, and detection head, it effectively improves the detection performance of small targets in pavement structural defects.

[0137] In a preferred embodiment, in step S103, the training of the pavement defect identification model employs a phased optimization strategy to balance convergence speed and model performance. The phased approach includes a first phase and a second phase. The training process for the first phase is as follows:

[0138] The Adam optimizer was used to achieve fast convergence. The initial learning rate was set to 0.001, the weight decay parameter was set to 0.0005, the momentum parameters β1 and β2 were 0.9 and 0.999 respectively, the training batch size was 256, and 200 iterations were performed, with the learning rate dynamically adjusted in each iteration.

[0139] A cosine annealing scheduling strategy is adopted to gradually reduce the learning rate in order to train the road surface defect identification model as a whole until the model converges, so as to ensure that the model can efficiently explore the solution space and quickly converge to a better solution.

[0140] In the first stage, the entire network is involved in training, including the backbone feature extraction layer (i.e., the feature extraction layer of the main network), the neck network, and the detection head. By comprehensively training the backbone feature extraction layer, the neck network, and the detection head, the network parameters are optimized comprehensively and rapidly, enabling the model to initially establish the ability to extract disease features.

[0141] The second phase of training is as follows:

[0142] The parameters of the feature extraction layer of the backbone network of the road surface defect identification model are frozen. The SGD optimizer is used, and the learning rate is gradually reduced by adopting a polynomial learning rate scheduling strategy. The neck network and the detection head are fine-tuned until the model converges.

[0143] Specifically, the second stage aims at fine-tuning by freezing the backbone network parameters and only fine-tuning the neck network and detection head. This stage focuses on optimizing the multi-scale feature fusion capability of the detection head to enable it to have better detection capabilities for small targets and difficult-to-segment disease areas in complex scenarios.

[0144] For example, the hyperparameter settings in the second stage include:

[0145] Using the SGD optimizer, the initial learning rate was reduced to 0.0001, and the momentum parameter was set to 0.9 to reduce the risk of overfitting while ensuring the stability of gradient updates. The number of iterations was reduced to 100. A Poly learning rate scheduling strategy was adopted, in which the learning rate decreases non-linearly with the number of iterations. The Poly learning rate scheduling strategy optimizes the multi-scale feature fusion capability and improves the performance of small object detection, so as to ensure that the later stages of training can focus on the fine-grained improvement of model performance.

[0146] Throughout the training process, the loss function consists of localization loss, class loss, and confidence loss. Localization loss uses Generalized Intersection over Union (GIoU), class loss uses cross-entropy loss, and confidence loss is calculated using binary cross-entropy. These three loss components are combined in a weighted ratio of 5:1:1 to ensure a balance between the model's capabilities in classification, localization, and confidence assessment.

[0147] After each iteration, the model performance is automatically evaluated on the validation set, and the training strategy is adjusted by monitoring changes in the validation set loss and performance metrics. During training, the validation set performance is monitored in real time using a loss curve. If the validation set performance does not improve for several consecutive epochs, training is stopped early to avoid overfitting.

[0148] Finally, the trained model weights are saved, and inference is performed on the test set to verify the detection results. Key performance indicators such as mAP, localization error, and recall are recorded to verify the accuracy and efficiency of the detection process.

[0149] It should be understood that the training process of the above-mentioned pavement defect identification model is carried out iteratively. After iterative training, the best pavement defect identification model is finally obtained.

[0150] Specifically, after the initial training of the model, the model is evaluated based on performance metrics on the validation and test sets, including key indicators such as loss value change, detection accuracy (mAP), localization error, and recall. If the model performance does not meet expectations, repeated training is performed by adjusting the training hyperparameters and network structure to obtain the optimal model.

[0151] The following is combined Figure 2 The method provided in this embodiment will be described in further detail.

[0152] like Figure 2 As shown, the objective of this embodiment is to provide a high-precision algorithm for detecting internal road surface defects, mainly including the following two aspects: construction and expansion of a road surface ground-penetrating radar (GPR) B-scan dataset, and establishment of a road surface defect detection algorithm. The dataset construction and expansion further includes: adjusting the existing generative adversarial network (GAN) structure to form a BL-GAN model, and then using the BL-GAN model to generate simulated defect images. The training set of the BL-GAN model comes from two sources: firstly, pre-processed GPR data collected in the field; and secondly, simulated images generated using forward modeling with gprMax3.0 software and the FDTD method. After the dataset expansion is completed, the expanded simulated defect images are combined with measured defect images to construct the training dataset for the defect detection model (road surface defect identification model). The pavement defect identification model includes an internal pavement defect detection algorithm, which is built on YOLOv5. It comprehensively considers and compares the performance of the Transformer vision module, attention mechanism, and small target detection under different conditions, and finally forms a defect detection model that combines the MobileViTv2 module, ECA attention module, and three detection heads (detection head 1, detection head 2, and detection head 3), thereby constructing a high-precision internal pavement defect detection algorithm.

[0153] In terms of development environment and tools, the model was developed using the Python programming language, with Visual Studio Code as the primary development tool. A modular design was adopted to achieve efficient development and maintenance of functions. To ensure the management of dependency libraries and the independence of the environment, a virtual environment was created to uniformly install and manage the required tool libraries. Dependencies include: PyTorch and its extension packages; OpenCV; NumPy; SciPy; TqDM; and Matplotlib. The hardware environment was configured with a computer equipped with an NVIDIA RTX 3060 GPU, supporting CUDA 11.3 and cuDNN 8.7, using GPU acceleration to improve training efficiency.

[0154] The generator part of the BL-GAN model is implemented through a separate function file. Its core is a multi-layer deconvolutional network used to generate high-resolution disease images from random noise.

[0155] The generator implementation is organized as a separate file, allowing you to focus on defining the network structure and call it as a separate functional module in the project.

[0156] During the generator training phase, 100-dimensional Gaussian random noise is used as input. The initial noise is mapped to an initial feature map of 32×32×512 through a fully connected layer. The feature map is gradually enlarged to 512×512 using a multi-layer deconvolutional network. The number of channels in each layer are 512, 256, 128, 64 and 32, respectively, and finally a high-resolution image of 512×512 is generated.

[0157] Self-attention mechanisms are embedded in the 256×256 and 128×128 layers to enhance global feature representation. In the deconvolutional layers, the kernel size is set to 4×4, the stride is 2, and the padding is 1. Batch normalization is applied after each layer to stabilize the training process, and the ReLU activation function is used for non-linear transformation. Finally, the output layer normalizes the pixel values ​​to the range [-1, 1] using the tanh activation function.

[0158] The discriminator is also implemented in a separate file and is mainly used to evaluate the authenticity of the input image.

[0159] Specifically, the discriminator downsamples the input image to a low-dimensional representation through multi-layer convolution, incorporates a self-attention mechanism in key layers, and combines the LeakyReLU activation function to improve robustness. The output image represents the probability value of being real or generated. The data loading module integrates actual acquired data and simulation data, standardizes them into a unified format, and dynamically segments them into batches for input.

[0160] Furthermore, the discriminator takes a high-resolution 512×512 image as input and progressively downsamples it to a low-dimensional representation through multiple convolutional operations. Each convolutional layer has a 4×4 kernel size, a stride of 2, and padding of 1, corresponding to 64, 128, 256, and 512 channels respectively. Self-attention mechanisms are embedded in the 256×256 and 128×128 layers to enhance the correlation between local details and global features. The convolutional layers are followed by a LeakyReLU activation function (with a negative slope α = 0.1) to effectively prevent gradient vanishing and enhance the network's robustness. The output layer flattens the feature map into a one-dimensional vector through a fully connected layer and outputs a probability value between 0 and 1 through a sigmoid activation function, representing the likelihood that the input image is real data. The discriminator is designed for efficient implementation and is used as a separate file throughout the project to complete classification and training tasks.

[0161] The BL-GAN model training employs an adversarial optimization strategy between the generator and discriminator, with a total of 500 training iterations. Sample images are generated periodically to monitor performance, effectively improving the realism and diversity of the generated images.

[0162] The code for the road surface defect identification model is organized into modular files, with each part independently encapsulated. The main training script is responsible for integrating all modules and implementing training and testing. The training process is executed in accordance with the process described in the aforementioned embodiments, and will not be elaborated here.

[0163] After the pavement defect identification model is trained, the model weights are saved, and inference verification is performed on the test set. The detection results are output, and key performance indicators such as mAP, localization error, and recall are recorded. The accuracy comparison between the existing model (YOLOv5s) and the improved model (pavement defect identification model) obtained from inference on the test set is shown in the table below:

[0164] Table 1 Accuracy Comparison

[0165]

[0166] As can be seen from Table 1, compared with the traditional YOLOv5s, the road surface defect identification model provided in this embodiment has significantly improved accuracy in all aspects.

[0167] The efficiency comparison between the existing model (YOLOv5s) and the improved model (road surface defect identification model) obtained based on the test set inference is shown in the table below:

[0168] Table 2 Efficiency Comparison

[0169]

[0170] In Table 2, GFLOPs stands for "billions of floating-point operations per second," used to measure the computational complexity of a neural network during inference; FPS stands for "frames per second," measuring the speed of the neural network during inference; and Weights is the total number of all learnable parameters in the neural network, including the weights and biases of each layer. These parameters are learned by the model through training and are used for feature abstraction of input data and prediction of output. As can be seen from Table 2, although the performance indicators of the pavement defect identification model decreased somewhat after embedding the Transformer module in the YOLOv5s framework, the decrease was not significant. This fully demonstrates that the method provided in this embodiment can balance the accuracy and computational efficiency of identifying pavement structural defects with poor interlayer bonding, significantly improving detection accuracy, robustness, and adaptability to practical deployment while ensuring detection efficiency.

[0171] In summary, the method provided in this embodiment has the following beneficial effects:

[0172] (1) Improve the detection performance of complex diseases while taking into account efficiency and training difficulty: The detection of reflected waves of diseases with poor interlayer bonding requires accurate feature extraction in complex backgrounds and multi-scale scenes, while existing methods have obvious contradictions between detection accuracy and computational efficiency.

[0173] This embodiment proposes an optimized design that balances detection accuracy and training feasibility by embedding the lightweight Transformer module MobileViTv2 into the neck network of YOLOv5. Compared to replacing the same 512-layer C3 module in the backbone network, the neck embedding of MobileViTv2 offers significant computational efficiency advantages: the module is only called three times during inference, effectively controlling computational overhead and avoiding dependence on edge device performance after embedding MobileViTv2 in the backbone. Furthermore, embedding MobileViTv2 in the backbone typically requires training a large-scale model from scratch, but data acquisition for poorly bonded disease reflection waves is costly and the dataset size is limited, making backbone optimization impractical. By embedding MobileViTv2 into the neck network, this invention fully leverages the advantages of transfer learning, achieving efficient training with only minor adjustments to the backbone network.

[0174] This design not only significantly improves the accuracy and robustness when dealing with multi-scale disease reflection waves, but also reduces the hardware requirements of detection equipment and the difficulty of model training in small-scale dataset scenarios, thus solving the practical application bottleneck in the detection of complex disease reflection waves.

[0175] (2) Enhance the model’s ability to learn the characteristics of rare diseases: The characteristics of reflected waves from diseases with poor interlayer bonding are usually manifested as complex local details and irregular macroscopic stratigraphic structures, while existing generative adversarial networks have significant limitations in modeling these characteristics.

[0176] The BL-GAN model proposed in this embodiment effectively balances the ability to model global features and capture local details by optimizing the embedding method of the self-attention mechanism and the network structure design, achieving high-resolution generation and optimization of disease images. Specifically, BL-GAN embeds a self-attention mechanism after the 256×256×64 and 128×128×128 feature layers in the generator, enhancing the global feature association and local detail representation of the generated images, enabling the generated disease images to more realistically reproduce complex geological structures. In the discriminator, the self-attention mechanism is embedded in the 128×128×64 and 256×256×32 feature layers, further strengthening the correlation between macroscopic structure and microscopic details, thereby significantly improving the ability to discriminate the realism of the generated images. Compared with SAGAN, BL-GAN achieves a better balance between global features and local details, enabling the generated images to have both global structural consistency and the ability to delicately reproduce the local details of complex reflected waves.

[0177] Furthermore, SAGAN itself relies on large-scale datasets for training, while BL-GAN aims to provide training support for models with small-scale complex datasets such as poor interlayer bonding. The model optimizes the embedding layers and network depth to reduce its own training set requirements and better meet the needs of actual data expansion scenarios.

[0178] (3) Optimizing the model's ability to detect small targets: Addressing the difficulty in accurately detecting small target defects in complex scenarios, this embodiment adjusts the detection head to a new scale specifically for small target feature extraction, replacing the original YOLOv5s scale. Simultaneously, multi-layer convolutional operations refine small target features, combined with a multi-scale feature fusion strategy to retain information from low-level feature maps, ensuring accurate localization and classification of small target defects in complex backgrounds. This optimized design significantly improves the model's ability to detect small target defects, meeting the needs for accurate identification of small defects in practical engineering.

[0179] (4) Improve detection efficiency and maintain high performance: In this embodiment, the model is developed based on the lightweight algorithm YOLOv5s. Under the premise of meeting the above requirements through optimization and improvement, it can still maintain high detection efficiency, ensuring that the model can quickly process a large amount of data in complex disease identification tasks and adapt to real-time detection application scenarios, further improving the practicality and engineering application value of the model.

[0180] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A training method for a pavement structural defect identification model with poor interlayer bonding, characterized in that, include: Step S101: Use ground-penetrating radar to detect defects in the road surface structure and perform multi-step signal processing to obtain measured images of road surface defects; Step S102: Expand the measured disease images using the BL-GAN model to generate simulated disease images; Step S103: Input the measured defect image and the simulated defect image into the road surface defect recognition model for training, and obtain the trained road surface defect recognition model; The BL-GAN model is an adversarial network consisting of a generator and a discriminator. The generator uses a 5-layer deconvolutional network to progressively increase the resolution of the initial feature map, and embeds self-attention mechanisms after the second and third layers to obtain high-resolution images. The discriminator uses a 4-layer convolutional network to downsample the high-resolution image layer by layer, and embeds a self-attention mechanism after the second and third layers to obtain low-dimensional feature maps. The road surface defect identification model is constructed by embedding a lightweight visual Transformer module into the neck network of YOLOv5s, based on the YOLOv5s framework. The lightweight visual Transformer module is specifically the MobileViTv2 module. Correspondingly, the lightweight visual Transformer module is embedded in the neck network of YOLOv5, specifically by replacing the C3 module with 512 channels in the neck network of YOLOv5 with the MobileViTv2 module.

2. The method according to claim 1, characterized in that, The training dataset for the BL-GAN model includes the measured disease images and simulated images, with the simulated images generated using the finite difference time-domain (FDTD) method.

3. The method according to claim 2, characterized in that, The BL-GAN model adopts a modular design, with the generator and the discriminator written into different function files.

4. The method according to claim 3, characterized in that, The BL-GAN model also includes a data loading and preprocessing module. The data loading and preprocessing module is defined as an independent function file, used to read the measured disease images and the simulated images from the memory, and convert them into the input format required by the generator and the discriminator.

5. The method according to claim 2, characterized in that, During the training process of the BL-GAN model, a set of sample images is generated every N iterations, and t-SNE clustering analysis is performed on the sample images and the real images.

6. The method according to claim 1, characterized in that, The road surface defect identification model also includes: An efficient channel attention (EAC) module is embedded after the feature layer with 1024 channels in the YOLOv5s backbone network.

7. The method according to claim 1, characterized in that, The road surface defect identification model also includes: The YOLOv5s detection head was adjusted to the following scales: 512×512×2, 256×256×2, and 128×128×2. Multi-layer convolution operations were performed on these scales to obtain detailed features of small-target pavement defects.

8. The method according to claim 1, characterized in that, In step S103, the training of the pavement defect identification model adopts a phased optimization strategy, including a first phase and a second phase. The first phase of training is as follows: The Adam optimizer is used, and a cosine annealing scheduling strategy is employed to gradually reduce the learning rate in order to train the entire road surface defect identification model until the model converges. The second phase of training is as follows: The parameters of the feature extraction layer of the backbone network of the road surface defect identification model are frozen. The SGD optimizer is used, and the learning rate is gradually reduced by a polynomial learning rate scheduling strategy. The neck network and the detection head are fine-tuned until the model converges.

9. The method according to any one of claims 1 to 7, characterized in that, In step S101, the multi-step signal processing includes the following steps: The original images acquired by the ground-penetrating radar were processed sequentially using a Gaussian filter and a median filter to obtain the filtered images. An adaptive background removal algorithm is applied to eliminate the background signal in the filtered image to obtain a background-removed image, which is an image that retains only the abnormal waveform regions related to road surface defects. The grayscale distribution of the image after background removal is optimized by using a histogram equalization algorithm and then normalized to obtain the measured disease image.

Images