A multi-scene road surface pci prediction method based on cross-region transfer learning

Abstract

The present application relates to a kind of multi-scene road surface PCI prediction method based on cross-region migration learning, belong to artificial intelligence traffic field.It includes the following steps: by gathering multi-scene road surface image to construct data set: fixed scene image is obtained by ConvNeXt-Tiny uncertainty road feature, training set image is obtained by ResNet50 pre-training model and road surface image feature extractor road feature;Two inputs transfer adaptation module generate transfer adaptation feature, and with uncertainty feature jointly input road surface uncertainty estimation module, output uncertainty feature;Subsequently fusion road feature, uncertainty feature and transfer adaptation feature, by subdomain adaptation branch processing and input PCI output layer, generate road surface PCI prediction value;By loss function and optimizer iterative training model, test set is input into the model trained and outputs prediction result.The present application can adapt to different road conditions, improve the prediction accuracy.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence in traffic science, specifically involving a multi-scenario road PCI prediction method based on cross-regional transfer learning. Background Technology

[0002] With the acceleration of urbanization and the increase in road traffic demand, the importance of road maintenance and management is becoming increasingly prominent. The Road Surface Condition Index (PCI), as a key indicator for measuring road performance and maintenance condition, plays a crucial role in road maintenance decisions. Traditional PCI assessments mainly rely on manual inspections or statistical analysis-based models. These methods are inefficient, costly, and highly subjective, making them difficult to meet the needs of large-scale, high-frequency applications. Currently, some research has introduced intelligent technologies such as image recognition and sensor data acquisition to achieve automated, data-driven road surface condition assessment. However, these methods still face many challenges in practical applications, particularly in terms of model regional generalization ability and multi-scenario adaptability. The distribution characteristics of road surface data vary significantly across different regions, meaning models are often only applicable to the region where the training data is located, making it difficult to generalize to other regions.

[0003] Furthermore, real-world road environments are diverse, with urban roads, highways, and rural roads differing in structural form, usage intensity, and damage types. A single model cannot adequately cover all scenarios, limiting its effectiveness in complex environments. Additionally, in situations where data is insufficient or unlabeled in some target areas, model training faces the "data cold start" problem.

[0004] To address these issues, transfer learning has become a crucial tool for enhancing the cross-regional adaptability of models. It achieves effective prediction from data-rich regions to data-scarce regions by transferring existing knowledge. Combined with the feature extraction capabilities of deep learning models, it can better capture complex nonlinear relationships and achieve unified modeling of various road surface scenarios. Therefore, there is an urgent need to construct a multi-scenario PCI prediction model based on cross-regional transfer learning to solve the problems of poor regional adaptability and weak scenario generalization in existing methods, improve prediction accuracy and model practicality, and provide technical support for intelligent road maintenance. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a multi-scenario road PCI prediction method based on cross-regional transfer learning.

[0006] To achieve the above objectives, the present invention employs the following technical solution:

[0007] This invention provides a multi-scenario road PCI prediction method based on cross-regional transfer learning, comprising the following steps:

[0008] S1. Collect road surface images from multiple scenes to construct a multi-source road surface image dataset (MSRD); preprocess the images in the dataset to obtain road surface image training data. Fixed images in multiple scenes and road surface image test data ;

[0009] S2. Fixed Images in Multiple Scenes The input is fed into the ConvNeXt-Tiny model for visual feature extraction to obtain the features of the uncertain road surface. ;

[0010] S3. Transfer road image data The input is fed into the ResNet50 pre-trained model to obtain the initial features. ; Initial features of the road surface The image is input into a road surface image feature extractor to obtain road surface features. ;

[0011] S4. Road surface features and uncertain road surface characteristics The input is fed into the transfer adaptation module to obtain transfer adaptation features. ; Migration adaptation characteristics and uncertain road surface characteristics The input is fed into the road surface uncertainty estimation module to obtain the uncertainty characteristics. Road surface characteristics Uncertainty characteristics and migration adaptation characteristics The input is processed in the subdomain adaptation branch to obtain the subdomain adaptation output features. ;

[0012] S5. Adapt subdomains to output features The results are input into the PCI output layer to obtain the road surface PCI prediction results;

[0013] S6. Calculate the error between the road PCI prediction result and the true value using the loss function, and use the Adam optimizer to iteratively optimize the parameters of each module in steps S3 to S5 until the model converges and the trained model is obtained.

[0014] S7. Transfer road surface image test data The data is input into the trained model to obtain the road surface PCI prediction results.

[0015] Further, in step S1, the multi-source road surface image dataset MSRD is constructed:

[0016] Multiple representative regions were selected based on different geographical locations, including urban roads, highways, and rural roads. Each region was divided into multiple scenes based on road surface material, usage frequency, and climate conditions. The PCI labels for each image were collected by a laser road surface detection system and labeled using a combination of manual verification. The images were processed using the Laplacian transform variance method to obtain cleaned images. A multi-source road surface image dataset MSRD was constructed and divided into training set, test set, and fixed set.

[0017] Furthermore, in step S1, the images in the dataset are preprocessed:

[0018] The images in the multi-source road surface image dataset MSRD are uniformly scaled using an image processing library; data augmentation is performed on the images in the training and test sets to generate road surface image training data. and road surface image test data Data augmentation includes random rotation, random horizontal flipping, and color jitter; M images are uniformly extracted from each scene category in a fixed set, and the images are processed using scaling and center cropping to obtain fixed images for multiple scenes. .

[0019] Furthermore, step S3 specifically includes:

[0020] The ResNet50 pre-trained model removes the fully connected layers from the original model;

[0021] The road surface image feature extractor includes a multi-scale residual fusion module, a channel attention enhancement module, and a linear compression module;

[0022] The multi-scale residual fusion module includes three parallel convolutional branches: a first branch, a second branch, and a third branch; initial features The input is fed into three parallel convolutional branches. The outputs of the three branches are concatenated along the channel dimension and then fused through convolutional kernels to obtain the channel-fused features. Initial features and channel fusion features Residual enhancement features are obtained through residual connection. ;

[0023] The channel attention enhancement module includes a global average pooling layer, a first fully connected layer, and a second fully connected layer; the residual enhancement feature... The dimensionality is compressed using a global average pooling layer to obtain compressed features. The compression feature The channel weights are obtained through nonlinear channel weight learning via the first and second fully connected layers. Channel weights With residual enhancement features Perform channel-by-channel multiplication to obtain the weighted features. ;

[0024] In the linear compression module, the weighted features After a linear transformation layer, followed by LayerNorm regularization and GELU activation function, the road surface features are obtained. .

[0025] Furthermore, the migration adaptation module in step S4 specifically includes:

[0026] The transfer adaptation module includes a subdomain mapping layer, a domain contrast encoder, and a self-supervised feature adjustment layer. The subdomain mapping layer includes a linear transformation layer, a normalization layer, and a GELU activation function. The domain contrast encoder includes a main encoder and a shadow encoder. The main encoder includes a third fully connected layer, a normalization layer, and a ReLU activation function. The shadow encoder has the same structure as the main encoder but does not share parameters. The self-supervised feature adjustment layer includes a feature concatenation layer, a fourth fully connected layer, a SiLU activation function, and a Dropout layer.

[0027] The road surface features After processing by the subdomain mapping layer, the subdomain mapping features are obtained. The subdomain mapping features Input the main encoder to obtain the main encoded features The encoding features and main coding features After concatenation using the Concat function in the feature concatenation layer, the features are obtained through a fourth fully connected layer, the SiLU activation function, and a Dropout layer to obtain the transfer adaptation features. .

[0028] Furthermore, the road surface uncertainty estimation module in step S4 specifically includes:

[0029] The road surface uncertainty estimation module is implemented based on the weighted correlation estimation method, and the formula is expressed as follows:

[0030] ,

[0031] in, This represents the GELU activation function; Represents the weight matrix; This indicates element-wise multiplication; b represents the feature concatenation operation; b represents the bias vector. These are the K uncertain road surface reference features extracted under the current prediction target scenario; The uncertainty attention aggregation function is expressed by the following formula:

[0032] ,

[0033] in, represents the attention weight of the k-th reference feature; sim represents the cosine similarity function; This represents the k-th uncertain road surface reference feature extracted in the scenario corresponding to the current prediction target; This represents the j-th uncertain road surface reference feature extracted under the current prediction target scenario; This represents an exponential function with the natural constant e as its base.

[0034] Furthermore, the subdomain adaptation branch in step S4 specifically includes:

[0035] In the subdomain adaptation branch, uncertainty features and migration adaptation characteristics The summation characteristic is obtained after performing element-level addition. Addition characteristics and road surface features The concatenation feature is obtained by concatenating the elements using the Concat function. ; splicing features A linear transformation is performed to reduce the dimensionality, followed by a normalization layer and a ReLU activation function to obtain the desired result. .

[0036] Furthermore, in step S5, the PCI output layer employs a single linear regression unit, and a fully connected layer is used to... The dimension is reduced to 1.

[0037] Furthermore, the loss function described in step S6 uses the mean squared error (MSE), expressed by the following formula:

[0038]

[0039] in, This represents the actual PCI value. This represents the model's predicted value, and N represents the amount of training data used.

[0040] Furthermore, a self-supervised contrastive learning mechanism is introduced into the domain contrastive encoder to address uncertain road surface features. The shadow encoding features are obtained through the shadow encoder. Through main coding features and shadow encoding features Calculate the InfoNCE contrast loss.

[0041] The advantages of this invention are:

[0042] This invention constructs a multi-source road surface image dataset by collecting road surface images in multiple scenarios, effectively improving the diversity and representativeness of the data and enhancing the model's generalization ability in different road surface environments. This enables the model to adapt to road conditions in different regions and with different road surface types, improving prediction accuracy. By using existing base models and ResNet pre-trained models, the method leverages transfer learning to reduce dependence on large amounts of labeled data, especially in cross-regional environments where obtaining labeled data is often difficult. Freezing part of the structure and using the pre-trained model as the initial feature extractor helps improve the model's training efficiency and reduce training time. Uncertain road surface features are introduced and combined with other features through a transfer adaptation module, thereby accurately estimating the uncertainty of road surface conditions. This is significant for high-risk areas or scenarios lacking labeled data, effectively handling the uncertainty of prediction results and enhancing the model's credibility in practical applications. By combining transfer adaptation features with road surface features through a subdomain adaptation branch, the method further optimizes the model's adaptability in the cross-regional transfer learning process. This design enables the model to dynamically adjust according to the characteristics of different regions, ensuring adaptability and prediction accuracy across different regions. This method is not only applicable to traditional road PCI prediction scenarios, but also to effective transfer learning in new scenarios. It is particularly suitable for use across regions, environments, or new environments, solving the problem that traditional models are difficult to adapt to in new scenarios, and has strong versatility and flexibility. Attached Figure Description

[0043] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0044] Figure 1 This is a flowchart of the steps of the method of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Example 1

[0047] In this embodiment, as Figure 1 As shown, this invention provides a multi-scenario road PCI prediction method based on cross-regional transfer learning, the specific steps of which include:

[0048] S1. Collect road surface images from multiple scenes to construct a multi-source road surface image dataset (MSRD); preprocess the images in the dataset to obtain road surface image training data. Fixed images in multiple scenes and road surface image test data .

[0049] Specifically, multiple representative areas were selected based on different geographical locations, including urban roads, highways, and rural roads. Each area was divided into multiple scenarios based on pavement material, usage frequency, and weather conditions. These scenarios included six pavement conditions: clean road, cracks, potholes, patches, slippery surfaces, and water accumulation. A total of 18 specific scenarios were covered. High-definition cameras were uniformly used for data collection, mounted on the front centerline of the data collection vehicle and secured with a shock-resistant mounting bracket. The PCI label for each image was collected by a laser pavement detection system and labeled using a combination of manual verification. The PCI scoring standard followed the "Technical Specification for Maintenance of Asphalt Pavement on Highways" (JTG). H40-2004); using the Laplacian transform variance method, images with sharpness below the threshold (variance value <100), overexposure (average brightness >240), and underexposure (average brightness <30) were removed to obtain cleaned images; finally, no less than 500 valid images were retained for each scene, and all images were stored in a structured image dataset, named as: <region number>_<scene number>_<timestamp>.jpg; the cleaned images were organized according to region and scene to construct the multi-source road surface image dataset MSRD, and divided into training set, test set, and fixed set in a ratio of 8:1:1.

[0050] S2. Fixed Images in Multiple Scenes The input is fed into the ConvNeXt-Tiny model for visual feature extraction to obtain the features of the uncertain road surface. .

[0051] Specifically, an image processing library was used to uniformly scale the images in the multi-source road surface image dataset MSRD to 224×224 pixels; data augmentation was performed on the images in the training and test sets to generate road surface image training data. and road surface image test data Data augmentation includes random rotation, random horizontal flipping, and color jitter; M images are uniformly extracted from each scene category in a fixed set, and the images are processed using scaling and center cropping to obtain fixed images for multiple scenes. .

[0052] S3. Transfer road image data The input is fed into the ResNet50 pre-trained model to obtain the initial features. ; Initial features of the road surface The image is input into a road surface image feature extractor to obtain road surface features. .

[0053] Specifically, the ResNet50 pre-trained model removes the fully connected layers from the original model;

[0054] Remove the fully connected layers (i.e., "fc layers") from the original model, retaining only the global average pooling layer (avgpool): freeze the initial convolutional layer conv1; freeze the first batch normalization layer bn1; freeze the first stage residual block layer1 (containing 3 bottleneck modules); freeze the second stage residual block layer2 (containing 4 bottleneck modules); retain only the third stage (layer3) and the fourth stage (layer4) as trainable states, and only perform gradient updates on these two parts during training to enhance the adaptability to specific tasks.

[0055] The following is the structure of the modified "ResNet50 pre-trained model":

[0056] Convolutional layer (conv1, frozen): kernel size 7×7, stride 2, output channels 64, padding size 3; Batch normalization layer (bn1, frozen); Activation function (ReLU); Max pooling layer: pooling window size 3×3, stride 2; First-stage residual module (layer1, frozen): contains 3 "bottleneck" modules, the structure of which is as follows: Step 1: 1×1 convolution (dimensionality reduction); Step 2: 3×3 convolution (spatial feature extraction); Step 3: 1×1 convolution (dimensionality increase); the first-stage residual module has 512 output channels; Second-stage residual module (layer2, frozen): contains 4 "bottleneck" modules, the second-stage residual module has 512 output channels; Third-stage residual module (layer3, trainable): contains 6... The third-stage residual module has 1024 output channels; the fourth-stage residual module (layer 4, trainable) contains 3 "bottleneck" modules, with 2048 output channels; and a global average pooling layer.

[0057] The road surface image feature extractor includes a multi-scale residual fusion module, a channel attention enhancement module, and a linear compression module;

[0058] Specifically, the multi-scale residual fusion module includes three parallel convolutional branches, namely a first branch, a second branch, and a third branch; the first branch uses a 1×1 convolutional kernel, the second branch uses a 3×3 convolutional kernel, and the third branch uses a 5×5 convolutional kernel; initial features The input is fed into three parallel convolutional branches. The outputs of the three branches are concatenated along the channel dimension and then fused through a 1×1 convolutional kernel to obtain the channel-fused features. Initial features and channel fusion features Residual enhancement features are obtained through residual connection. ;

[0059] The channel attention enhancement module includes a global average pooling layer, a first fully connected layer, and a second fully connected layer. The first fully connected layer has an input dimension of 2048 and an output dimension of 128, and uses the ReLU activation function. The second fully connected layer has an input dimension of 128 and an output dimension of 2048, and uses the Sigmoid activation function. The residual enhancement feature... After being compressed to 2048 dimensions by a global average pooling layer, the compressed features are obtained. The compression feature The channel weights are obtained through nonlinear channel weight learning via the first and second fully connected layers. Channel weights With residual enhancement features Perform channel-by-channel multiplication to obtain the weighted features. ;

[0060] In the linear compression module, the weighted features The dimensionality is reduced from 2048 to 1024 after a linear transformation layer, and then the road surface features are obtained after LayerNorm regularization and GELU activation function. .

[0061] S4. Road surface features and uncertain road surface characteristics The input is fed into the transfer adaptation module to obtain transfer adaptation features. ; Migration adaptation characteristics and uncertain road surface characteristics The input is fed into the road surface uncertainty estimation module to obtain the uncertainty characteristics. Road surface characteristics Uncertainty characteristics and migration adaptation characteristics The input is processed in the subdomain adaptation branch to obtain the subdomain adaptation output features. .

[0062] Specifically, the transfer adaptation module includes a subdomain mapping layer, a domain contrast encoder, and a self-supervised feature adjustment layer; the subdomain mapping layer includes a linear transformation layer, a normalization layer, and a GELU activation function; the linear transformation layer has an input dimension of 1024 and an output dimension of 512; the domain contrast encoder includes a main encoder and a shadow encoder; the main encoder includes a third fully connected layer, a normalization layer, and a ReLU activation function; the third fully connected layer has an input dimension of 512 and an output dimension of 256; the shadow encoder has the same structure as the main encoder but does not share parameters; the self-supervised feature adjustment layer includes a feature concatenation layer, a fourth fully connected layer, a SiLU activation function, and a Dropout layer; the fourth fully connected layer has an input dimension of 768 and an output dimension of 512;

[0063] The road surface features After processing by the subdomain mapping layer, the subdomain mapping features are obtained. The subdomain mapping features Input the main encoder to obtain the main encoded features The encoding features and main coding features After concatenation using the Concat function in the feature concatenation layer, the features are obtained through a fourth fully connected layer, the SiLU activation function, and a Dropout layer to obtain the transfer adaptation features. .

[0064] A self-supervised contrastive learning mechanism is introduced into the domain contrastive encoder to enhance the model's adaptability to cross-regional road surface images. In this mechanism, road surface features extracted from the road surface image are first... After passing through the subdomain mapping layer, the subdomain mapping features are obtained. The data is then fed into the main encoder to generate the main encoded features. Meanwhile, uncertain road surface characteristics Generating shadow-encoded features using a shadow encoder The main encoded features and the shadow encoded features form a contrast sample pair;

[0065] To improve the discriminative power and robustness of the main encoder's feature extraction, the InfoNCE contrastive loss function was used for training. This loss function narrows the distance between the positive sample features of the main encoder and the subdomain mapping, and widens the distance between the main encoder and the negative samples generated by the shadow encoder, thus enabling the model to learn to identify stable PCI features even in the presence of uncertainty and domain shift.

[0066] Finally, after training, only the output of the main encoder is retained as the primary basis for self-supervised feature adjustment. The encoded features and shadow encoded features are concatenated using the Concat function, and then processed through a fourth fully connected layer, the SiLU activation function, and Dropout regularization to obtain more adaptive transfer features. This structure effectively improves the model's prediction accuracy and stability in cross-domain environments.

[0067] Because the goal of contrastive learning is to train the main encoder to learn robust representations, and the shadow encoder only exists as a "contrast reference during training," the main encoder has good generalization ability after training, and the shadow encoder is no longer needed.

[0068] The formula for the InfoNCE loss function is as follows:

[0069] ,

[0070] This represents the features extracted by the main encoder; This represents a positive sample (such as the same image after subdomain mapping). represents negative samples (features extracted by the shadow encoder); sim represents the similarity metric (using cosine similarity); τ=0.29 is the temperature coefficient.

[0071] Specifically, the road surface uncertainty estimation module is implemented based on the weighted correlation estimation method, as expressed in the following formula:

[0072] ,

[0073] in, This represents the GELU activation function; Represents the weight matrix; This indicates element-wise multiplication; b represents the feature concatenation operation; b represents the bias vector. These are the K uncertain road surface features extracted from the scenario corresponding to the current prediction target; The uncertainty attention aggregation function is expressed by the following formula:

[0074] ,

[0075] in, represents the attention weight of the k-th reference feature; sim represents the cosine similarity function; This represents the k-th uncertain road surface reference feature extracted in the scenario corresponding to the current prediction target; This represents the j-th uncertain road surface reference feature extracted under the current prediction target scenario; This represents an exponential function with the natural constant e as its base.

[0076] Specifically, in the subdomain adaptation branch, uncertainty features and migration adaptation characteristics The summation characteristic is obtained after performing element-level addition. Addition characteristics and road surface features The concatenation feature is obtained by concatenating the elements using the Concat function. ; splicing features A linear transformation is performed, reducing the dimension to 256. After passing through a normalization layer and the ReLU activation function, the result is... .

[0077] S5. Adapt subdomains to output features The results are input into the PCI output layer to obtain the road surface PCI prediction results;

[0078] Specifically, the PCI output layer uses a single linear regression unit and a fully connected layer to... The dimension is reduced to 1.

[0079] S6. Calculate the error between the road PCI prediction result and the true value using the loss function, and use the Adam optimizer to iteratively optimize the parameters of each module in steps S3 to S5 until the model converges and the trained model is obtained.

[0080] Specifically, the loss function uses the mean squared error (MSE), and the formula is as follows:

[0081] ,

[0082] in, These are the actual PCI values. is the model's predicted value, and N is the amount of training data.

[0083] S7. Transfer road surface image test data The data is input into the trained model to obtain the road surface PCI prediction results.

[0084] Example 2

[0085] To more comprehensively demonstrate the effectiveness of the method of the present invention, this embodiment compares it with existing methods in specific application scenarios. The advantages of the method of the present invention will be demonstrated below through several specific application scenarios and experimental results.

[0086] This invention enables road surface PCI prediction in different geographical regions (urban, suburban, and mountainous areas). Traditional road surface PCI prediction models typically rely on training with data from a single scene and cannot effectively handle the differences between different regions. To address this issue, the method of this invention significantly improves the cross-regional adaptability of the model by utilizing a multi-source road surface image dataset (MSRD) and cross-regional transfer learning.

[0087] Scene comparison:

[0088] Urban roads: The roads are relatively smooth, but the road surface types are uniform.

[0089] Mountain roads: The road surface conditions are complex, with various types of damage such as potholes and cracks.

[0090] Suburban roads: Roads are often affected by changes in climate and traffic flow, resulting in more complex road surface damage.

[0091] The method of this invention was experimentally compared with existing single-scene prediction methods. For a fair comparison, the performance of the two methods in road PCI prediction was tested under the same scenario.

[0092] Experimental setup: Road surface image data from three regions—urban, mountainous, and suburban—were used (road surface images from different scenarios). Each scenario contained 500 road surface images.

[0093] Model:

[0094] Existing methods: These methods train the model using only a single scene and then predict other scenes. In the comparison with these existing methods, this invention selects a representative, publicly published convolutional neural network (CNN)-based approach as the benchmark. This model has a relatively standard structure, is trained only on a single scene dataset, and does not incorporate transfer learning or uncertainty modeling; it belongs to a typical "single-scene supervised learning" framework.

[0095] The method of this invention: training is performed using a multi-source road surface image dataset (MSRD), and cross-regional transfer learning is used for model adaptation;

[0096] The performance metrics are as follows: Accuracy: the proportion of samples correctly predicted by the model out of the total number of samples; Root Mean Square Error (RMSE): the square root of the error between the predicted value and the true value, reflecting the model's prediction accuracy; Adaptability Score: evaluates the model's performance in different scenarios, i.e., the model's performance when migrating from urban scenes to mountainous or suburban areas. Table 1 shows a comparison of experimental data between the method of this invention and existing methods.

[0097] Table 1 Experimental data of the method of the present invention and existing methods

[0098]

[0099] This table compares the accuracy of existing methods and the method of this invention in different road surface scenarios (urban, mountainous, and suburban). It can be seen that the method of this invention achieves high accuracy in all scenarios, especially in mountainous and suburban scenarios, where the accuracy is significantly higher than that of existing methods. The table also shows a comparison of the RMSE of existing and new methods in different scenarios, reflecting the magnitude of the prediction error. The comparison reveals that the RMSE value of the method of this invention is significantly lower than that of existing methods, indicating its superior accuracy. An adaptability score evaluates the model's transferability across different scenarios. The score ranges from 1 to 10, with 1 indicating poor adaptability and 10 indicating strong adaptability. It can be seen that the method of this invention also significantly outperforms existing methods in terms of adaptability, especially in mountainous and suburban scenarios.

[0100] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-scenario road PCI prediction method based on cross-regional transfer learning, characterized in that, Includes the following steps: S1. Collect road surface images from multiple scenes to construct a multi-source road surface image dataset (MSRD); preprocess the images in the dataset to obtain road surface image training data. Fixed images in multiple scenes and road surface image test data ; S2. Fixed Images in Multiple Scenes The input is fed into the ConvNeXt-Tiny model for visual feature extraction to obtain the features of the uncertain road surface. ; S3. Transfer road image data The input is fed into the ResNet50 pre-trained model to obtain the initial features. ; Initial features of the road surface The image is input into a road surface image feature extractor to obtain road surface features. ; S4. Road surface features and uncertain road surface characteristics The input is fed into the transfer adaptation module to obtain transfer adaptation features. ; to adapt migration features and uncertain road surface characteristics The input is fed into the road surface uncertainty estimation module to obtain the uncertainty characteristics. ; Road surface features Uncertainty characteristics and migration adaptation characteristics The input is processed in the subdomain adaptation branch to obtain the subdomain adaptation output features. ; The road surface uncertainty estimation module specifically includes: The road surface uncertainty estimation module is implemented based on the weighted correlation estimation method, and the formula is expressed as follows: ， in, This represents the GELU activation function; Represents the weight matrix; This indicates element-wise multiplication; b represents the feature concatenation operation; b represents the bias vector. These are the K uncertain road surface reference features extracted under the current prediction target scenario; The uncertainty attention aggregation function is expressed by the following formula: ， in, represents the attention weight of the k-th reference feature; sim represents the cosine similarity function; This represents the k-th uncertain road surface reference feature extracted in the scenario corresponding to the current prediction target; This represents the j-th uncertain road surface reference feature extracted under the current prediction target scenario; This represents an exponential function with the natural constant e as its base. The subdomain adaptation branch specifically includes: In the subdomain adaptation branch, uncertainty features and migration adaptation characteristics The summation characteristic is obtained after performing element-level addition. Addition characteristics and road surface features The concatenation feature is obtained by concatenating the elements using the Concat function. ; splicing features Perform linear transformation to reduce dimensionality, then pass through a normalization layer and the ReLU activation function to obtain... ; S5. Adapt subdomains to output features The results are input into the PCI output layer to obtain the road surface PCI prediction results; S6. Calculate the error between the road PCI prediction result and the true value using the loss function, and use the Adam optimizer to iteratively optimize the parameters of each module in steps S3 to S5 until the model converges and the trained model is obtained. S7. Transfer road surface image test data The data is input into the trained model to obtain the road surface PCI prediction results.

2. The multi-scenario road PCI prediction method based on cross-regional transfer learning according to claim 1, characterized in that, In step S1, the multi-source road surface image dataset MSRD is constructed: Multiple representative regions were selected based on different geographical locations, including urban roads, highways, and rural roads. Each region was divided into multiple scenes based on road surface material, usage frequency, and climate conditions. The PCI labels for each image were collected by a laser road surface detection system and labeled using a combination of manual verification. The images were processed using the Laplacian transform variance method to obtain cleaned images. A multi-source road surface image dataset MSRD was constructed and divided into training set, test set, and fixed set.

3. The multi-scenario road PCI prediction method based on cross-regional transfer learning according to claim 2, characterized in that, In step S1, the images in the dataset are preprocessed: The images in the multi-source road surface image dataset MSRD are uniformly scaled using an image processing library; data augmentation is performed on the images in the training and test sets to generate road surface image training data. and road surface image test data Data augmentation includes random rotation, random horizontal flipping, and color jitter; M images are uniformly extracted from each scene category in a fixed set, and the images are processed using scaling and center cropping to obtain fixed images for multiple scenes. .

4. The multi-scenario road PCI prediction method based on cross-regional transfer learning according to claim 3, characterized in that, Step S3 specifically includes: The ResNet50 pre-trained model removes the fully connected layers from the original model; The road surface image feature extractor includes a multi-scale residual fusion module, a channel attention enhancement module, and a linear compression module; The multi-scale residual fusion module includes three parallel convolutional branches: a first branch, a second branch, and a third branch; initial features The input is fed into three parallel convolutional branches. The outputs of the three branches are concatenated along the channel dimension and then fused through convolutional kernels to obtain the channel-fused features. Initial features and channel fusion features Residual enhancement features are obtained through residual connection. ; The channel attention enhancement module includes a global average pooling layer, a first fully connected layer, and a second fully connected layer; the residual enhancement feature... The dimensionality is compressed using a global average pooling layer to obtain compressed features. The compression feature The channel weights are obtained by nonlinear channel weight learning through the first and second fully connected layers. Channel weights With residual enhancement features Perform channel-by-channel multiplication to obtain the weighted features. ; In the linear compression module, the weighted features After a linear transformation layer, followed by LayerNorm regularization and GELU activation function, the road surface features are obtained. .

5. The multi-scenario road PCI prediction method based on cross-regional transfer learning according to claim 4, characterized in that, The migration adaptation module in step S4 specifically includes: The transfer adaptation module includes a subdomain mapping layer, a domain contrast encoder, and a self-supervised feature adjustment layer. The subdomain mapping layer includes a linear transformation layer, a normalization layer, and a GELU activation function. The domain contrast encoder includes a main encoder and a shadow encoder. The main encoder includes a third fully connected layer, a normalization layer, and a ReLU activation function. The shadow encoder has the same structure as the main encoder but does not share parameters. The self-supervised feature adjustment layer includes a feature concatenation layer, a fourth fully connected layer, a SiLU activation function, and a Dropout layer. The road surface features After processing by the subdomain mapping layer, the subdomain mapping features are obtained. The subdomain mapping features Input the main encoder to obtain the main encoded features The encoding features and main coding features After concatenation using the Concat function in the feature concatenation layer, the features are passed through a fourth fully connected layer, a SiLU activation function, and a Dropout layer to obtain the transfer adaptation features. .

6. The multi-scenario road PCI prediction method based on cross-regional transfer learning according to claim 5, characterized in that, The PCI output layer described in step S5 uses a single linear regression unit and a fully connected layer to... The dimension is reduced to 1.

7. The multi-scenario road PCI prediction method based on cross-regional transfer learning according to claim 6, characterized in that, The loss function described in step S6 uses the mean squared error (MSE), and the formula is as follows: in, This represents the actual PCI value. This represents the model's predicted value, and N represents the amount of training data used.

8. The multi-scenario road PCI prediction method based on cross-regional transfer learning according to claim 7, characterized in that, Introducing a self-supervised contrastive learning mechanism into the domain contrastive encoder to address uncertain road surface features The shadow encoding features are obtained through the shadow encoder. Through main coding features and shadow coding features Calculate the InfoNCE contrast loss.

Images