Multi-scene pavement PCI (Peripheral Component Interconnect) prediction method based on cross-regional transfer learning

Abstract

The invention relates to a multi-scene pavement PCI (Peripheral Component Interconnect) prediction method based on cross-regional transfer learning, and belongs to the field of artificial intelligence traffic science. The method comprises the following steps that a data set is constructed by collecting multi-scene road surface images, uncertain road surface features are obtained from fixed scene images through ConvNeXt-Tiny, and road surface features are obtained from training set images through a ResNet50 pre-training model and a road surface image feature extractor; the two features are input into a migration adaptation module to generate migration adaptation features, the migration adaptation features and the uncertainty features are jointly input into a road surface uncertainty estimation module, and the uncertainty features are output; fusing the pavement features, the uncertainty features and the migration adaptive features, performing sub-domain adaptive branch processing and inputting the features into a PCI output layer, and generating a pavement PCI predicted value; and iteratively training the model through a loss function and an optimizer, inputting the test set into the trained model, and outputting a prediction result. The method can adapt to different road conditions, and the prediction accuracy is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of artificial intelligence traffic, and particularly relates to a multi-scene road surface PCI prediction method based on cross-region transfer learning. BACKGROUND

[0002] With the acceleration of urbanization and the increase of road traffic demand, the importance of road maintenance and management is increasingly prominent. As a key indicator for measuring road use performance and maintenance condition, pavement condition index (PCI) plays an important role in road maintenance decision-making. Traditional PCI evaluation mainly relies on manual inspection or statistical analysis-based models, which are inefficient, costly and subjective, and are difficult to meet the actual demand of large-scale and frequent use. At present, some studies have introduced intelligent technologies such as image recognition and sensor collection to realize automatic and data-driven road condition evaluation. However, these methods still face many challenges in practical application, especially in the aspects of regional generalization ability and multi-scene adaptability of the model. The distribution characteristics of road data in different regions are significantly different, and the model can only be applied to the region where the training data is located, and it is difficult to be extended to other regions.

[0003] In addition, the real road environment is diverse, such as urban roads, highways, and rural roads, which differ in structure, use intensity, and damage type. A single model cannot take into account all kinds of scenes, limiting its application effect in complex environments. At the same time, in some target regions where data is insufficient or unlabeled, the model training faces the problem of "data cold start".

[0004] To solve these problems, transfer learning technology has become an important means to improve the cross-region adaptability of the model. It transfers existing knowledge to realize effective prediction from data-rich regions to data-scarce regions. Combined with the feature extraction capability of deep learning models, it can better capture complex nonlinear relationships and realize unified modeling of multiple road surface scenes. Therefore, it is urgent to build a multi-scene PCI prediction model based on cross-region transfer learning to solve the problems of poor regional adaptability and weak scene generalization of existing methods, improve the prediction accuracy and model practicability, and provide technical support for intelligent road maintenance. SUMMARY

[0005] The application provides a multi-scene road surface PCI prediction method based on cross-region transfer learning to solve the above problems.

[0006] To achieve the above purpose, the application realizes the following technical scheme: The application provides a multi-scene road surface PCI prediction method based on cross-region transfer learning, which includes the following steps: S1. Collect multi-scene road surface images to construct a multi-source road surface image dataset MSRD; pre-process the images in the dataset to obtain road surface image training data , multi-scene fixed images , and road surface image test data ; S2. The multi-scene fixed images are input into a ConvNeXt-Tiny model for visual feature extraction to obtain uncertain road surface features ; S3. The road surface image data is input into a ResNet50 pre-trained model to obtain initial features ; the road surface initial features are input into a road surface image feature extractor to obtain road surface features ; S4. The road surface features and the uncertain road surface features are input into a transfer adaptation module to obtain transfer adaptation features ; the transfer adaptation features and the uncertain road surface features are input into a road surface uncertainty estimation module to obtain uncertainty features ; the road surface features , the uncertainty features , and the transfer adaptation features are input into a sub-domain adaptation branch for processing to obtain sub-domain adaptation output features ; S5. The sub-domain adaptation output features are input into a PCI output layer to obtain road surface PCI prediction results S6. Use a loss function to calculate the error between the road surface PCI prediction results and the true values, and use an Adam optimizer to iteratively optimize the parameters of each module in steps S3 to S5 until the model converges, obtaining a trained model S7. The road surface image test data is input into the trained model to obtain road surface PCI prediction results

[0007] Further, in step S1, the multi-source road surface image dataset MSRD is constructed as follows: According to different geographical positions, a plurality of representative areas are selected, including urban roads, highways and rural roads; each area is divided into a plurality of scenes according to road surface materials, use frequency and climate conditions; the PCI label of each image is collected by a laser road detection system and labeled by combining an artificial checking method; the image is processed by using a Laplace transform variance method to obtain a cleaned image; a multi-source road surface image dataset MSRD is constructed and divided into a training set, a test set and a fixed set.

[0008] Further, the images in the dataset are preprocessed in step S1: The images in the multi-source road surface image dataset MSRD are uniformly scaled by using an image processing library; the images in the training set and the test set are data enhanced to generate road surface image training data and road surface image test data ; the data enhancement includes random rotation, random horizontal flipping and color jittering; M images are uniformly extracted from each type of scene in the fixed set, and the images are processed by using scaling and center cropping to obtain multi-scene fixed images .

[0009] Further, step S3 specifically includes: The ResNet50 pre-training model deletes the fully connected layer in 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 convolution branches, and the convolution branch includes a first branch, a second branch and a third branch; the initial feature is input into the three parallel convolution branches, the outputs of the three branches are spliced in the channel dimension and then subjected to channel fusion by a convolution kernel to obtain channel fusion features ; the initial feature and the channel fusion feature obtain residual enhanced features 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 enhanced feature is compressed in dimension by the global average pooling layer to obtain a compressed feature ; the compressed feature is subjected to nonlinear channel weight learning by the first fully connected layer and the second fully connected layer to obtain a channel weight The channel weight is multiplied with the residual enhanced feature channel by channel to obtain a weighted feature ; In the linear compression module, the weighted features After the linear transformation layer, the road surface features are obtained through LayerNorm normalization and a GELU activation function .

[0010] Further, the migration adaptation module in step S4 specifically includes: The migration adaptation module includes a sub-domain mapping layer, a domain contrast encoder, and a self-supervised feature adjustment layer; the sub-domain 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 splicing layer, a fourth fully connected layer, a SiLU activation function, and a Dropout layer; The road surface features are processed through the sub-domain mapping layer to obtain sub-domain mapping features ; the sub-domain mapping features are input into the main encoder to obtain main encoding features ; the encoding features and the main encoding features are spliced through a Concat function of the feature splicing layer, and then are processed through a fourth fully connected layer, a SiLU activation function, and a Dropout layer to obtain migration adaptation features .

[0011] Further, the road surface uncertainty estimation module in step S4 specifically includes: The road surface uncertainty estimation module is realized based on a weighted correlation estimation method, and the formula is as follows: , wherein, represents a GELU activation function; represents a weight matrix; represents element-wise multiplication; represents a feature splicing operation; b represents a bias vector; is K uncertainty road surface reference features extracted under the scene corresponding to the current prediction target; represents an uncertainty attention aggregation function, and the formula is as follows: , wherein, represents the attention weight of the kth reference feature; sim represents a cosine similarity function; represents the kth uncertainty road surface reference feature extracted under the scene corresponding to the current prediction target; represents the jth uncertainty road reference feature extracted under the current prediction target corresponding scene; represents an exponential function with a natural constant e as the base.

[0012] Further, the sub-domain adaptation branch in step S4 specifically includes: In the sub-domain adaptation branch, the uncertainty feature and the migration adaptation feature are added to obtain an added feature ; the added feature and the road feature are spliced through a Concat function to obtain a spliced feature ; the spliced feature is linearly transformed to reduce the dimension, and then passes through a normalization layer and a ReLU activation function to obtain .

[0013] Further, the PCI output layer in step S5 adopts a linear regression unit, and uses a full connection layer to reduce the dimension of to 1.

[0014] Further, the loss function in step S6 uses a mean square error (MSE), and the formula is as follows: wherein, represents the true PCI value, represents the predicted value of the model, and N represents the number of training data.

[0015] Further, a self-supervised contrast learning mechanism is introduced in the domain contrast encoder, and the uncertainty road feature passes through a shadow encoder to obtain a shadow encoding feature , and the main encoding feature and the shadow encoding feature are used to calculate an InfoNCE contrast loss.

[0016] The present application has the following advantages: The present application effectively improves the diversity and representativeness of the data by collecting road surface images in multiple scenes and constructing a multi-source road surface image dataset, and enhances the generalization ability of the model in different road surface environments. This enables the model to adapt to different road conditions in different regions and different road types, improving the prediction accuracy; by using existing basic models and ResNet pre-training models, the method reduces the dependence on a large amount of labeled data through transfer learning, especially in cross-regional environments, where it is often difficult to obtain labeled data. Freezing part of the structure and using the pre-training model as an initial feature extractor helps to improve the training efficiency of the model and reduce the training time; uncertainty road features are introduced and combined with other features through a transfer adaptation module, so that the uncertainty of the road condition can be accurately estimated. This is of great significance for high-risk areas or scenes lacking labeled data, as it effectively handles the uncertainty of the prediction results and improves the credibility of the model in practical applications; the sub-domain adaptation branch combines the transfer adaptation features with the road features, further optimizing the model 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 between different regions; this method is not only suitable for traditional road PCI prediction scenarios, but also can effectively perform transfer learning in new scenarios, and is particularly suitable for use in cross-regional, cross-environment or new environments, solving the problem that traditional models are difficult to adapt to new scenarios, and has strong versatility and flexibility. BRIEF DESCRIPTION OF DRAWINGS

[0017] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application, and are used to explain the present application, and do not constitute a limitation on the present application.

[0018] Figure 1 A flowchart of the steps of the method of the present application. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be described in detail below with reference to the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0020] Embodiment 1 In this embodiment, as shown in the following table, the present application provides a multi-scene road PCI prediction method based on cross-regional transfer learning, and the specific steps include: Figure 1 ​S1. Collect multi-scene road surface images to construct a multi-source road surface image dataset MSRD; preprocess the images in the dataset to obtain road surface image training data , multi-scene fixed images , and road surface image test data .

[0021] Specifically, a plurality of representative areas are selected according to different geographical locations, including urban roads, highways, and rural roads; each area is divided into a plurality of scenes according to road surface materials, usage frequency, and weather conditions, and the scenes include six road surface states of clean road, crack, pothole, patch, wet, and water; a total of 18 specific scenes are covered; the collection equipment uniformly adopts a high-definition camera device, which is mounted on the middle axis position of the front part of the collection vehicle and is fixed by a shockproof mounting bracket; the PCI label of each image is collected by a laser road detection system and labeled by combining manual verification; the PCI scoring standard follows the “Highway Asphalt Pavement Maintenance Technical Specification” (JTG H40-2004); images with a clarity lower than a threshold value (variance value <100), overexposure (average brightness value >240), and overdarkness (average brightness value <30) are removed using the Laplace transform variance method to obtain cleaned images; finally, no less than 500 valid images are retained for each scene, and all images are saved in a structured image dataset with a naming method of <region number>_<scene number>_<timestamp>.jpg; the cleaned images are organized according to regions and scenes to construct a multi-source road surface image dataset MSRD, and are divided into a training set, a test set, and a fixed set according to a ratio of 8:1:1.

[0022] S2. Multi-scene fixed images are input into a ConvNeXt-Tiny model for visual feature extraction to obtain uncertain road surface features .

[0023] Specifically, the images in the multi-source road surface image dataset MSRD are uniformly scaled to 224x224 pixels using an image processing library; the images in the training set and the test set are data-augmented to generate road surface image training data and road surface image test data ; data augmentation includes random rotation, random horizontal flipping, and color jittering; M images are uniformly extracted from each type of scene in the fixed set, and the images are processed using scaling and center cropping to obtain multi-scene fixed images .

[0024] S3. The road surface image data is input into a ResNet50 pre-trained model to obtain initial features ; the road surface initial features The input is input into the road surface image feature extractor to obtain road surface features .

[0025] Specifically, the ResNet50 pre-training model deletes the fully connected layer in the original model; Delete the fully connected layer (i.e. "fc layer") in the original model, and only keep 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 (contains 3 bottleneck modules); freeze the second stage residual block layer2 (contains 4 bottleneck modules); only keep the third stage (layer3) and the fourth stage (layer4) in a trainable state, and only update the gradients of these two parts during training to enhance the adaptability to specific tasks.

[0026] The following is the structure of the modified "ResNet50 pre-training model": Convolutional layer (conv1, frozen): convolutional kernel size 7x7, stride 2, output channel number 64, padding size 3; Batch normalization layer (bn1, frozen); activation function (ReLU); Max pooling layer: pooling window size 3x3, stride 2; First stage residual module (layer1, frozen): contains 3 "bottleneck structure (Bottleneck)" modules, the structure of the bottleneck structure module is as follows: first step: 1x1 convolution (dimension reduction); second step: 3x3 convolution (extract spatial features); third step: 1x1 convolution (dimension increase); The output channel number of the first stage residual module is 512; Second stage residual module (layer2, frozen): contains 4 "bottleneck structure (Bottleneck)" modules, the output channel number of the second stage residual module is 512; Third stage residual module (layer3, trainable): contains 6 "bottleneck structure (Bottleneck)" modules, the output channel number of the third stage residual module is 1024; Fourth stage residual module (layer4, trainable): contains 3 "bottleneck structure (Bottleneck)" modules, the output channel number of the fourth stage residual module is 2048; Global average pooling layer (avgpool).

[0027] The road surface image feature extractor includes a multi-scale residual fusion module, a channel attention enhancement module, and a linear compression module; Specifically, the multi-scale residual fusion module includes three parallel convolution branches, including a first branch, a second branch, and a third branch; the first branch uses a convolution kernel with a size of 1x1, the second branch uses a convolution kernel with a size of 3x3, and the third branch uses a convolution kernel with a size of 5x5; The initial feature The input is input into three parallel convolution branches, the outputs of the three branches are spliced in the channel dimension, and then a 1*1 convolution kernel is used for channel fusion to obtain a channel fusion feature ; 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 input dimension of the first fully connected layer is 2048, the output dimension is 128, and the activation function is a ReLU activation function; the input dimension of the second fully connected layer is 128, the output dimension is 2048, and the activation function is a Sigmoid activation function; the residual enhancement features are compressed to 2048 dimensions through the global average pooling layer to obtain compressed features ; the compressed features are input into the first fully connected layer and the second fully connected layer for nonlinear channel weight learning to obtain channel weights The channel weights are multiplied with the residual enhancement features channel by channel to obtain weighted features ; In the linear compression module, the weighted features are input into a linear transformation layer to reduce the dimension from 2048 to 1024, and then LayerNorm regularization and a GELU activation function are used to obtain road surface features .

[0028] S4. The road surface features and the uncertainty road surface features are input into a transfer adaptation module to obtain transfer adaptation features ; the transfer adaptation features and the uncertainty road surface features are input into a road surface uncertainty estimation module to obtain uncertainty features ; the road surface features , the uncertainty features , and the transfer adaptation features are input into a subfield adaptation branch for processing to obtain subfield adaptation output features .

[0029] Specifically, the migration adaptation module comprises a sub-domain mapping layer, a domain contrast encoder, and a self-supervised feature adjustment layer; the sub-domain mapping layer comprises a linear transformation layer, a normalization layer, and a GELU activation function; the input dimension of the linear transformation layer is 1024, and the output dimension is 512; the domain contrast encoder comprises a main encoder and a shadow encoder; the main encoder comprises a third fully connected layer, a normalization layer, and a ReLU activation function; the input dimension of the third fully connected layer is 512, and the output dimension is 256; the shadow encoder has the same structure as the main encoder but does not share parameters; the self-supervised feature adjustment layer comprises a feature splicing layer, a fourth fully connected layer, a SiLU activation function, and a Dropout layer; the input dimension of the fourth fully connected layer is 768, and the output dimension is 512; the road surface feature is processed through the sub-domain mapping layer to obtain a sub-domain mapping feature ; the sub-domain mapping feature is input into the main encoder to obtain a main encoding feature ; the encoding feature and the main encoding feature are spliced through a Concat function of the feature splicing layer, and then pass through the fourth fully connected layer, the SiLU activation function, and the Dropout layer to obtain a migration adaptation feature .

[0030] A self-supervised contrast learning mechanism is introduced in the domain contrast encoder; the purpose is to enhance the migration adaptation ability of the model to cross-region road surface images. In this mechanism, first, the road surface feature extracted from the road surface image is processed through the sub-domain mapping layer to obtain a sub-domain mapping feature , and is input into the main encoder to generate a main encoding feature ; at the same time, the uncertain road surface feature is input into the shadow encoder to generate a shadow encoding feature . The main encoding feature and the shadow encoding feature constitute a contrast sample pair; In order to improve the discrimination ability and robustness of the feature extracted by the main encoder, an InfoNCE contrast loss function is used for training. This loss function narrows the distance between the positive sample features of the main encoder and the sub-domain mapping, and widens the distance between the negative samples generated by the shadow encoder, so that the model learns to recognize stable PCI features in the presence of uncertainty and domain shift; Finally, only the output of the main encoder is reserved as the main basis for self-supervised feature adjustment after training. The encoded features and shadow encoded features are concatenated through the Concat function, and then processed through the fourth fully connected layer, SiLU activation function and Dropout regularization to obtain more adaptive transfer features. This structure effectively improves the prediction accuracy and stability of the model in the cross-domain environment; Because the goal of contrastive learning is to train the main encoder to learn a robust representation; the shadow encoder only exists as a "contrast reference during training"; so after training, the main encoder has good generalization ability, and the shadow encoder is no longer needed; The formula of InfoNCE loss function is as follows: , represents the feature extracted by the main encoder; represents the positive sample (such as the same image after sub-domain mapping); represents the negative sample (the feature extracted by the shadow encoder); sim represents the similarity measure (using cosine similarity); τ=0.29 is the temperature coefficient.

[0031] Specifically, the road uncertainty estimation module is realized based on a weighted correlation estimation method, and the formula is as follows: , wherein, represents the GELU activation function; represents the weight matrix; represents element-wise multiplication; represents the feature concatenation operation; b represents the bias vector; is the K uncertainty road features extracted in the scene corresponding to the current prediction target; represents the uncertainty attention aggregation function, and the formula is as follows: , wherein, represents the attention weight of the kth reference feature; sim represents the cosine similarity function; represents the kth uncertainty road reference feature extracted in the scene corresponding to the current prediction target; represents the jth uncertainty road reference feature extracted in the scene corresponding to the current prediction target; represents the exponential function with natural constant e as the base.

[0032] Specifically, in the sub-domain adaptation branch, the uncertainty feature and the transfer adaptation feature are added element by element to obtain the added feature ; add the features and road features Concatenate by Concat function to obtain concatenated features ; add the concatenated features Linear transformation, dimension reduction to 256, through the normalization layer and the ReLU activation function to obtain .

[0033] S5. The sub-domain adaptive output features are input into the PCI output layer to obtain the road PCI prediction result; Specifically, the PCI output layer adopts a linear regression unit, and uses a full connection layer to reduce the dimension of to 1.

[0034] S6. The error between the road PCI prediction result and the true value is calculated by using a loss function, and the parameters of each module in steps S3 to S5 are iteratively optimized by using an Adam optimizer until the model converges, thereby obtaining a trained model; Specifically, the loss function uses mean square error (MSE), which is expressed by the following formula: , wherein, is the true PCI value, is the predicted value of the model, and N is the number of training data.

[0035] S7. The road image test data are input into the trained model to obtain the road PCI prediction result.

[0036] Embodiment 2 In order to more comprehensively demonstrate the effectiveness of the method of the present application, this embodiment compares the present application with existing methods in combination with specific application scenarios. The following will prove the advantages of the method of the present application through several specific application scenarios and in combination with experimental results.

[0037] Road PCI prediction in different geographical areas (city, suburb and mountainous area). Traditional road PCI prediction models usually rely on single-scene data training and cannot effectively cope with the differences between different areas. To solve this problem, the method of the present application can significantly improve the cross-region adaptability of the model through multi-source road image data set (MSRD) and cross-region transfer learning.

[0038] Scenario comparison: Urban road: the road flatness is good, and the road type is single.

[0039] Mountainous road: the road condition is complex, and the road has various damage types such as potholes and cracks.

[0040] Suburban road: The road is often affected by weather and traffic flow changes, and the road surface damage is complex.

[0041] The method of the present application is compared with the existing single-scene prediction method. In order to make a fair comparison, the present application will test the performance of the two methods in road surface PCI prediction under the same scene.

[0042] Experimental setup: Use road image data in urban, mountainous and suburban areas (road image in different scenes). Each scene contains 500 road images.

[0043] Model: Existing method: Only use single scene to train the model and make prediction in other scenes; In the "existing method" compared in the present application, a representative and publicly published convolutional neural network (CNN) based method is selected as the benchmark method. The model structure is relatively standard, only trained on a single scene dataset, without introducing transfer learning or uncertainty modeling, belonging to the typical "single scene supervised learning" framework.

[0044] Method of the present application: Use multi-source road image dataset (MSRD) for training and use cross-region transfer learning for model adaptation; Performance indicators are as follows: Accuracy: the proportion of correctly predicted samples to 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 prediction accuracy of the model; Adaptability score: evaluate the effect of model transfer between different scenes, i.e. the performance of the model from urban scene to mountainous or suburban area. The experimental data of the method of the present application and the existing method are shown in Table 1.

[0045] Table 1 Experimental data of the method of the present application and the existing method The table shows the accuracy comparison of the existing method and the method of the present application in different road surface scenes (urban, mountainous and suburban). It can be seen that the method of the present application has achieved high accuracy in each scene, especially in mountainous and suburban scenes, the accuracy is significantly higher than that of the existing method. The table also shows the RMSE comparison of the existing method and the method of the present application in different scenes, reflecting the size of the prediction error. Through comparison, it can be found that the RMSE value of the method of the present application is significantly lower than that of the existing method, indicating that it is more accurate. Adaptability score evaluates the effect of model transfer between different scenes. The score is from 1 to 10, 1 is poor adaptability, and 10 is strong adaptability. It can be seen that the method of the present application is also significantly better than the existing method in adaptability, especially in mountainous and suburban areas.

[0046] Finally, it should be noted that the above only describes the preferred embodiments of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art will appreciate that the technical solutions described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalent ones. 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 multi-scene road surface PCI prediction method based on cross-region transfer learning, characterized in that, The method comprises the following steps: S1. Collect multi-scene road surface images, and construct a multi-source road surface image dataset MSRD; pre-process the images in the dataset to obtain road surface image training data , multi-scene fixed images , and road surface image test data ; S2. Multi-scene fixed image Input into the ConvNeXt-Tiny model for visual feature extraction to obtain uncertain road features ; S3. input the road surface image data into a ResNet50 pre-training model to obtain initial features ; input the road surface initial features into a road surface image feature extractor to obtain road surface features ; input the road surface initial features into a road surface image feature extractor to obtain road surface features ; input the road surface initial features into a road surface image feature extractor to obtain road surface features ; input the road surface initial features into a road surface image feature extractor to obtain road surface features S4. inputting the road surface feature and the uncertainty road surface feature into a migration adaptation module to obtain a migration adaptation feature ; inputting the migration adaptation feature and the uncertainty road surface feature into a road surface uncertainty estimation module to obtain an uncertainty feature ; Road surface features , uncertainty features , and migration adaptation features are input into the subfield adaptation branch for processing to obtain subfield adaptation output features ; S5. Adapt the output features of the sub-domain Input into the PCI output layer to obtain the road surface PCI prediction result; S6. Calculate the error between the road surface 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, obtaining a trained model; S7. The road surface image test data is input into the trained model to obtain the road surface PCI prediction result.

2. The multi-scene road surface PCI prediction method based on cross-region transfer learning according to claim 1, characterized in that, In step S1, a multi-source road surface image dataset MSRD is constructed: According to different geographical positions, a plurality of representative areas are selected, including urban roads, highways and rural roads; each area is divided into a plurality of scenes according to road surface material, use frequency and climate conditions; the PCI label of each image is collected by a laser road detection system and labeled by combining manual verification; the image is processed using the Laplace transform variance method to obtain a cleaned image; a multi-source road surface image dataset MSRD is constructed and divided into a training set, a test set and a fixed set.

3. The multi-scene road surface PCI prediction method based on cross-region transfer learning according to claim 2, characterized in that, In step S1, the images in the dataset are preprocessed: The image processing library is used to uniformly scale the images in the multi-source road image dataset MSRD; data augmentation is performed on the images in the training set and the test set to generate road image training data and road image test data ; the data augmentation includes random rotation, random horizontal flipping and color jittering; M images are uniformly extracted from each type of scene in the fixed set, and the scaling and center cropping method is used to process the images to obtain multi-scene fixed images .

4. The multi-scene road surface PCI prediction method based on cross-region transfer learning according to claim 3, characterized in that, Step S3 specifically includes: The ResNet50 pre-training model deletes the fully connected layer in 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 convolution branches, and the convolution branches include a first branch, a second branch, and a third branch The input is input into the three parallel convolution branches, the outputs of the three branches are spliced in the channel dimension, and then channel fusion is performed through a convolution kernel to obtain a channel fusion feature ; the initial feature and the channel fusion feature The residual enhancement feature is obtained through a 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 the linear transformation layer, the road surface features are obtained through LayerNorm regularization and a GELU activation function .

5. The multi-scene road surface PCI prediction method based on cross-region transfer learning according to claim 4, characterized in that, The migration adaptation module in step S4 specifically includes: The migration adaptation module includes a sub-domain mapping layer, a domain contrast encoder and a self-supervised feature adjustment layer; the sub-domain 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 feature The sub-domain mapping feature is obtained by processing through a sub-domain mapping layer ; the sub-domain mapping feature The main coding feature is obtained by inputting a main coder ; the coding feature And the main coding feature After being spliced through the Concat function of the feature splicing layer, the migration adaptation feature is obtained through a fourth full connection layer, a SiLU activation function and a Dropout layer .

6. The multi-scene road surface PCI prediction method based on cross-region transfer learning according to claim 5, characterized in that, The road surface uncertainty estimation module in step S4 specifically includes: The road surface uncertainty estimation module is realized based on a weighted correlation estimation method, and the formula is as follows: ， wherein, represents a GELU activation function; represents a weight matrix; represents an element-wise multiplication; represents a feature concatenation operation; b represents a bias vector; is the K uncertain road reference features extracted under the current prediction target corresponding scene; represents an uncertain attention aggregation function, and the formula is as follows: ， wherein, denotes the attention weight of the kth reference feature; sim denotes a cosine similarity function; denotes the kth uncertainty road reference feature extracted under the scene corresponding to the current prediction target; denotes the jth uncertainty road reference feature extracted under the scene corresponding to the current prediction target; denotes an exponential function with a natural constant e as the base.

7. The multi-scene road surface PCI prediction method based on cross-region transfer learning according to claim 6, characterized in that, The sub-domain adaptation branch in step S4 specifically includes: In the sub-domain adaptation branch, the uncertainty feature and the migration adaptation feature After element-level addition, the added feature is obtained ; the added feature and the road surface feature are spliced by the Concat function to obtain the spliced feature ; the spliced feature is linearly transformed and dimensionally reduced, and then passes through the normalization layer and the ReLU activation function to obtain .

8. The multi-scene road surface PCI prediction method based on cross-region transfer learning according to claim 7, characterized in that, The PCI output layer in step S5 employs a linear regression unit using a fully connected layer to reduce the dimensionality of the input to 1.

1.

9. The multi-scene road surface PCI prediction method based on cross-region transfer learning according to claim 8, characterized in that, In step S6, the loss function uses mean square error MSE, and the formula is as follows: wherein, represents the true PCI value, represents the predicted value of the model, N represents the number of training data.

10. The multi-scene road surface PCI prediction method based on cross-region transfer learning according to claim 9, characterized in that, Introducing self-supervised contrastive learning mechanism in domain contrastive encoder, uncertain road features Get shadow coding features through shadow encoder , calculate InfoNCE contrastive loss through main coding features and shadow coding features .

Images