Pavement structure disease recognition model for poor interlayer bonding and training method

Abstract

The invention provides a pavement structure disease identification model for poor interlayer bonding and a training method, and belongs to the technical field of image data processing. The method comprises the following steps: step S101, carrying out defect detection and multi-step signal processing on a pavement structure by adopting a ground penetrating radar to obtain an actually measured disease image of the pavement structure; step S102, expanding the actually measured disease image by using a BL-GAN model to generate a simulated disease image; s103, inputting the actually measured disease image and the simulated disease image into a pavement disease recognition model for training to obtain a trained pavement disease recognition model; the BL-GAN model is an adversarial network composed of a generator and a discriminator; the generator adopts a five-layer deconvolution network structure; the discriminator adopts a four-layer convolutional network structure, and the generator and the discriminator are embedded with a self-attention mechanism behind a specific feature layer; and the pavement disease recognition model is constructed and obtained by embedding a lightweight visual Transform module into a neck network of YOLOv5s. The method can balance the efficiency and precision of the model.

Description

Technical Field

[0001] The present application relates to the technical field of image data processing, and in particular to a pavement structure disease identification model and training method for poor interlayer bonding. Background Art

[0002] The reflected wave signal of ground penetrating radar (GPR) can provide key information about poor interlayer adhesion. Detecting it is one of the effective means to identify pavement structure diseases with poor interlayer adhesion.

[0003] The reflected wave detection of interlayer poor adhesion defects faces technical difficulties such as signal complexity, noise interference and weak features. In some existing studies, convolutional neural networks (including deep neural networks) are used to extract features from defect images obtained by ground penetrating radar to achieve automatic identification of defects. However, convolutional neural networks (including deep neural networks) are limited by their limited local receptive fields and hierarchical structures, which hinder their ability to capture global features and their insufficient ability to extract defect features in complex backgrounds, resulting in low recognition accuracy of the model.

[0004] In recent years, the Transformer model has shown significant advantages in identifying disease features and small target diseases in complex backgrounds with its powerful global feature modeling capabilities. Applying the Transformer to disease detection tasks provides a new idea for improving detection accuracy and adaptability. However, compared with traditional convolutional neural networks or deep neural networks, the complexity and computational complexity of the Transformer network are greatly increased. Its training requires large-scale labeled data as support, and the model convergence is difficult. Existing research focuses on using the Transformer to optimize complex disease recognition, but fails to effectively solve the Transformer model's demand for large-scale labeled data, and does not consider the problem of adapting to deployment in resource-constrained environments. Specifically: on the one hand, the pavement disease data obtained by ground penetrating radar is scarce and the labeling cost is high. The existing Transformer model is prone to overfitting when the data is insufficient, which affects the application effect of actual engineering; on the other hand, the Transformer model usually has a large number of parameters and high computational complexity, which is not conducive to deployment in resource-constrained environments, limiting its promotion in actual engineering. The above problems significantly restrict the efficient application of Transformer technology in pavement disease detection and the full realization of its engineering value.

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

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

[0007] The present disclosure provides a pavement structure disease identification model and training method for poor interlayer bonding, and the technical solution is as follows:

[0008] Step S101: using ground penetrating radar to perform defect detection and multi-step signal processing on the pavement structure to obtain a measured disease image of the pavement structure;

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

[0010] Step S103: inputting the measured disease image and the simulated disease image into a pavement disease recognition model for training to obtain a trained pavement disease 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 deconvolution network to enlarge the resolution of the initial feature map layer by layer, and embeds a self-attention mechanism after the second and third layers respectively to obtain a high-resolution image;

[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 a low-dimensional feature map;

[0014] The pavement disease recognition model is constructed on the basis of the YOLOv5s framework by embedding a lightweight vision transformer (ViT) module in the neck network of YOLOv5s.

[0015] Preferably, the training data set of the BL-GAN model includes the measured disease images and simulated images, and the simulated images are generated by a finite difference time domain FDTD method.

[0016] Preferably, the BL-GAN model adopts a modular design, and the generator and the discriminator are respectively written into different functional 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 functional file, which is used to read the measured disease image and the simulated image 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 group of sample images is generated every N iterations, and t-SNE clustering analysis is performed on the sample images and real images.

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

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

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

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

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

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

[0026] The detection head of YOLOv5 is adjusted to the following scales: 512×512×2, 256×256×2, and 128×128×2, and multi-layer convolution operations are used at the above scales to obtain detailed features of small target road surface diseases.

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

[0028] The training process of the first stage is as follows:

[0029] Using the Adam optimizer and adopting a cosine annealing scheduling strategy to gradually reduce the learning rate, so as to train the entire pavement disease recognition model until the model converges;

[0030] The training process of the second stage is as follows:

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

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

[0033] The original image acquired by the ground penetrating radar is processed by using a Gaussian filter and a median filter in sequence to obtain a filtered image;

[0034] Applying an adaptive background removal algorithm to eliminate background signals in the filtered image to obtain a background-removed image, wherein the background-removed image is an image that only retains abnormal waveform areas related to pavement damage;

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

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

[0037] In this embodiment, in the scenario of pavement structure disease identification with poor interlayer bonding, the training data is extremely scarce, the reflection wave noise interference of the road surface with poor interlayer bonding is large, and the reflection wave characteristics are complex, resulting in low recognition accuracy. An overall solution is provided: the constructed BL-GAN model is used to achieve high-quality expansion of reflection wave data, and the lightweight visual Transformer (VIT) module is embedded in the neck network of YOLOv5s. The Transformer technology is used to achieve accurate extraction of pavement disease features in complex backgrounds and multi-scale scenes, thereby achieving a good balance between disease identification accuracy and computational efficiency. This not only reduces the demand for detection equipment hardware, making the model applicable to more resource-constrained edge devices, but also, in the scenario of a small-scale data set of pavement structure diseases with poor interlayer bonding, only limited hardware resources and a small number of training samples are required to achieve model training, solving the practical application bottleneck problem in complex disease reflection wave detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A schematic flow chart of a pavement structure defect identification model and training method for poor interlayer adhesion provided according to some embodiments of the present application.

[0039] Figure 2 A technical logic diagram of a pavement structure defect identification model and training method for poor interlayer adhesion provided according to some embodiments of the present application.

[0040] Figure 3 A schematic diagram of a BL-GAN network structure provided according to some embodiments of the present application.

[0041] Figure 4An exemplary image of interlayer bonding poor reflection waves generated by a BL-GAN network according to some embodiments of the present application.

[0042] Figure 5 A schematic diagram of the network structure of a pavement damage identification model provided according to some embodiments of the present application. DETAILED DESCRIPTION

[0043] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0044] It should be noted that some existing solutions also mention the use of lightweight visual Transformer modules (such as MobileViT) to improve the YOLO model, but the existing improvements mainly adopt the method of backbone network transformation, that is, directly replacing the backbone network of the YOLO model with the MobileViT model. However, such an improvement method cannot be applied to the pavement structure disease identification scenario with poor interlayer bonding. The reason is that replacing the backbone network with the MobileViT model and retraining the backbone network requires a large number of training samples. Since there are a large number of public data sets in the general field, the retraining of the backbone network will not have the problem of lack of training data. However, in the pavement structure disease identification scenario with poor interlayer bonding, the training data uses the reflection wave data obtained by the ground penetrating radar. This type of data has no public data set as support, is difficult to collect, and the scale of the labeled data set is limited. If the method of MobileViT is used to transform the backbone network, 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 structure disease recognition model and training method for poor interlayer bonding. On the basis of the YOLOv5s framework, a visual lightweight Transformer module is embedded in the neck network of YOLOv5s. By improving the embedding position, the training efficiency of the backbone network is significantly improved. At the same time, by only embedding the key feature extraction module in the original MobileViTv2 network, instead of replacing the MobileViTv2 network as the YOLO backbone feature network as a whole, the complexity and parameter scale of the model are greatly reduced, the difficulty of model training is reduced, and the model's demand for the scale of the training data set is reduced. At the same time, the number of calls to the lightweight Transformer module is reduced, and the model training efficiency and operation efficiency are significantly improved, so that the constructed detection network (i.e., the pavement disease recognition model) can also run smoothly on edge devices with weaker performance, expanding the applicable scenarios of the pavement disease recognition model, so that it can be efficiently applied in engineering and play a better engineering value.

[0046] It should also be noted that although the improvement of the above model reduces the demand for the size of the training data set, due to the extreme scarcity of the reflected wave data obtained by the ground penetrating radar, it is still impossible to achieve rapid fitting of the training process by only training the model through measured disease images. In order to better fit the training process of the road surface disease recognition model built based on YOLOv5s, this embodiment uses the BL-GAN model to expand the reflected wave data obtained by the ground penetrating radar, and retrains the road surface disease recognition model using the expanded simulated disease images and measured disease images, so as to achieve the purpose of rapidly training the road surface disease recognition model.

[0047] The existing technology also provides some solutions to the problem of data expansion. For example, in the field of computer vision, conventional expansion methods include: image flipping, cropping, scaling, translation, etc. These methods are all geometric transformations of images, and the core feature of the reflected wave data obtained by the ground penetrating radar is the reflected wave signal, not just the geometric features of the image. Therefore, the above methods are not applicable to the scenario of this embodiment. The BL-GAN model provided in this embodiment is improved on the basis of the adversarial network. Compared with the traditional adversarial network (such as the SA-GAN network), the BL-GAN model reduces the use of the self-attention mechanism and moderately increases the network depth, making it less dependent on the amount of training data, and can generate high-quality pavement disease images with a small-scale data set.

[0048] Specifically, the design goal of SA-GAN is to improve the model's ability to capture complex features, especially in the application of large-scale data sets, to generate more diverse samples, such as portraits, natural images, etc. To achieve this goal, SA-GAN itself usually relies on a large amount of training data, and introduces a self-attention mechanism for each layer of feature maps to generate richer sample diversity and details. However, this method has a poor application effect on small-scale ground penetrating radar data sets in the case of poor interlayer bonding, because SA-GAN itself is highly dependent on the amount of data, and the images of pavement structure diseases with poor interlayer bonding usually have very little data. When the amount of data is insufficient, the quality and diversity of the generated images often cannot meet the actual needs, and may even lead to overfitting problems.

[0049] In contrast, the design goal of BL-GAN focuses on generating high-quality disease images under scarce data conditions, and is mainly used in disease detection tasks under small-scale data sets. BL-GAN embeds the self-attention mechanism in two layers with 256 and 128 channels in the generator and discriminator, reduces the use of the self-attention mechanism, and moderately increases the network depth, balancing the global structure and detail expression of the generated image. In particular, in the modeling of complex stratum disease features, it can generate images with high-quality local details. This design enables BL-GAN to ensure training efficiency and improve the effectiveness of training data under scarce data, avoiding the low training efficiency and overfitting problems of SA-GAN when data is insufficient.

[0050] The embodiments of the present application are described below in conjunction with the accompanying drawings.

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

[0052] Step S101: using ground penetrating radar to perform defect detection and multi-step signal processing on the pavement structure to obtain a measured disease image of the pavement structure.

[0053] It should be noted that during the data collection stage, the sampling pavement structure was designed as an asphalt surface layer with a cement-stabilized gravel subbase underneath. Its multi-layer structure makes the reflective characteristics of the disease complex and diverse.

[0054] In the technical solution provided in this embodiment, in the data collection stage, based on radar technology, 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 the ground penetrating radar include: 270MHz, 400MHz and 1500MHz. By using a multi-frequency antenna, it is possible to provide high-resolution reflection images for the reflection characteristics of damage at different depths.

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

[0057] Among them, the multiple time windows can be set to 80, 45 and 12 nanoseconds respectively. By setting multiple time windows, the signal characteristics of different depth ranges can be captured more comprehensively, and the detection capability of different underground layers and structures can be enhanced.

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

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

[0060] In this embodiment, the B-Scan image is stored in a high-resolution 16-bit data format so as to completely record the temporal and spatial characteristics of the disease reflection waveform.

[0061] Considering that the raw data acquired by the ground penetrating radar usually contains noise, artifacts and other unnecessary interference, in this embodiment, after acquiring the raw data, a data preprocessing step is also included. By using a multi-step signal processing process in the data preprocessing stage to process the raw data acquired by the ground penetrating radar, the quality and availability of the collected data can be improved, thereby effectively improving 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 image acquired by the ground penetrating radar is processed by using a Gaussian filter and a median filter in sequence to obtain a filtered image;

[0064] 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 only retains the abnormal waveform area related to the pavement disease;

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

[0066] In the above multi-step signal processing steps, the Gaussian filter can remove high-frequency noise in the image (such as noise caused by radar wave reflection, weather and other factors) and retain the low-frequency information of the image. The sliding window median filter can remove extreme noise in the image. The combination of the two can effectively remove noise and smooth the signal. The adaptive background removal algorithm can eliminate the background signal in the filtered image. By removing irrelevant background information in the image, it can enhance the focus on target features (such as pavement diseases, cracks, etc.), and only retain abnormal waveform areas to enhance feature focus. Through histogram equalization processing, the gray value distribution of the image can be optimized, and the visibility and contrast of the disease features can be improved. Finally, the processed image is normalized, and the gray value is linearly mapped to the [0,1] interval to provide a unified high-quality input for the subsequent deep learning model.

[0067] Furthermore, the multi-step signal processing is specifically performed 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 implemented through the sliding window method, and outliers are replaced pixel by pixel to ensure image smoothness and retain the main signal features.

[0070] (3) Apply the adaptive background removal algorithm to eliminate useless background signals in the image. The specific operations include:

[0071] (4) Calculate the mean of each column of pixel values ​​to generate a background template;

[0072] (5) Subtract the background template from the original image and retain only the abnormal waveform area related to the disease;

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

[0074] (7) Normalize the processed data and linearly map the grayscale range of all pixel values ​​to the interval [0,1]. The 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 value range.

[0077] Step S102: using the BL-GAN model to expand the measured disease image to generate a simulated disease image.

[0078] In this embodiment, the BL-GAN model is proposed to solve the problem of scarcity of actual disease data. By using the BL-GAN model to expand the measured disease images, a large number of simulated disease images with diversity and high authenticity can be generated, thereby effectively increasing the size of the training data set, helping the pavement disease recognition model to better adapt to various disease detection scenarios and improve its generalization ability and accuracy.

[0079] Here, the ground penetrating radar reflection wave data is extremely scarce, which means that the reflection wave data obtained by relying solely on the ground penetrating radar can neither meet the training requirements of the Transformer architecture nor the training requirements of the traditional GAN ​​network. For example, in the test scenario of this embodiment, only about 400 ground penetrating radar reflection wave images can be collected through various channels such as actual measurement, literature, and public data. This amount of data is far from the amount of data required for Transformer network training, and even fails to meet the amount of data required for traditional GAN ​​network training. Therefore, in the scenario of pavement structure disease identification with poor interlayer bonding, if you want to apply Transformer in the network architecture to improve the ability to extract complex features, you must first expand the data volume of the ground penetrating radar reflection wave image. The training process of the traditional GAN ​​network itself requires the support of a large-scale data set, and it is difficult to directly apply it to data expansion in the scenario where ground penetrating radar reflection wave data is extremely scarce. Based on this, the traditional GAN ​​network must be improved 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 disease images when the images of pavement structure diseases with poor interlayer bonding are extremely scarce. The BL-GAN model is proposed to ensure that after the Transformer architecture is embedded in the pavement disease recognition model, it can quickly converge during the training process, and then accurately identify the characteristics of pavement structure diseases with poor interlayer bonding. The two are technically related and are a unified whole.

[0080] Specifically, the BL-GAN model is an adversarial network composed of a generator and a discriminator; the generator uses a 5-layer deconvolution network to enlarge the resolution of the initial feature map layer by layer, and embeds a self-attention mechanism after the second and third layers to obtain a high-resolution image; 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 a low-dimensional feature map.

[0081] Refer to the following Figure 3, the BL-GAN model structure provided in this embodiment is described in detail.

[0082] like Figure 3 As shown in Figure 1, the BL-GAN model includes a generator and a discriminator. The generator includes a 5-layer deconvolution network, which enlarges the resolution of the feature map layer by layer through multi-layer deconvolution operations, and finally obtains a high-resolution image. The discriminator includes a 4-layer convolution network, which gradually downsamples the input high-resolution image to a low-dimensional representation through multi-layer convolution operations.

[0083] For example, in the 5-layer deconvolution network of the generator, the feature map sizes of each layer are: 32×32×512, 64×64×256, 128×128×128, 256×256×64, 512×512×32, that is, the number of channels of each layer are: 512, 256, 128, 64 and 32. Through the 5-layer deconvolution network of the generator, the resolution of the initial feature map is enlarged layer by layer, and finally a high-resolution image with a size of: 512×512×32 is obtained.

[0084] In the 4-layer convolutional network of the discriminator, the feature map sizes of each layer are: 256×256×64, 128×128×128, 64×64×256, 32×32×512, and the number of channels of each layer is: 64, 128, 256 and 512. The high-resolution image (such as 512×512×32) is downsampled layer by layer through the 4-layer convolutional network to obtain a low-dimensional feature map (size: 32×32×512).

[0085] In this embodiment, a self-attention mechanism is provided after the second layer and the third layer of the generator and the discriminator respectively.

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

[0087] This embodiment uses the self-attention mechanism at appropriate positions of the BL-GAN network instead of using it for all feature layers, thereby reducing the complexity of the network and the demand for training data, thereby ensuring training efficiency. Compared with the traditional adversarial network using the self-attention mechanism, the improved BL-GAN model reduces the use of the self-attention mechanism and appropriately increases the network depth, thereby balancing the quality of generated images and the demand for training data. In the case of extremely scarce ground penetrating radar reflection wave data, it is still possible to expand high-quality training data to meet the training requirements of the road disease recognition model embedded with a lightweight visual Transformer module.

[0088] To improve the quality of images generated by the BL-GAN model, in some embodiments, the training data set of the BL-GAN model includes measured disease images and simulated images, and the simulated images are generated by a finite difference time domain FDTD method.

[0089] Specifically, the simulated image is a 100-dimensional Gaussian random noise generated by the finite difference time domain FDTD method, and the noise distribution is a Gaussian distribution with 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 reflection waveform characteristics of the disease in different scenarios can be simulated to fill the feature gaps in the actual data.

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

[0091] The input image of the discriminator consists of two parts, one is the high-resolution image generated by the generator, and the other is the measured disease image. 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, and the generator and discriminator are written into different functional files. Optionally, the BL-GAN model also includes a data loading and preprocessing module, which is defined as an independent functional file for reading measured disease images and simulated images from the memory and converting them into the input format required by the generator and discriminator.

[0093] By adopting a modular design and writing the generator, discriminator, 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 submodules: data preprocessor and loader. The data preprocessor is used to adjust the image size to 512×512 and standardize the pixel value to the range of [-1,1] to ensure that it meets the input requirements of the generator and discriminator. The loader is used to dynamically segment the preprocessed image to convert it into batch input of the model.

[0095] Exemplarily, the batch size may be set to 16. It is understandable that the batch size may also be set to other values ​​according to research needs, which is not limited in this embodiment.

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

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

[0098] In this embodiment, measured disease images and simulated images are used as training data sets, and BL-GAN model training is implemented by integrating independent files of generator, discriminator and data loading functions, and network parameters are updated in an alternating optimization manner to enhance the model's learning ability for scarce disease features.

[0099] Based on the above description, it can be known that the measured defect images are B-Scan images after multi-step signal processing, and these images are marked with small cracks and delamination defect information; the simulation images are generated by the FDTD method.

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

[0101] The learning rate of the generator is 0.0001;

[0102] The learning rate of the discriminator is 0.0004;

[0103] The batch size is 16;

[0104] The total number of training iterations is 500;

[0105] The loss function uses binary cross entropy (BCE) to balance the distribution difference between real images and generated images.

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

[0107] During training, a set of 3×3 sample images (such as Figure 4 As shown in the figure, the generator performance can be observed in real time to ensure the training effect of the generator at each stage and provide a basis for further adjusting the model.

[0108] By performing t-SNE cluster analysis on the generated sample images and the real images, it is evaluated whether the generated sample images are close to the real images in the feature space, and its authenticity and feature retention ability are verified.

[0109] Step S103: inputting the measured disease images and the simulated disease images into a pavement disease recognition model for training to obtain a trained pavement disease recognition model.

[0110] Among them, the road surface disease recognition model is constructed based on the YOLOv5s framework by embedding a lightweight visual Transformer (VIT) module in the neck network of YOLOv5s.

[0111] In this embodiment, the pavement defect recognition model is also called a detection network, and its function is to analyze the input image to be recognized in the detection stage of the model, and to recognize and locate the pavement structure defect targets therein.

[0112] In the model training phase, the pavement disease recognition model is trained based on measured disease images and simulated disease images. The measured disease images are real pavement disease data collected by ground penetrating radar, including cracks, delamination and other diseases of the actual pavement structure. The simulated disease images are images generated by the BL-GAN network, which are used to simulate the reflection waves of diseases of different types and forms of pavement structures and different environmental conditions.

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

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

[0115] (1) Images of defects such as small cracks and delamination are collected from field collection and public datasets and used as real images for model training. The resolution is uniformly adjusted to 512×512.

[0116] (2) The BL-GAN model is used to generate simulated disease images to expand the data scale, and the generated simulated disease images are screened to eliminate samples with poor quality to ensure the authenticity and diversity of the data set.

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

[0118] (4) Use the professional labeling software LabelImg to manually label all images and mark the reflected wave position and shape information of the disease to ensure the accuracy and consistency of the labeling.

[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 category label, so as to facilitate direct loading during subsequent training.

[0120] (6) According to the COCO dataset standard, the sorted data is stored as 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: instances_train.json, instances_val.json, and instances_test.json, corresponding to the annotation information of the training set, validation set, and test set, respectively.

[0122] In this embodiment, the detection network (road surface disease recognition model) is implemented based on the YOLOv5s framework. It should be noted that YOLOv5 is a target detection model designed based on a convolutional neural network (CNN), which is suitable for efficient object detection. YOLOv5s is a lightweight version of YOLOv5, which has fewer parameters and lower computational complexity, and is suitable for running on devices with limited hardware resources.

[0123] Based on the description in the background technology section, it can be seen that the convolutional neural network is limited by the local receptive field and hierarchical structure, and its ability to extract disease features in complex backgrounds is insufficient. Therefore, this embodiment improves it by embedding a lightweight visual Transformer module based on the YOLOv5s framework to enhance its ability to extract pavement structure disease features in complex backgrounds.

[0124] In order to further reduce the complexity of the network and ease the difficulty of model training, in this embodiment, a lightweight visual Transformer module is embedded in the neck network of YOLOv5s. Compared with the existing YOLO model improvement scheme, the purpose of improving model training efficiency and operation efficiency is achieved by changing the embedding position from the backbone network to the neck network.

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

[0126] As a preferred embodiment, the lightweight visual Transformer module is specifically: MobileViTv2 module. Correspondingly, the lightweight visual Transformer module is embedded in the neck network of YOLOv5s, specifically: the MobileViTv2 module replaces the C3 module with 512 channels in the neck network of YOLOv5.

[0127] Compared with other lightweight visual transformers, the MobileViTv2 module is selected to replace the C3 module with 512 channels in the neck network of YOLOv5. The advantages are: through MobileViTv2, the convolutional network (CNN) is combined with the visual transformer (VIT). The convolution layer is used to extract local features, while the transformer module is used to capture global features. Therefore, the pavement disease recognition model can combine the computational efficiency of CNN and the global modeling ability of Transformer, and make full use of the lightweight transformer architecture to realize the global encoding ability of feature maps, which is beneficial in the case of extremely scarce data and enhances its feature modeling ability in complex backgrounds and multi-scale disease scenes. At the same time, the combination of CNN and VIT can reduce the computational overhead, enable it to be deployed and used on resource-constrained edge devices, and improve the overall performance of the model. In contrast, models such as EfficientViT and T2T-ViT rely more on the transformer layer to process the feature information of the entire image, which may lead to higher computational overhead in the process of pavement disease recognition, which is not conducive to deployment and application on edge devices.

[0128] Specifically, the MobileViTv2 module globally encodes feature maps through a lightweight Transformer architecture. The input feature map first extracts local features through a 3×3 convolution, and then the feature map is divided into 4×4 windows through a slicing operation. Self-attention calculation is performed on each window to capture global information. The updated feature map fuses global and local information through a 1×1 convolution, and finally restores the number of channels to match the input requirements of the subsequent detection head. The addition of this module enhances the modeling capability of multi-scale feature maps, especially when dealing with the segmentation and positioning of smaller targets in complex backgrounds.

[0129] As a preferred embodiment, the road surface disease recognition model also includes: embedding an efficient channel attention EAC (Channel Attention Module) module after the feature layer with 1024 channels in the YOLOv5 backbone network to dynamically adjust the channel feature weights.

[0130] In this embodiment, after the EAC module is embedded in the feature layer with 1024 channels, the correlation between local channels is calculated through 1D convolution to optimize the key channel expression of the deep feature map, enhance the representation ability of the deep feature map for small targets, and suppress the interference of useless information.

[0131] Specifically, in the feature extraction process of YOLOv5, the previous network layers are mainly responsible for extracting low-level features (such as edges, textures, etc.), and as the network level deepens, the feature map becomes more abstract and complex. After embedding the EAC module into the layer with 1024 channels, it can help the network to further adaptively select more meaningful features, thereby more effectively improving the detection accuracy in complex backgrounds and small targets. In other words, in the feature representation of the 1024-layer, the network has captured relatively rich high-level feature information. By weighting the channels after the 1024-layer channel number, it can significantly enhance the focus on important channels, help the model better identify important features related to the task, and suppress unimportant features, and improve the ability to identify diseases in complex backgrounds. At the same time, in the deep layers with large feature map dimensions (such as the layer with 1024 channels), directly applying the attention mechanism can reduce the computational burden of unimportant channels and significantly improve the expressiveness of the feature map, thereby optimizing the performance and efficiency of the model.

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

[0133] By adjusting the detection head to a scale suitable for small target detection and combining multi-layer convolution operations to extract detail features, the model can enhance the reflection wave recognition capability of complex and subtle defects such as poor interlayer bonding.

[0134] The detection head is modified to detect small targets at different scales: 512×512×2, 256×256×2, and 128×128×2. At this scale, multiple layers of convolution operations are added to extract the detailed features of small targets. The optimized detection head retains the information of low-level feature maps during feature fusion, which is used to locate and classify small targets.

[0135] The detection head includes multiple layers of convolution operations, with the convolution kernel set to alternate 1×1 and 3×3, a step size of 1, and an activation function of SiLU, which ultimately outputs the category probability, confidence, and bounding box position of each candidate box. The feature maps at different levels are optimized and combined through a multi-scale feature fusion strategy to ensure the detection accuracy of small and medium-sized objects.

[0136] In summary, the pavement disease recognition model provided in this embodiment is built based on the YOLOv5s framework, and effectively improves the detection performance of small and medium-sized targets in pavement structure diseases by optimizing the neck network, backbone network and detection head.

[0137] As a preferred embodiment, in step S103, the training of the pavement disease recognition model adopts a phased optimization strategy to balance the convergence speed and model performance. The phases include the first phase and the second phase. The training process of 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, and the training batch size was 256. The training was repeated 200 times, and the learning rate was adjusted dynamically in each iteration.

[0139] The cosine annealing scheduling strategy is used to gradually reduce the learning rate to train the entire pavement disease recognition model until the model converges, ensuring 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 trained, including the backbone feature extraction layer (i.e., the feature extraction layer of the backbone 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 comprehensively and quickly optimized, so that the model can initially establish the ability to extract disease features.

[0141] The training process of the second stage is as follows:

[0142] The parameters of the feature extraction layer of the backbone network of the pavement disease recognition model are frozen, and the SGD optimizer is used. The polynomial Poly learning rate scheduling strategy is adopted to gradually reduce the learning rate, and the neck network and detection head are fine-tuned until the model converges.

[0143] Specifically, in the second stage, with the goal of refined optimization, the backbone network parameters are frozen, and only the neck network and the detection head are fine-tuned. This stage focuses on optimizing the multi-scale feature fusion capability of the detection head, so that it has better detection capabilities for small targets and difficult-to-segment diseased areas in complex scenes.

[0144] Exemplarily, the hyperparameter settings of 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. The Poly learning rate scheduling strategy was adopted, and the learning rate decreased nonlinearly with the number of iterations. The Poly learning rate scheduling strategy was used to optimize the multi-scale feature fusion capability and improve the performance of small target detection, ensuring that the later stage of training can focus on the refined improvement of model performance.

[0146] During the entire training process, the loss function consists of positioning loss, category loss, and confidence loss. Among them, the positioning loss uses the generalized intersection-over-union (GIoU), the category loss uses the cross entropy loss, and the confidence loss is calculated by binary cross entropy. The three parts of the loss are combined in a weighted manner with a weight distribution of 5:1:1 to ensure that the model's capabilities in classification, positioning, and confidence assessment are balanced.

[0147] After each iteration, the model performance is automatically evaluated on the validation set, and the training strategy is adjusted by monitoring the validation set loss and performance indicator changes. During the training process, the validation set performance is monitored in real time through the loss curve graph. If the validation set performance does not improve for multiple epochs in a row, the 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 and record key performance indicators such as mAP, positioning error, and recall rate to verify the accuracy and efficiency of the detection process.

[0149] It should be understood that the training process of the above-mentioned pavement disease recognition model is repeated and iterated. After iterative training, a trained pavement disease recognition model, that is, an optimal model, is finally obtained.

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

[0151] Combine the following Figure 2 The method provided in this embodiment is described in further detail.

[0152] like Figure 2 As shown, the goal of this embodiment is to provide a high-precision pavement internal disease detection algorithm, which mainly includes the following two aspects: construction and expansion of pavement ground penetrating radar B-scan data set, and construction of pavement internal disease detection algorithm. Among them, the data set construction and expansion further include: adjustment of the existing generative adversarial network (GAN) structure to form a BL-GAN model, and then use the BL-GAN model to generate simulated disease images. The training set of the BL-GAN model comes from two aspects, one is the ground penetrating radar data collected on-site after data preprocessing, and the other is the simulated image generated by the forward simulation of gprMax3.0 software and the FDTD method. After the data set expansion is completed, the expanded simulated disease image is combined with the measured disease image to construct a training data set for the disease detection model (pavement disease recognition model). The pavement defect recognition model includes an internal pavement defect detection algorithm, which is built based on YOLOv5. It comprehensively considers and compares the performance of the Transformer visual module, attention mechanism, and small target detection in different situations, and finally forms a defect detection model that combines the MobileViTv2 module, the ECA attention module, and three detection heads (detection head 1, detection head 2, and detection head 3), thereby constructing a high-precision internal pavement detection algorithm.

[0153] In terms of the development environment and tool selection, the model is developed based on the Python programming language, using Visual Studio Code as the main development tool, and adopting a modular design to achieve efficient development and maintenance of functions. To ensure the management of dependent libraries and the independence of the environment, a virtual environment is created to uniformly install and manage the required tool libraries. Dependent libraries include: Pytorch and its extended toolkit; OpenCV; numpy; scipy; tqdm; matplotlib. The hardware environment is configured as a computer equipped with an NVIDIA RTX3060 GPU, supporting CUDA11.3 and cuDNN8.7, and improving training efficiency through GPU acceleration.

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

[0155] The implementation of the generator is organized in the form of an independent file, which can focus on the definition of the network structure and be called as a separate functional module in the project.

[0156] In the generator training stage, 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. A multi-layer deconvolution network is used to gradually enlarge the feature map to 512×512. The number of channels in each layer is 512, 256, 128, 64 and 32 respectively, and finally a 512×512 high-resolution image is generated.

[0157] The self-attention mechanism is embedded in the 256×256 and 128×128 layers to enhance the global feature expression. In the deconvolution layer, the convolution kernel size is set to 4×4, the stride is 2, and the padding is 1. Batch normalization is performed after each layer to stabilize the training process, and the ReLU activation function is used for nonlinear transformation. Finally, the output layer uses the tanh activation function to normalize the pixel value to the range of [-1,1].

[0158] The discriminator part is also implemented in an independent file, which 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, adds a self-attention mechanism in the key layer, combines the LeakyReLU activation function to improve robustness, and outputs the probability value of whether the image is real or generated. The data loading module integrates the actual collected data and the simulated data, standardizes them into a unified format, and dynamically divides them into batch inputs.

[0160] Furthermore, the input of the discriminator is a 512×512 high-resolution image, which is gradually downsampled to a low-dimensional representation through multi-layer convolution operations. The convolution kernel size of each layer is 4×4, the stride is 2, and the padding is 1. The corresponding number of channels is 64, 128, 256, and 512, respectively. The self-attention mechanism is embedded in the 256×256 and 128×128 layers to enhance the correlation between local details and global features. The convolution layer is followed by a LeakyReLU activation function (negative slope α=0.1) to effectively prevent the gradient from disappearing and enhance the robustness of the network. 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 to indicate the possibility that the input image is real data. The design of the discriminator focuses on efficient implementation and is called as an independent file throughout the project to complete classification and training tasks.

[0161] The BL-GAN model training adopts the adversarial optimization strategy between the generator and the discriminator, with a total training iteration of 500 times. Sample images are generated regularly to monitor the performance, effectively improving the authenticity and diversity of the generated images.

[0162] The code of the pavement disease identification model is organized as a modular file, and each part is encapsulated independently. The main training script is responsible for integrating all modules and implementing training and testing. The training process is executed with reference to the process described in the above embodiment, and will not be described in detail here.

[0163] After the pavement damage recognition model training is completed, the trained model weights are saved, and the model is inferred and verified on the test set. The detection results are output and key performance indicators such as mAP, positioning error and recall rate are recorded. The accuracy comparison between the existing model (YOLOv5s) and the improved model (pavement damage recognition model) obtained by inference on the test set is shown in the following table:

[0164] Table 1 Accuracy comparison

[0165]

[0166] It can be seen from Table 1 that compared with the traditional YOLOv5s, the road surface disease recognition model provided by this embodiment has significantly improved accuracy indicators in all aspects.

[0167] The efficiency comparison between the existing model (YOLOv5s) and the improved model (pavement disease recognition model) obtained by reasoning based on the test set is shown in the following table:

[0168] Table 2 Efficiency comparison

[0169]

[0170] In Table 2, GFLOPs is "billion floating point operations per second", which is used to measure the computational complexity of the neural network when performing reasoning; FPS means "frames per second", which measures the speed of the neural network when performing reasoning; Weights is the total number of all learnable parameters in the neural network, including the weights, biases and other parameters of each layer in the network, which 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, after the Transformer module is embedded in the YOLOv5s framework, the performance indicators of the pavement disease recognition model in various aspects have decreased, but the decline is not large, which fully demonstrates that the method provided in this embodiment can balance the accuracy and computational efficiency of the identification of pavement structure diseases with poor interlayer bonding, and significantly improve the detection accuracy, robustness and adaptability to actual deployment while ensuring detection efficiency.

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

[0172] (1) Improving the performance of complex disease detection while taking into account efficiency and training difficulty: The detection of reflection waves of interlayer poor bonding defects requires accurate feature extraction in complex backgrounds and multi-scale scenes, but the existing methods have an obvious contradiction between detection accuracy and computational efficiency.

[0173] This embodiment proposes an optimization design that takes into account both detection accuracy and training feasibility by embedding a lightweight Transformer module MobileViTv2 in the neck network of YOLOv5. Compared with replacing the same 512-layer C3 module into the backbone network, the solution of embedding MobileViTv2 in the neck has a significant advantage in computational efficiency: the module only needs to be called three times during the inference process, thereby effectively controlling the computational overhead and avoiding the dependence of the backbone on the performance of the edge device after embedding MobileViTv2. In addition, embedding MobileViTv2 in the backbone usually requires training a large-scale model from scratch, but the data collection cost of the reflected wave of the interlayer poor bonding disease is high and the data set size is limited. This limitation makes it difficult for the backbone optimization to achieve practicality. By embedding MobileViTv2 into the neck network, the present invention can fully utilize the advantages of transfer learning and only needs to fine-tune the backbone network to achieve efficient training.

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

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

[0176] The BL-GAN model proposed in this embodiment effectively balances the ability of global feature modeling and local detail capture by optimizing the embedding method and network structure design of the self-attention mechanism, and realizes high-resolution generation and optimization of disease images. Specifically, BL-GAN embeds the self-attention mechanism after the 256×256×64 and 128×128×128 feature layers of the generator, which enhances the global feature association and local detail performance of the generated image, so that the generated disease image can more realistically reproduce the complex stratigraphic structure. 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 authenticity of the generated image. Compared with SAGAN, BL-GAN achieves a better balance between global features and local details, so that the generated image has both global structural consistency and can delicately reproduce the local details of complex reflection waves.

[0177] In addition, SAGAN itself relies on large-scale dataset training, while BL-GAN is designed to provide training support for small-scale complex dataset models such as poor inter-layer adhesion. The model optimizes the embedding level 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: In order to address the problem that small target diseases are difficult to detect accurately in complex scenes, the technical solution of this embodiment adjusts the detection head to a new scale specifically used for small target feature extraction, replacing the original YOLOv5s scale. At the same time, the multi-layer convolution operation is used to refine the small target features and combine the multi-scale feature fusion strategy to retain the information of the low-level feature map, ensuring that small target diseases can be accurately located and classified in complex backgrounds. This optimized design significantly improves the model's ability to detect small target diseases and meets the needs for accurate identification of small diseases in actual engineering.

[0179] (4) Improving detection efficiency and maintaining high performance: In the embodiments of the present application, a model is developed based on the lightweight algorithm YOLOv5s. Through optimization and improvement, while meeting the above requirements, it can still maintain a high detection efficiency, ensuring that the model can quickly process large amounts of data in complex disease identification tasks and adapt to application scenarios of real-time detection, further improving the practicality of the model and its engineering application value.

[0180] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A pavement structure disease identification model and training method for poor interlayer adhesion, characterized in that: include: Step S101: using ground penetrating radar to perform defect detection and multi-step signal processing on the pavement structure to obtain a measured disease image of the pavement structure; Step S102: using the BL-GAN model to expand the measured disease image to generate a simulated disease image; Step S103: inputting the measured disease image and the simulated disease image into a pavement disease recognition model for training to obtain a trained pavement disease recognition model; The BL-GAN model is an adversarial network consisting of a generator and a discriminator; The generator uses a 5-layer deconvolution network to enlarge the resolution of the initial feature map layer by layer, and embeds a self-attention mechanism after the second and third layers respectively to obtain a high-resolution image; 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 a low-dimensional feature map; The pavement disease recognition model is constructed based on the YOLOv5s framework by embedding a lightweight visual Transformer module in the neck network of YOLOv5s.

2. The method according to claim 1, characterized in that The training data set of the BL-GAN model includes the measured disease image and the simulated image, and the simulated image is generated by a finite difference time domain FDTD method.

3. The method according to claim 1, characterized in that The BL-GAN model adopts a modular design, and the generator and the discriminator are written into different functional files respectively.

4. The method according to any one of claims 1 to 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 functional file, which is used to read the measured disease image and the simulated image from the memory and convert them into the input format required by the generator and the discriminator.

5. The method according to claim 1, 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 real images.

6. The method according to claim 1, characterized in that The lightweight visual Transformer module is specifically: MobileViTv2 module, Correspondingly, a lightweight visual Transformer module is embedded in the neck network of YOLOv5s, specifically: The MobileViTv2 module replaces the C3 module with 512 channels in the neck network of YOLOv5.

7. The method according to claim 1, characterized in that The pavement disease identification model further includes: The efficient channel attention EAC module is embedded after the feature layer with 1024 channels in the YOLOv5 backbone network.

8. The method according to claim 1, characterized in that The pavement disease identification model further includes: The detection head of YOLOv5 is adjusted to the following scales: 512×512×2, 256×256×2, and 128×128×2, and multi-layer convolution operations are used at the above scales to obtain detailed features of small target road surface diseases.

9. The method according to claim 1, characterized in that: In step S103, the training of the pavement disease recognition model adopts a phased optimization strategy, including a first phase and a second phase; The training process of the first stage is as follows: Using the Adam optimizer and adopting a cosine annealing scheduling strategy to gradually reduce the learning rate, so as to train the entire pavement disease recognition model until the model converges; The training process of the second stage is as follows: The feature extraction layer parameters of the backbone network of the pavement disease recognition model are frozen, and the SGD optimizer is used. The learning rate is gradually reduced by adopting a polynomial Poly learning rate scheduling strategy, and the neck network and the detection head are fine-tuned until the model converges.

10. The method according to any one of claims 1 to 8, characterized in that: In step S101, the multi-step signal processing includes the following steps: The original image acquired by the ground penetrating radar is processed by using a Gaussian filter and a median filter in sequence to obtain a filtered image; Applying an adaptive background removal algorithm to eliminate background signals in the filtered image to obtain a background-removed image, wherein the background-removed image is an image that only retains abnormal waveform areas related to pavement damage; The grayscale value distribution of the image after background removal is optimized by using a histogram equalization algorithm, and normalized to obtain the measured disease image.

Images