A method for identifying road hidden diseases based on a visual basic model

Abstract

The application discloses a road hidden disease identification method based on a visual basic model, and relates to the technical field of intelligent transportation systems and road intelligent maintenance. The application generates a two-dimensional gray-scale image containing hidden disease and normal area gray-scale contrast through non-invasive scanning of the deep layer of the road surface by a ground penetrating radar; combines the typical spatial scale of the road disease to divide it into multiple image blocks containing complete visual features; extracts fine-grained visual features in each image block through a visual basic model DINOv2, and converts the fine-grained visual features into a high-dimensional feature vector that can accurately present the internal structure of the road; inputs the high-dimensional feature vector into a table data basic model TabPFN classifier, reasons the feature vector through a Bayesian prior network, and outputs a final disease classification label and a confidence score. The road hidden disease identification framework constructed by combining the DINOv2 and TabPFN models breaks through the generalization ability and training efficiency bottleneck of the traditional machine learning model on the small sample data.

Description

Technical Field

[0001] This application relates to the field of intelligent transportation systems technology, and in particular to a method for identifying hidden road defects based on a visual basic model. Background Technology

[0002] Modern transportation infrastructure is a core support for economic and social operations, and its structural stability directly affects traffic safety and transportation efficiency. Roads, as the fundamental carriers of the transportation network, endure the combined effects of vehicle loads, environmental erosion, and material aging over long periods, making them prone to hidden defects. These defects are concealed within the road structure or beneath the pavement, without obvious surface damage, but already pose a potential threat to the structural load-bearing capacity and stability. Common types include early cracks, voids, interlayer delamination, and underground cavities. With the continuous growth of social development and transportation demands, the scale of road construction continues to expand, and the frequency of use increases significantly, leading to a large number of roads gradually entering the aging stage, making hidden defects increasingly prominent. If these defects are not identified and addressed in a timely manner, they will severely weaken the road structure's strength, induce pavement damage, and even traffic accidents, not only increasing maintenance costs but also potentially causing traffic accidents or even major engineering disasters. Therefore, accurately identifying hidden road defects is one of the core tasks of transportation infrastructure management.

[0003] Traditional road defect detection relies on manual visual inspection and basic sensor technology. Manual inspection requires a large workforce, is inefficient and highly subjective, and struggles to detect underground or internal hazards. Basic sensors, such as strain gauges and pressure sensors, can only monitor the stress state at localized points, failing to achieve wide-area coverage. Ground-penetrating radar (GPR), as a non-destructive detection technology, has become a key means of detecting hidden road defects due to its advantages of being unaffected by weather and lighting conditions and its ability to penetrate the road surface to obtain information about underground structures. GPR emits high-frequency electromagnetic pulses into the ground and receives reflected waves from interfaces or defects in different materials to form radar images, which contain information about the location, morphology, and structural characteristics of the defects. However, GPR image interpretation is highly dependent on the operator's professional experience; manual analysis is inefficient, inconsistent, and prone to overlooking critical defects due to subjective judgment errors, making it difficult to meet the needs of large-scale road inspection.

[0004] Ground-penetrating radar (GPR) technology for detecting hidden road defects can achieve rapid and accurate detection of the internal structure of roads without damaging the road surface. Existing GPR image-based hidden road defect identification technologies are mostly supported by traditional image processing methods or deep learning neural networks. However, these methods are largely based on supervised learning frameworks, and their recognition accuracy is highly dependent on the scale and quantity of manually labeled data. Furthermore, when faced with real-world scenarios involving diverse road materials, complex environmental interference, and diverse defect types, they often exhibit limitations such as insufficient feature extraction capabilities from complex radar images and weak model generalization performance. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, device, medium, and equipment for identifying hidden road defects based on a visual basic model to address the aforementioned technical problems.

[0006] The following technical solution is adopted in this specification:

[0007] This manual provides a method for identifying hidden road defects based on a visual fundamental model, including:

[0008] By using ground-penetrating radar to perform non-invasive scanning of the deep layers of the target road surface, structural reflection information below the road surface layer is captured, and a two-dimensional grayscale image containing grayscale contrast between hidden defects and normal areas is generated.

[0009] The two-dimensional grayscale image is sampled at equal intervals using the tile embedding module of the visual basic model DINOv2, and divided into multiple image blocks containing complete local disease features. Then, the image blocks are encoded into vector tokens through linear projection to achieve structured semantic mapping of the image.

[0010] The location encoding module of the visual basic model DINOv2 processes multiple tokens to capture the local features of each image block and the dependencies between image blocks, thereby obtaining the spatial structure of the image blocks. The location encoding module consists of a layer normalization layer, a multi-head self-attention mechanism module, a secondary normalization layer, and a multilayer perceptron module.

[0011] The spatial structure of image blocks is progressively processed using the multi-layer Transformer encoder of the visual basic model DINOv2 to obtain high-dimensional visual feature vectors.

[0012] The high-dimensional visual feature vector is input into the TabPFN classifier, and the high-dimensional visual feature vector is inferred through the Bayesian prior network to obtain the image-level road defect category label and corresponding confidence score of the target road surface.

[0013] Furthermore, the non-invasive scanning of the road surface depth using ground-penetrating radar to capture structural reflection information below the road surface layer specifically includes:

[0014] Ground-penetrating radar (GPR) is used to perform non-invasive scanning of the deep layers of the road surface, utilizing the difference in dielectric constant between different media to capture structural reflection information at a certain depth below the road surface layer.

[0015] Furthermore, the high-dimensional visual feature vector includes the encoding of electromagnetic scattering patterns and local disease texture feature information in the image patch.

[0016] Furthermore, the Transformer encoder includes:

[0017] The encoders have an identical multilayer structure, with each encoder stacked on top of the others. Each encoder layer includes two components: a self-attention mechanism and a feedforward neural network. The feedforward neural network includes two fully connected layers and a ReLU activation function, with the ReLU activation function connected to each of the two fully connected layers.

[0018] The global information is processed progressively through the multi-layer Transformer encoder.

[0019] Attention weights are generated by calculating the dot product similarity between the query, key, and value matrices, and then the value vectors are weighted and aggregated to capture the global contextual information of the image.

[0020] By learning the attention distribution of multiple feature subspaces in parallel, the model's ability to represent multi-dimensional features of images and its ability to transform nonlinear features are enhanced, resulting in high-dimensional visual feature vectors.

[0021] Furthermore, the complete local disease features of the image block specifically include the gray-level gradient direction features and the continuity features of the reflected waves.

[0022] Furthermore, the fitting and inference process of the TabPFN classifier includes:

[0023] Local sample data is input into the TabPFN base model, and the input data undergoes feature format verification, automatic inference and encoding of category features, and preprocessing of missing values; the local sample data includes structured feature vectors and corresponding road defect category labels;

[0024] Load the TabPFN base model weights to build the inference engine for inference and configure the corresponding integration strategy, and complete the parameter configuration and caching of the TabPFN classifier on local sample data;

[0025] Using the inference engine, based on the distribution of the pre-trained Bayesian prior and the local training dataset, the posterior probability distribution of each disease category of the hidden road surface defects to be identified is calculated, and the disease category label with the highest posterior probability and the corresponding confidence score are output.

[0026] The inference engine includes:

[0027] The one-dimensional feature attention module models the preprocessed high-dimensional feature vector based on the dependencies between feature dimensions, thereby strengthening the feature weights related to the identification of hidden road defects.

[0028] The one-dimensional sample attention module combines the sample representation of local sample data to extract statistical correlation information between the unidentified road surface hidden defects sample and the local sample, forming a sample-level attention representation.

[0029] The multilayer perceptron module takes the comprehensive feature representation obtained based on feature attention and sample attention as input into the multilayer perceptron module, and completes feature reconstruction through nonlinear transformation to obtain high-dimensional discriminative features for classification.

[0030] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects:

[0031] The method for identifying hidden road defects provided in this specification extracts high-dimensional robust deep features with both spatial structure and texture details from ground-penetrating radar images using the DINOv2 model. It also leverages the general data distribution adaptation capabilities accumulated through pre-training of the TabPFN model to effectively capture the complex patterns and subtle differences of hidden defects based on the high-dimensional robust features, thereby achieving accurate type identification of hidden road defects. The two work together and fully utilize the excellent performance of rapid generalization without a large number of annotations in small sample scenarios, breaking through the generalization bottleneck of traditional machine learning models on small sample data. Attached Figure Description

[0032] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0033] Figure 1 This document provides a flowchart illustrating a method for identifying hidden road defects based on a visual fundamental model.

[0034] Figure 2 This is a schematic diagram of a common GPR image category provided in this specification, wherein Figure 2 (a) is a GPR image of the road integrity layer, where Figure 2 (b) is a GPR image of road defects, in which Figure 2 (c) is a GPR image of the loose layer of the road;

[0035] Figure 3 This is a schematic diagram of DINOv2 image feature extraction provided in this specification;

[0036] Figure 4 This document presents a schematic diagram of a method for identifying hidden road defects based on TabPFN. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0038] In recent years, the development of artificial intelligence, especially deep learning technology, has driven the transformation of GPR (Gross Residue Permeable) image recognition for hidden road defects from experience-driven to data-driven approaches. Convolutional Neural Networks (CNNs), as the core architecture of deep learning, have become the mainstream method for processing GPR images due to their advantages in local weight sharing and hierarchical feature extraction. Their technological evolution has undergone an iterative process from basic feature extraction to deep, complex feature modeling.

[0039] LeNet, as an early classic CNN model, was the first to introduce an alternating stacked structure of convolutional and pooling layers into image recognition. By reducing model parameters through local receptive fields and weight sharing, LeNet can effectively extract fundamental features from GPR images. For example, convolutional layers capture the waveform texture and phase changes of reflected waves, while pooling layers compress data dimensionality and retain key features (such as the peak position of reflected waves). The application of LeNet has shifted GPR image recognition from manual feature engineering to automatic feature learning, initially realizing the automated identification of basic diseases and providing a basic paradigm for feature extraction in subsequent models.

[0040] To address the limited feature extraction capabilities of LeNet, AlexNet significantly enhances the model's nonlinear expressive power and generalization performance by increasing network depth and introducing the ReLU activation function and Dropout regularization. In GPR image recognition, AlexNet effectively distinguishes the hallmark features of typical diseases, including the hyperbolic reflection characteristics of underground cavities in GPR images and anomalous amplitude regions in delamination areas. The ReLU function enhances the contrast between features and background noise through nonlinear activation, while Dropout reduces the risk of overfitting. The successful application of AlexNet signifies that deep CNNs are beginning to play a central role in GPR image analysis.

[0041] To address the issue of AlexNet's limited feature scale, VGGNet replaces large convolutional kernels with stacked, standardized small convolutional kernels (3×3). This multi-layered small convolution simulates the receptive field of large convolutions while simultaneously increasing network depth. This structure allows the model to extract more refined local details from GPR images. For example, road layer stripping can cause phase distortion in reflected waves; VGGNet's deep small convolutions can capture these subtle phase changes. The diffusion effect at the edges of underground cavities can also be gradually abstracted into identifiable features through successive small convolutional layers. VGGNet significantly improves the model's accuracy in identifying early-stage latent defects and represents a classic architecture for local feature extraction.

[0042] To address the issue of insufficient adaptability of CNNs to multi-scale features, the Inception architecture (GoogleNet) incorporates a multi-branch convolutional module that performs 1×1, 3×3, and 5×5 convolutions and pooling operations in parallel within the same layer, simultaneously capturing features at different scales. In GPR images, the feature scales of diseases vary significantly. For example, scattering spots caused by cracks belong to micro-texture, while continuous weakly reflective areas in voided regions belong to macro-distribution patterns. Inception's multi-branch structure can extract both types of features simultaneously, enhancing the model's adaptability to multi-scale diseases. Furthermore, the introduction of 1×1 convolutions reduces model parameters and improves computational efficiency, enabling Inception to maintain high accuracy while being suitable for large-scale GPR data processing.

[0043] As network depth increases, the vanishing gradient problem becomes increasingly prominent, leading to a decline in model performance. ResNet effectively alleviates this problem by introducing residual connections, allowing for network depths of hundreds of layers. Deep ResNet possesses hierarchical feature abstraction capabilities, enabling it to handle complex lesions in GPR images. Shallow convolutions extract basic reflected wave features, mid-level convolutions integrate local structural information, and deep convolutions fuse multiple feature classes into a global lesion pattern. The application of ResNet has elevated the accuracy and generalization ability of GPR image recognition to new heights, making it the mainstream model for complex lesion identification.

[0044] The iterative development of existing CNN models has gradually solved the key problems of GPR image feature extraction, namely, from basic waveforms to complex patterns, and from single scale to multi-scale. However, the aforementioned methods for identifying hidden road defects based on convolutional neural network (CNN) models rely heavily on large-scale, high-quality manually labeled data for supervised learning, which faces the following significant limitations in practical engineering applications:

[0045] (1) Strong reliance on manual annotation incurs significant costs.

[0046] Training and optimizing the model requires domain experts to accurately label diseased areas in massive ground-penetrating radar (GPR) images to establish a "feature-label" association. This approach is costly in terms of professional human resources and time, and the labeling process is complex, time-consuming, and prone to subjective biases. Obtaining a sufficient number of labeled samples that comprehensively cover potential disease types and complex scenarios is extremely costly, making it difficult to quickly transfer the technology to new areas or new disease types, severely restricting its large-scale application.

[0047] (2) Insufficient generalization ability makes it difficult to adapt to complex scenarios

[0048] The feature correlations learned by the model are mainly based on limited labeled samples, making it difficult to fully model and adapt to the differentiated electromagnetic wave propagation characteristics and background noise interference caused by diverse roadbed structures and environmental conditions. When dealing with complex scenarios (such as permafrost areas, hot and humid areas, saline areas, etc.) or new disease patterns (such as atypical electromagnetic scattering characteristics caused by special geological structures and the application of new materials), the model's recognition accuracy and reliability are prone to significant decline.

[0049] (3) Performance degrades significantly in small sample scenarios

[0050] When faced with new types of diseases or new areas, labeled data is often insufficient. In such cases, the model lacks enough samples to learn features, resulting in a significant drop in recognition accuracy and a marked increase in false alarm rate. For example, when dealing with unlabeled new types of diseases or new road areas, the model struggles to distinguish between the disease and background noise, failing to meet actual detection needs.

[0051] To address the limitations of existing technologies, the core objective of this invention is to construct a method for identifying hidden road defects that does not rely on large-scale manual annotation, has strong generalization ability, and is suitable for small sample scenarios. This method aims to solve the problems of insufficient sample annotation and poor generalization ability in the identification of hidden road defects (such as underground cavities and cracks). A collaborative identification framework is constructed using the visual base model DINOv2 and the tabular base model TabPFN to achieve high-precision classification. DINOv2 extracts high-dimensional robust features from ground-penetrating radar (GPR) images, while TabPFN accurately classifies the extracted features. Together, they overcome the generalization bottleneck of traditional machine learning models on small sample data.

[0052] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0053] Figure 1 This is a flowchart illustrating a method for identifying hidden road defects based on a visual model, as described in this specification. The method includes the following steps:

[0054] S101: Non-invasive scanning of the deep layers of the road surface is performed using ground penetrating radar to capture structural reflection information below the road surface layer and generate a two-dimensional grayscale image containing grayscale contrast between hidden defects and normal areas.

[0055] S102: The DINOv2 model divides the two-dimensional grayscale image into multiple image blocks containing complete local disease features according to the spatial scale of road defects; and extracts fine-grained features from each image block and converts them into high-dimensional visual feature vectors.

[0056] refer to Figure 3 DINOv2 is a self-supervised vision model based on the Transformer architecture. Its core design goal is to learn general and robust image feature representations under unlabeled data conditions. In the feature extraction process for GPR images, DINOv2's patch embedding module discretizes the input GPR image into fixed-size image patches and encodes each image patch into a vector token through a linear projection operation, achieving structured semantic mapping of the image. These tokens are then sequentially processed by the DINOv2 positional encoding module through layer normalization, multi-head self-attention mechanism, secondary normalization, and multilayer perceptron, gradually completing the extraction of local image features and the integration of global information.

[0057] The Transformer encoder consists of multiple stacked encoder layers with the same structure. Each layer comprises two core components: a self-attention mechanism and a feed-forward network (FNN). The self-attention mechanism generates attention weights by calculating the dot product similarity between the query, key, and value matrices, and then weights and aggregates the value vectors to capture global contextual information of the image. The multi-head attention mechanism further enhances the model's ability to represent multi-dimensional image features by learning the attention distribution across multiple feature subspaces in parallel. The feed-forward network consists of two fully connected layers and an intermediate ReLU activation function, which enhances the non-linear transformation capability of features and deepens the understanding of complex image patterns.

[0058] After progressive processing by multiple Transformer encoders, the model finally outputs a high-dimensional visual feature vector of dimension 384. This feature vector effectively encodes electromagnetic scattering patterns and local disease features in GPR images, possessing both global contextual relevance and local detail discriminative power. Based on a self-supervised learning paradigm, DINOv2 can learn highly discriminative semantic features without relying on manually labeled data, providing a high-quality feature foundation for subsequent disease identification tasks.

[0059] S103: The TabPFN classifier uses a Bayesian prior network to infer the high-dimensional visual feature vector and outputs image-level road defect classification labels and confidence scores.

[0060] TabPFN is a table-based model based on Bayesian prior networks, designed to solve classification tasks under limited sample conditions. Built around a Bayesian inference framework, it introduces prior distributions into the model parameters and, combined with the likelihood function of the input data, derives the posterior probability distributions for each category based on Bayes' theorem, thus achieving accurate predictions with limited supervised information. Unlike traditional classification models that rely on large amounts of labeled data to fit parameters, TabPFN's prior distributions incorporate domain knowledge and possess a significant regularization effect, effectively mitigating overfitting in small-sample scenarios, making it particularly suitable for applications with scarce labeled data.

[0061] In this invention, reference Figure 4 TabPFN takes a 384-dimensional visual feature vector extracted from DINOv2 as input. This feature vector encodes the electromagnetic scattering patterns and disease texture features in the GPR image. The input features are sequentially processed through three core modules of the two-dimensional TabPFN layer: 1D Feature Attention, 1D Sample Attention, and Multilayer Perceptron (MLP). The 1D Feature Attention module focuses on modeling the dependencies between feature dimensions, enhancing the effectiveness of feature representation by strengthening the weights of discriminative features. The 1D Sample Attention module aims to capture statistical associations between samples, enhancing the model's small-sample generalization ability by mining common information in a limited number of samples. The MLP module completes feature reconstruction through nonlinear transformations, deepening the association between features and disease categories. Finally, the model combines the outputs of the above modules with a pre-set Bayesian prior to calculate the posterior probability distribution of each category and outputs the disease category label with the highest posterior probability and its corresponding confidence score. Thanks to the organic integration of structured attention mechanism and Bayesian inference principle, TabPFN can still achieve high-precision disease identification even in scenarios with limited labeled samples.

[0062] To verify the effectiveness of this technical solution, the present invention designed a three-stage progressive experiment and conducted tests based on the GPR image dataset. The experiment consisted of two parts: a training set and a test set. The size of the training set gradually increased, while the size of the test set remained constant. The evaluation metrics included accuracy and F1 score.

[0063] refer to Figure 2The GPR image dataset used in the experiment covers three typical sample types: intact layer (pavement structure without defects), damaged layer (containing hidden defects such as voids and cracks), and loose layer (base material with insufficient compaction). The training set has three gradient sample sizes: dataset 1 (100 samples), dataset 2 (200 samples), and dataset 3 (300 samples); the three training sets share the same validation set (100 samples) to eliminate the interference of validation set differences on performance evaluation.

[0064] Table 1 Experimental Results

[0065]

[0066] The experimental results in Table 1 show that the model performance steadily increases with the increase of the training sample size. DINOv2, based on self-supervised pre-training, can effectively capture robust features in GPR images without additional fine-tuning, and TabPFN can achieve high-precision identification of hidden road defects with only a small number of samples in small-sample scenarios.

[0067] The key feature of this invention lies in addressing the core scientific challenge of achieving high-precision identification of hidden road defects under small sample conditions—a crucial step in the application of ground-penetrating radar (GPR) detection technology to intelligent road maintenance. Existing methods for identifying hidden road defects are largely based on supervised deep learning frameworks, heavily reliant on training models with numerous manually labeled radar image samples. However, the difficulty and cost of collecting and labeling hidden defect samples hinder the rapid deployment and limit the generalization capabilities of these methods. This invention addresses this pain point by integrating the self-supervised feature extraction capabilities of the visual model DINOv2 with the small-sample classification advantages of the tabular model TabPFN. It constructs an end-to-end identification framework of "self-supervised feature learning - precise small-sample classification," focusing on high-dimensional feature mining and small-sample classification of hidden defects in GPR images. This alleviates the contradiction between insufficient feature extraction and low classification accuracy in small-sample scenarios, achieving efficient and accurate identification of hidden road defects.

[0068] The innovation of this invention is reflected in:

[0069] (1) In view of the limitations of existing methods that rely on a large number of manually labeled samples, this invention draws on the research ideas of self-supervised learning and makes full use of the self-supervised pre-training mechanism of DINOv2. It can extract high-dimensional robust deep features from ground-penetrating radar images without manually labeled data. Through self-supervised contrastive learning tasks, it learns the general visual representation of hidden diseases in radar images, accurately captures the spatial structure features and gray-scale texture features of disease areas, breaks through the strong dependence of supervised methods on labeled data, and provides low-labeling cost and high-discrimination feature support for disease classification under small sample conditions.

[0070] (2) To address the problem of insufficient classification accuracy in small sample scenarios, this invention combines Bayesian inference with the small sample learning capability of TabPFN to propose a Bayesian-enhanced small sample classification method. By utilizing the general data distribution adaptation capability of TabPFN pre-training and integrating the uncertainty estimation and probabilistic modeling of Bayesian inference, the association boundary between disease features and types is quickly constructed by probabilistically learning the feature distribution of a small number of labeled samples. This solves the problems of model overfitting and weak generalization ability in small sample scenarios, and achieves high-precision discrimination of latent disease types. Even under the condition of a small number of labeled samples, it can still maintain high classification accuracy.

[0071] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for identifying hidden road defects based on a visual fundamental model, characterized in that, include: By using ground-penetrating radar to perform non-invasive scanning of the deep layers of the target road surface, structural reflection information below the road surface layer is captured, and a two-dimensional grayscale image containing grayscale contrast between hidden defects and normal areas is generated. The two-dimensional grayscale image is sampled at equal intervals using the tile embedding module of the visual basic model DINOv2, and divided into multiple image blocks containing complete local disease features. Then, the image blocks are encoded into vector tokens through linear projection to achieve structured semantic mapping of the image. The location encoding module of the visual basic model DINOv2 processes multiple tokens to capture the local features of each image block and the dependencies between image blocks, thereby obtaining the spatial structure of the image blocks. The location encoding module consists of a layer normalization layer, a multi-head self-attention mechanism module, a secondary normalization layer, and a multilayer perceptron module. The spatial structure of image blocks is progressively processed using the multi-layer Transformer encoder of the visual basic model DINOv2 to obtain high-dimensional visual feature vectors. The high-dimensional visual feature vector is input into the TabPFN classifier, and the high-dimensional visual feature vector is inferred through the Bayesian prior network to obtain the image-level road defect category label and corresponding confidence score of the target road surface. The fitting and inference process of the TabPFN classifier includes: Local sample data is input into the TabPFN base model, and the input data undergoes feature format verification, automatic inference and encoding of category features, and preprocessing of missing values; the local sample data includes structured feature vectors and corresponding road defect category labels; Load the TabPFN base model weights to build the inference engine for inference and configure the corresponding integration strategy, and complete the parameter configuration and caching of the TabPFN classifier on local sample data; The inference engine calculates the posterior probability distribution of each disease category of the road surface hidden disease sample to be identified based on the distribution of the pre-trained Bayesian prior and the local training dataset, and outputs the disease category label with the highest posterior probability and the corresponding confidence score. The inference engine includes: The one-dimensional feature attention module models the preprocessed high-dimensional feature vector based on the dependencies between feature dimensions, thereby strengthening the feature weights related to the identification of hidden road defects. The one-dimensional sample attention module combines the sample representation of local sample data to extract statistical correlation information between the unidentified road surface hidden defects sample and the local sample, forming a sample-level attention representation. The multilayer perceptron module takes the comprehensive feature representation obtained based on feature attention and sample attention as input into the multilayer perceptron module, and completes feature reconstruction through nonlinear transformation to obtain high-dimensional discriminative features for classification. The high-dimensional visual feature vector includes the encoding of electromagnetic scattering patterns and local disease texture feature information in the image patch; The complete local disease features of the image block specifically include the gray-level gradient direction features and the continuity features of the reflected waves.

2. The method for identifying latent road defects based on a visual fundamental model as described in claim 1, characterized in that, The method of using ground-penetrating radar to perform non-invasive scanning of the deep layers of the road surface and capture structural reflection information below the road surface layer specifically includes: Ground-penetrating radar (GPR) is used to perform non-invasive scanning of the deep layers of the road surface, utilizing the difference in dielectric constant between different media to capture structural reflection information at a certain depth below the road surface layer.

3. The method for identifying latent road defects based on a visual fundamental model as described in claim 1, characterized in that, The Transformer encoder includes: The encoders have an identical multilayer structure, with each encoder stacked on top of the others. Each encoder layer includes two components: a self-attention mechanism and a feedforward neural network. The feedforward neural network includes two fully connected layers and a ReLU activation function, with the ReLU activation function connected to the two fully connected layers respectively. The global information is processed progressively through the multi-layer Transformer encoder. Attention weights are generated by calculating the dot product similarity between the query, key, and value matrices, and then the value vectors are weighted and aggregated to capture the global contextual information of the image. By learning the attention distribution of multiple feature subspaces in parallel, the model's ability to represent multi-dimensional features of images and its ability to transform nonlinear features are enhanced, resulting in high-dimensional visual feature vectors.

Images