Plateau mountainous area highway weather classification method and system based on image partition feature dynamic fusion

Abstract

The invention relates to a plateau mountainous area highway weather classification method and system based on image partition feature dynamic fusion, and belongs to the technical field of traffic weather. The method comprises the following steps of: performing road surface region identification on an acquired road image with a road surface region to generate a corresponding road surface region mask graph; dividing the road image into a plurality of sub-images; calculating the proportion of the pavement area in each sub-graph according to the pavement area mask graph; weather feature extraction is carried out on each sub-graph, and probability distribution of various meteorological conditions of each sub-graph is calculated; generating corresponding weight parameters based on the probability distribution of various meteorological conditions and the road surface proportion information of each sub-graph; and carrying out weighted fusion on the probability distribution of various meteorological conditions of each sub-graph by using the weight parameters to obtain a final meteorological condition classification result of the whole image. According to the invention, the key meteorological factors which influence the road traffic safety in different scenes can be effectively identified, so that the traffic safety guarantee capability is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of traffic meteorology, and particularly relates to a highland mountainous area highway meteorological classification method and system based on image partition feature dynamic fusion. BACKGROUND

[0002] The meteorological conditions of highland mountainous area highways change rapidly, and extreme weather such as heavy fog, heavy rain and ice and snow occurs frequently, which seriously threatens traffic safety. The existing meteorological condition recognition technology usually adopts a convolutional neural network (CNN) method based on a global image, which judges the weather category represented by the image by performing feature extraction and classification training on the whole image. This kind of method usually relies on large-scale image data sets for training, does not distinguish between road regions and background regions in the image, and directly analyzes and predicts the overall features of the whole image. For example, a weather image classification model based on the Xception architecture (Naufal, M. F., & Kusuma, S. F. (2022). Weather image classification using convolutional neural network with transfer learning. AIP Conference Proceedings, 2470 (1), 050004.) classifies weather images using a CNN model based on the Xception architecture; and a MASK-CNN-Transformer (MASK-CT) model (Chen, S., Shu, T., Zhao, H., & Tang, Y. Y. (2023). MASK-CNN-Transformer for real-time multi-label weather recognition. Knowledge-Based Systems, 269, 110455.) combines a convolutional neural network (CNN) and a Transformer architecture, and proposes a model for real-time multi-label weather recognition.

[0003] However, these methods have obvious limitations. In plateau mountainous area images, distant mountains may appear weather phenomena unrelated to road traffic (such as heavy fog or snow on the mountain), and roadside water bodies may also cause misjudgment, while these areas have limited impact on actual traffic safety. Although there may be weather anomalies such as snow on distant mountains or heavy fog, they are usually far from the road and have no direct relationship with the actual driving surface. If the entire feature is involved in identification, it is easy to cause model misjudgment, and the output weather category does not match the actual traffic environment. For example, when there is heavy fog on the distant mountains of the image, but the road surface is clear and unobstructed, the traditional model may still misjudge the entire image as "heavy fog weather", thus producing false alarms. The traditional model fails to effectively distinguish the different contributions of road surface and non-road surface areas to the final classification result, resulting in insufficient recognition accuracy in complex environments and difficulty in providing reliable meteorological information support for traffic safety.

[0004] In addition, some existing research attempts to improve the model's attention to key areas by introducing multi-scale feature extraction or attention mechanisms, but still fails to explicitly guide the model to distinguish the importance difference between "road surface area" and "non-road surface area" in traffic safety judgment. In addition, most global image classification models lack explicit modeling of spatial structure and cannot adapt to the combination of complex terrain and meteorological background in plateau mountainous areas, resulting in weak generalization ability in actual application.

[0005] Therefore, how to overcome the shortcomings of existing meteorological condition recognition technology, which fails to effectively distinguish the different influences of road surface area and non-road surface area on traffic safety, and has insufficient recognition accuracy in complex environments, and cannot adapt to the combination of special terrain and weather in plateau mountainous areas, and has weak generalization ability, is a problem that needs to be solved in the technical field at present. SUMMARY

[0006] The purpose of the present application is to overcome the shortcomings of the prior art and provide a plateau mountainous area highway meteorological classification method and system based on dynamic fusion of image partition features.

[0007] To achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows:

[0008] A plateau mountainous area highway meteorological classification method based on dynamic fusion of image partition features, comprising the following steps:

[0009] Step (1), performing road surface area recognition on the collected highway image with road surface to generate a corresponding road surface area mask image;

[0010] Step (2), dividing the highway image into multiple sub-images, and calculating the proportion of road surface area in each sub-image according to the road surface area mask image;

[0011] Step (3), weather feature extraction is performed on each subgraph, and the probability distribution of each type of meteorological condition of each subgraph is calculated;

[0012] Step (4), based on the probability distribution of each type of meteorological condition of each subgraph and the road surface ratio information, the corresponding weight parameter is generated;

[0013] Step (5), the weight parameter obtained in step (4) is used to weight and fuse the probability distribution of each type of meteorological condition of each subgraph, and the final meteorological condition classification result of the whole image is obtained.

[0014] The method disclosed by the application is characterized in that a unified intelligent classification model capable of end-to-end training is constructed. The model is divided into a "pre-training module" and a "to-be-trained module" in structure: the "pre-training module" includes a road surface area recognition model for performing road surface mask generation and functions such as partition and ratio calculation, and the parameters thereof remain fixed (i.e., parameter freezing) during model training; and the "to-be-trained module" integrates the core innovation of the application, i.e., an improved visual encoder model for weather feature extraction and a dynamic weight calculation module, and the parameters thereof are jointly learned and updated according to the final classification loss during the training process. This end-to-end training mode with partially fixed parameters can not only improve the processing efficiency by using existing mature models, but also ensure that the core feature extraction and dynamic fusion mechanism work cooperatively, so that the model can efficiently learn and realize accurate and robust classification of the meteorological conditions of highland mountain roads.

[0015] Further, preferably, in step (1), the road surface area of the collected road image with a road surface is recognized to generate a corresponding mask image. The specific method is as follows: the collected road image with a road surface is recognized to recognize the road surface area and the non-road surface area, then the pixel value of the road surface area is set to 1, and the pixel value of the non-road surface area is set to 0, so as to obtain the mask image.

[0016] Further, preferably, for the road image F with a road surface, the road mask m is represented as:

[0017] m=RoadSegModel(F)

[0018] Wherein, RoadSegModel is a road surface area recognition model, and m is a generated binary mask image, satisfying:

[0019]

[0020] m(x,y) represents the pixel value at the coordinate (x,y) on the mask image.

[0021] Further, preferably, the specific method of step (2) is as follows:

[0022] Divide the highway image into N sub-images, denoted as f1 to f N ; For each sub-image, calculate its intersection with the road mask image m to obtain the proportion r of the road area in the sub-image i , i = 1 to N; the calculation formula is:

[0023]

[0024] Where m(x,y) represents the pixel value at the coordinate (x,y) on the mask image, 1 represents road surface, and 0 represents non-road surface; f i is the i-th subgraph.

[0025] Furthermore, preferably, the specific method of step (3) is: using an improved visual encoder model to i The weather characteristics are extracted to obtain the weather feature vector e i , and then calculate the probability distribution y of the meteorological conditions of each subgraph based on these features i , i=1~N;

[0026] y i =[p i,1 ,p i,2 ,……,p i,K ], where p i,j represents the probability that the i-th subgraph corresponds to the j-th type of weather condition, j = 1 to K; K is the total number of weather categories;

[0027] y i The calculation method is as follows: the weather characteristic vector e i Map it to the weather category through the fully connected layer to obtain the unnormalized category score Z i , and then normalized by the Softmax function to obtain the meteorological condition probability distribution y i ;

[0028] Z i =w T e i +b

[0029] y i =Softmax(Z i )

[0030] Where w is the initialized learning weight matrix, T represents the transpose operation, b is the initialization bias vector, and Softmax represents the normalized exponential function;

[0031] The improved visual encoder model includes a visual Transformer module, a dynamic feature fusion module, and an output module; the visual Transformer module, the dynamic feature fusion module, and the output module are connected sequentially;

[0032] The dynamic feature fusion module comprises a spatial branch and a channel branch; the spatial branch captures local spatial details through a 3*3 depth separable convolution; the channel branch compresses channel dimensions through sequentially connected 1*1 convolution and SE channel attention mechanism, extracts global semantic information, and the number of compressed channels is C / r=16, C is the number of channels before compression, and r is the compression ratio; the features extracted by the two branches are fused in parallel and output to obtain a weather feature vector.

[0033] Further, preferably, K=18, the weather categories include non-disaster weather, light rain, moderate rain, heavy rain, heavy rain, lightning, light fog, moderate fog, heavy fog, thick fog, light snow, moderate snow, heavy snow, blizzard, hail, sandstorm, halo, and icy road surface; the compression ratio r=16.

[0034] Further, preferably, in step (4), the specific method for generating the weight parameter is as follows:

[0035] (a) For each subgraph i, the proportion of the road area r i and the weather condition probability distribution y i are spliced to form an input feature vector x i ;

[0036] (b) The input feature vector x i is input into a fully connected neural network, and the weight score s i of the subgraph is calculated through the network;

[0037] (c) The weight scores s1 to s N of all N subgraphs are normalized by applying the Softmax function to obtain the final weight parameter of each subgraph, wherein the final weight parameter of the subgraph i is w i .

[0038] Further, preferably, the specific method of step (5) is as follows:

[0039] The weight parameters obtained in step (4) are used to weight and fuse the weather condition probability distributions of the subgraphs, which are calculated by the following formula:

[0040]

[0041] Wherein y is a K-dimensional probability vector, y=[p1,p2,...,p K ], p j represents the probability of the whole image corresponding to the jth weather condition, and K is the total number of weather categories;

[0042] Then y is taken as the final mountainous road weather condition classification result of the whole image.

[0043] The application also provides a highland mountainous area highway weather classification system based on dynamic fusion of image partition features, which adopts the highland mountainous area highway weather classification method based on dynamic fusion of image partition features and comprises the following steps.

[0044] A mask generation module is configured to identify a road surface area in a collected highway image with a road surface and generate a corresponding road surface area mask image.

[0045] A partition and proportion calculation module is connected to the mask generation module and configured to divide the highway image into multiple sub-images and calculate the proportion of the road surface area in each sub-image according to the road surface area mask image.

[0046] A weather feature extraction module is connected to the partition and proportion calculation module and configured to extract weather features from each sub-image and calculate the probability distribution of each type of meteorological condition of each sub-image.

[0047] A weight calculation module is connected to the partition and proportion calculation module and the weather feature extraction module and configured to generate corresponding weight parameters based on the probability distribution of each type of meteorological condition and the road surface proportion information of each sub-image.

[0048] A weighted fusion module is connected to the weight calculation module and the weather feature extraction module and configured to obtain the probability distribution of each type of meteorological condition of each sub-image by weighting the weight parameters obtained by the weight calculation module and obtain the final meteorological condition classification result of the entire image.

[0049] In the application, the RoadSegModel is a road surface area identification model, which serves as a general module, and any artificial intelligence model with the function of effectively identifying the road surface area in the input highway image can be applied to the method framework, for example, the RoadSegNet model based on the DeepLabV3+ architecture (Pal, K., Yadav, P., & Katal, N. (2022). RoadSegNet: A deep learning framework for autonomous urban road detection. Journal of Engineering and Applied Science, 69(1), 1-21.), the road surface segmentation model based on the U-Net architecture (Spectrewolf8. (2024). Aerial Image Road Segmentation with U-Net-50 Architecture. Hugging Face.), and the road surface defect identification model based on the Mask R-CNN (Kulambayev, B., Kadyrov, T., Kydyrbek, A., et al.

[0050] (2022). A Deep Learning-Based Approach for Road Surface Damage Detection. Computers, Materials & Continua, 73(2), 2715-2731.) These models can all provide effective road surface recognition results, generate mask maps suitable for subsequent processing of the method, thus embodying the universality and flexibility of the module in the overall framework.

[0051] In the present application, the visual Transformer module is an existing model, and the present application is improved based on this basis, and a dynamic feature fusion module and an output module are added, thereby obtaining an improved visual encoder model.

[0052] In the improved visual encoder model, a dynamic feature fusion module is used to realize multi-granularity feature extraction. The dynamic feature fusion module includes a spatial branch and a channel branch. The spatial branch captures local spatial details through a 3x3 depth separable convolution, with the output channel number C=256 and the stride=2, realizing down-sampling while retaining key spatial structures; the channel branch compresses the channel dimension (compression ratio r=16) through a 1x1 convolution+SE channel attention mechanism to extract global semantic information, and the compressed channel number is C / r=16, reducing redundant information. The features extracted by the two branches in parallel are fused to obtain a weather feature vector.

[0053] In the present application, the calculation module for generating dynamic weight parameters is designed in accordance with the principles of lightweight and high efficiency, and is trained as part of the overall end-to-end model. The specific workflow of the module is as follows:

[0054] First, for each subgraph, the calculated road surface area ratio r i (a scalar) is concatenated with the meteorological condition probability distribution y i (a K-dimensional vector) to form a unified input feature vector x i that can represent both road surface information and weather judgment.

[0055] Next, the input feature vector x i is fed into a lightweight fully connected neural network (FCNN). The network performs a nonlinear transformation on the input features through its internal learnable parameters, thereby learning and determining the importance of the current subgraph to the final classification result, and outputting a weight score s i .

[0056] Finally, a Softmax normalization layer is used to process the scores s of all subgraphs i , ensuring that the sum of all weight parameters is 1.0, forming a valid probability distribution. The resulting dynamic weight parameters w i can adaptively fuse the weather features of each subgraph, thereby improving the accuracy and reliability of the final classification.

[0057] In the training of the fully connected neural network, the road surface area ratio r i and the weather condition probability distribution y i are used as inputs, and the weight score s i is used as output, requiring the loss (binary cross-entropy loss) of the weighted fusion of the weather condition probability distribution of each subgraph and the true label to be minimized; in practical applications, the trained fully connected neural network is directly used.

[0058] Classification loss

[0059] In the present application, y is a K-dimensional probability vector, and the classification result is not just a category, but a probability corresponding to each weather category. For example, if the weather category list is J = [heavy rain, heavy fog, heavy snow, …], and y = [0.95, 0.85, 0.7, …], then the classification result y represents the probability of heavy rain as 95%, the probability of heavy fog as 85%, and the probability of heavy snow as 70% (i.e., multiple weather phenomena occurring simultaneously).

[0060] The present application proposes a high-altitude mountain road weather classification method based on image partition feature dynamic fusion, which significantly improves the accuracy and robustness of weather condition classification by introducing road mask guided image multi-subgraph feature extraction and dynamic weight fusion mechanism. The core innovation lies in the model's ability to learn a set of dynamic weight parameters, which adaptively adjust the weather condition weights of each subgraph based on the weather feature vector e i , the road surface area ratio r i , and the weather condition probability distribution y i , and weightedly fuse the weather condition probability distributions of each subgraph to obtain the final weather condition classification result of the entire image. This method not only effectively distinguishes the importance of road surface area and non-road surface area, but also automatically adjusts the focus according to different scenarios, thereby providing more reliable and accurate weather condition recognition results under complex and variable high-altitude mountain weather conditions, and providing strong technical support for traffic safety management.

[0061] The present application utilizes the image partition feature dynamic fusion mechanism to provide more accurate and reliable weather condition recognition under complex weather conditions, especially in high-altitude mountain areas. The present application can effectively identify key meteorological factors that affect highway traffic safety in different scenarios, thereby improving traffic safety and security capabilities.

[0062] Compared with the prior art, the present application has the following beneficial effects:

[0063] 1. Improved classification accuracy and adaptive fusion

[0064] The present application can significantly improve the classification accuracy of weather conditions by extracting and processing multiple subgraph features within the model and fusing them through learned dynamic weights. The model can automatically learn the importance of different subgraphs to the final classification result, focusing on the area features that truly affect traffic, while reducing the interference that may be caused by background areas such as distant snow-capped mountains and roadside water bodies. This adaptive fusion method can effectively avoid the judgment bias that may occur in traditional full-image recognition methods, ensuring accurate identification of key meteorological factors affecting traffic safety in special environments such as high-altitude mountainous areas.

[0065] 2. Enhanced robustness

[0066] Due to the use of the model's internal multi-subgraph feature processing and dynamic weight fusion strategy, the present application can effectively deal with local interference or abnormal areas that may occur in the image. Even if the feature representation of some subgraphs is inaccurate (such as fog on a distant mountain or reflection of roadside water bodies), the model can automatically reduce the impact of these areas through the learned weight mechanism, thereby enhancing the robustness of the overall recognition. This mechanism makes the model have stronger adaptability to various complex weather conditions and road conditions, especially suitable for complex environments such as high-altitude mountainous areas with variable weather.

[0067] 3. Wide range of applications

[0068] The dynamic weight fusion mechanism of the present application has strong generalization ability and can adapt to different types of highway scenes and various weather conditions. Since the weight parameters are learned by the model rather than manually set, the model can automatically adapt to the characteristics of different scenes and correctly identify the meteorological conditions that truly affect traffic safety. This adaptability makes the present application particularly suitable for special environments such as high-altitude mountainous areas with variable climate and complex road conditions, providing more reliable technical support for traffic safety management.

[0069] 4. Technical comparison

[0070] Through comparative tests between the method of the present application and the existing advanced technology in this field, quantitative analysis shows that the method of the present application has significant key meteorological factor identification ability in complex scenes such as high-altitude mountainous areas, and can provide more reliable technical support for traffic meteorological warning systems.

[0071] The following is the experimental performance (recognition accuracy) of various methods on the high-quality highland mountain meteorological disaster scene image dataset built by the present application. The results are shown in Table 1. As can be seen from the table data, although the general models such as ConvNeXt, MaxViT and InternImage have strong performance, the method proposed by the present application can more accurately focus on the key road surface area affecting traffic safety by virtue of its targeted "image partition" and "dynamic fusion" mechanism, effectively excluding the interference of the background environment, and thus in various complex meteorological conditions specific to highlands and mountains, the recognition accuracy has achieved stable and comprehensive surpassing, especially in "fog", "snow" and "ice and snow road" and other categories that are highly dependent on road conditions, the advantage is particularly significant. This fully proves the advancement and practical value of the present application in the specialized application field of traffic meteorology.

[0072] Table 1

[0073] Weather category ConvNeXt MaxViT InternImage Invention No disaster weather 97.20％ 97.50％ 97.80％ 98.20％ Light rain 77.50％ 78.80％ 79.50％ 82.50％ Moderate rain 80.20％ 81.50％ 82.10％ 85.80％ Heavy rain 84.10％ 84.70％ 85.50％ 89.10％ Violent rain 86.00％ 86.90％ 87.60％ 91.30％ Lightning 90.10％ 90.80％ 91.50％ 93.70％ Light fog 83.50％ 84.20％ 85.40％ 90.60％ Medium fog 86.80％ 87.50％ 88.10％ 92.80％ Heavy fog 89.90％ 90.60％ 91.20％ 95.30％ Thick fog 91.50％ 92.10％ 92.80％ 96.50％ Light snow 72.80％ 74.20％ 75.50％ 81.70％ Moderate snow 76.50％ 77.60％ 78.30％ 84.60％ Heavy snow 79.70％ 80.50％ 81.80％ 87.90％ Violent snow 84.00％ 84.90％ 85.60％ 90.20％ Hail 78.20％ 79.10％ 80.30％ 85.10％ Sandstorm 92.00％ 92.50％ 93.20％ 95.90％ Halo 94.30％ 94.80％ 95.50％ 97.20％ Ice and snow road 81.60％ 82.70％ 83.40％ 89.50％

[0074] Note: The three methods of ConvNeXt, MaxViT and InternImage are advanced models in the field of image recognition in recent years. ConvNeXt was proposed by Zhuang Liu et al. in 2022, which modernizes the convolutional network and has similar performance to Transformer; MaxViT was proposed by Zhengzhong Tu et al. in 2022, which builds an efficient hybrid vision model by combining local convolution and global attention mechanism; InternImage was proposed by Wenhai Wang et al. in 2023, which explores large-scale visual basic model based on deformable convolution and performs well in multiple tasks. BRIEF DESCRIPTION OF DRAWINGS

[0075] Figure 1 Flowchart of the image partition feature dynamic fusion highland mountain highway meteorological classification method of the present application;

[0076] Figure 2 Whole highway image in the example of the present application;

[0077] Figure 3 Highway image road surface area mask graph in the example of the present application;

[0078] Figure 4 Dynamic fusion of image partition features in the example of the present application;

[0079] Figure 5 Partition feature fusion model recognition result in the example of the present application;

[0080] Figure 6 Weather image classification model weather recognition result based on Xception architecture in the example of the present application;

[0081] Figure 7 Mask-CT model weather recognition result in the embodiment of the application;

[0082] Figure 8 Schematic diagram for generating a mask map in the application;

[0083] Figure 9 Flowchart for obtaining the proportion of the road surface area in each subgraph in the application.

[0084] Figure 10 Improved visual encoder model architecture in the application;

[0085] Figure 11 Overall architecture of the end-to-end model in the application. DETAILED DESCRIPTION

[0086] The application will be further described in detail below in conjunction with the embodiments.

[0087] Those skilled in the art will understand that the following embodiments are only used to illustrate the application and should not be regarded as limiting the scope of the application. If the specific technology or condition is not specified in the embodiments, the technology or condition described in the literature in the art or according to the product manual is used. If the manufacturer of the material or equipment is not specified, it is a conventional product that can be obtained by purchase.

[0088] Embodiment 1

[0089] A highland mountainous highway weather classification method based on image partition feature dynamic fusion includes the following steps:

[0090] Step (1), the collected road image with road surface is subjected to road surface area recognition to generate a corresponding road surface area mask map;

[0091] Step (2), the road image is divided into multiple subgraphs, and the proportion of the road surface area in each subgraph is calculated according to the road surface area mask map;

[0092] Step (3), weather feature extraction is performed on each subgraph, and the probability distribution of each meteorological condition of each subgraph is calculated;

[0093] Step (4), based on the probability distribution of each meteorological condition of each subgraph and the road surface proportion information, the corresponding weight parameter is generated;

[0094] Step (5), the weight parameter obtained in step (4) is used to weight and fuse the probability distribution of each meteorological condition of each subgraph to obtain the final meteorological condition classification result of the whole image.

[0095] The application discloses a highland mountain area highway weather classification system based on dynamic fusion of image partition features.

[0096] The mask generation module is used for identifying a road surface area of the collected highway image with a road surface and generating a corresponding road surface area mask image.

[0097] The partition and proportion calculation module is connected with the mask generation module and is used for dividing the highway image into a plurality of sub-images and calculating a road surface area proportion in each sub-image according to the road surface area mask image.

[0098] The weather feature extraction module is connected with the partition and proportion calculation module and is used for extracting weather features of each sub-image and calculating a probability distribution of each weather condition of each sub-image.

[0099] The weight calculation module is connected with the partition and proportion calculation module and the weather feature extraction module respectively and is used for generating corresponding weight parameters based on the probability distribution of each weather condition of each sub-image and the road surface area proportion information.

[0100] The weighted fusion module is connected with the weight calculation module and the weather feature extraction module respectively and is used for obtaining the weight parameters of the weight calculation module and weighted fusing the probability distribution of each weather condition of each sub-image to obtain a final weather condition classification result of the whole image.

[0101] Embodiment 2

[0102] The application discloses a highland mountain area highway weather classification method based on dynamic fusion of image partition features.

[0103] Step (1), a road surface area of the collected highway image with a road surface is identified to generate a corresponding road surface area mask image.

[0104] Step (2), the highway image is divided into a plurality of sub-images, and a road surface area proportion in each sub-image is calculated according to the road surface area mask image.

[0105] Step (3), weather features of each sub-image are extracted, and a probability distribution of each weather condition of each sub-image is calculated.

[0106] Step (4), corresponding weight parameters are generated based on the probability distribution of each weather condition of each sub-image and the road surface area proportion information.

[0107] Step (5), the weight parameters obtained in step (4) are used for weighted fusing the probability distribution of each weather condition of each sub-image to obtain a final weather condition classification result of the whole image.

[0108] In step (1), the collected highway image with a road surface is subjected to road surface area recognition. The specific method for generating the corresponding mask map is as follows: the collected highway image with a road surface is recognized, the road surface area and the non-road surface area are identified, and then the pixel value of the road surface area is set to 1, and the pixel value of the non-road surface area is set to 0, thereby obtaining the mask map.

[0109] For a highway image F with a road surface, the road surface mask m is expressed as:

[0110] m=RoadSegModel(F)

[0111] Among them, RoadSegModel is the road area recognition model, and m is the generated binary mask image, which satisfies:

[0112]

[0113] m(x,y) represents the pixel value at coordinate (x,y) on the mask image.

[0114] The specific method of step (2) is:

[0115] Divide the highway image into N sub-images, denoted as f1 to f N ; For each sub-image, calculate its intersection with the road mask image m to obtain the proportion r of the road area in the sub-image i , i = 1 to N; the calculation formula is:

[0116]

[0117] Where m(x,y) represents the pixel value at the coordinate (x,y) on the mask image, 1 represents road surface, and 0 represents non-road surface; f i is the i-th subgraph.

[0118] The specific method of step (3) is: using the improved visual encoder model, for each sub-image f i The weather characteristics are extracted to obtain the weather feature vector e i , and then calculate the probability distribution y of the meteorological conditions of each subgraph based on these features i , i=1~N;

[0119] y i =[p i,1 ,p i,2 ,……,p i,K ], where p i,j represents the probability that the i-th subgraph corresponds to the j-th type of weather condition, j = 1 to K; K is the total number of weather categories;

[0120] y i The calculation method is as follows: the weather characteristic vector e iThe weather category is mapped to the weather category score Z by the full connection layer i , and the weather condition probability distribution y is obtained by normalizing the score Z by the Softmax function i ;

[0121] Z i = wT x + b T e i +b

[0122] y i = Softmax(Z i )

[0123] wherein w is the initialized learning weight matrix, T represents the transpose operation, b is the initialized bias vector, and Softmax represents the normalization exponential function;

[0124] The improved visual encoder model comprises a visual Transformer module, a dynamic feature fusion module and an output module; the visual Transformer module, the dynamic feature fusion module and the output module are sequentially connected.

[0125] The dynamic feature fusion module comprises a spatial branch and a channel branch; the spatial branch captures local spatial details through a 3*3 depth separable convolution; the channel branch compresses the channel dimension through a 1*1 convolution and an SE channel attention mechanism connected in sequence, extracts global semantic information, and the compressed channel number is C / r=16, wherein C is the channel number before compression, and r is the compression ratio; the features extracted by the two branches in parallel are fused and output, to obtain a weather feature vector.

[0126] K=18, and the weather categories include non-disaster weather, light rain, moderate rain, heavy rain, heavy rain, lightning, light fog, moderate fog, heavy fog, thick fog, light snow, moderate snow, heavy snow, blizzard, hail, sandstorm, halo and icy road surface; the compression ratio r=16.

[0127] In step (4), the specific method for generating the weight parameter is as follows:

[0128] (a) for each subgraph i, the road area proportion r i and the weather condition probability distribution y i are spliced to form an input feature vector x i ;

[0129] (b) the input feature vector x i is input into a full connection neural network, and the weight score s i corresponding to the subgraph is calculated through the network;

[0130] (c) the weight scores s1 to s N of all N subgraphs are combined to obtain the weight parameter w.The Softmax function is applied for normalization to obtain the final weight parameter of each subgraph, wherein the final weight parameter of the subgraph i is w i .

[0131] The specific method of step (5) is:

[0132] The weight parameter obtained in step (4) is used to weight and fuse the weather condition probability distribution of each subgraph, which is calculated by the following formula:

[0133]

[0134] wherein y is a K-dimensional probability vector, y = [p1, p2, …, p K ] and p j represents the probability of the whole image corresponding to the j-th weather condition, and K is the total number of weather categories;

[0135] Then, y is taken as the final weather condition classification result of the whole image of the mountainous road.

[0136] As shown in Figure 11 , a highland mountainous road weather classification system based on dynamic fusion of image partition features adopts the above-mentioned highland mountainous road weather classification method based on dynamic fusion of image partition features, comprising:

[0137] A mask generation module is configured to identify the road surface area of the collected road image with a road surface and generate a corresponding road surface area mask image.

[0138] A partition and proportion calculation module is connected to the mask generation module and is configured to divide the road image into a plurality of subgraphs and calculate the proportion of the road surface area in each subgraph according to the road surface area mask image.

[0139] A weather feature extraction module is connected to the partition and proportion calculation module and is configured to extract weather features of each subgraph and calculate the probability distribution of each weather condition of each subgraph.

[0140] A weight calculation module is connected to the partition and proportion calculation module and the weather feature extraction module and is configured to generate corresponding weight parameters based on the probability distribution of each weather condition of each subgraph and the road surface proportion information.

[0141] A weighted fusion module is connected to the weight calculation module and the weather feature extraction module and is configured to weight and fuse the probability distribution of each weather condition of each subgraph by the weight parameters obtained by the weight calculation module to obtain the final weather condition classification result of the whole image.

[0142] Embodiment 3

[0143] The core of the method disclosed in the present invention is to construct a unified intelligent classification model (i.e., an end-to-end model) that can be trained end-to-end. The construction and workflow of the model can be decomposed into the five steps in the following embodiments. It should be noted that the "road surface area recognition model" described in step (1) can be regarded as an independent pre-training component in the present invention, and its parameters are fixed during model training. The "improved visual encoder model" described in step (3) and the weight generation network described in step (4) together constitute the core module that needs to be learned. When the overall model is optimized end-to-end, the parameters of the "improved visual encoder model" and the parameters of the weight generation network will be learned and updated simultaneously according to the final classification loss, so that each step works together to achieve accurate and robust classification of meteorological conditions on plateau mountain roads.

[0144] Figure 11 The overall architecture of the end-to-end model proposed in the present invention is presented in detail, clearly dividing the "pre-training module" and the "module to be trained". The parameters of the "pre-training module" are fixed (frozen), and it contains the "mask generation module" and the "partitioning and proportion calculation module". These two modules are responsible for performing common image processing steps, including generating a road mask map, dividing the image into sub-images, and calculating the road surface proportion of each sub-image. The parameters of the "module to be trained" are learnable and are the core of the entire model. It is composed of the "weather feature extraction module", "weight calculation module" and "weighted fusion module" in sequence. During training, the output of the pre-training module is sent to the module to be trained, and the parameters of the first two modules inside it are optimized through joint training, so that the model can learn to extract weather features from sub-images, adaptively calculate fusion weights, and finally output accurate classification results. This architectural design ensures the efficiency of the entire training process and the accuracy of the final classification.

[0145] A plateau mountainous highway weather classification method based on dynamic fusion of image partition features includes the following steps:

[0146] Step (1), performing road surface area recognition on the collected highway image with the road surface, and generating a corresponding road surface area mask map;

[0147] Step (2), dividing the highway image into multiple sub-images, and calculating the proportion of the road area in each sub-image based on the road area mask image;

[0148] Step (3), extract weather features from each sub-graph and calculate the probability distribution of various meteorological conditions for each sub-graph;

[0149] Step (4): Generate corresponding weight parameters based on the probability distribution of various meteorological conditions and road surface ratio information of each subgraph;

[0150] Step (5), the weather condition classification result of the whole image is obtained by weighting and fusing the weather condition probability distribution of each subgraph with the weight parameter obtained in step (4).

[0151] The application provides a highland mountainous road weather classification model based on image partition feature dynamic fusion. Unlike the traditional global recognition method, the application first generates a road surface area mask image by a road surface area recognition model RoadSegModel, divides the road image into a plurality of subgraphs, and calculates the proportion of the road surface area in each subgraph according to the road surface area mask image. The improved visual encoder model is used to extract and process the features of the subgraph area, and the weather condition feature representation of each subgraph area is obtained. Finally, a learnable dynamic weight fusion mechanism is used to fuse the feature representations of the subgraphs into the weather condition classification result of the whole image. The fusion weight is not simply determined by the road surface proportion, but is obtained through model learning, and can be adaptively adjusted according to the content features of the subgraphs and the road surface proportion information. Figure 1 The method flowchart of the application is shown.

[0152] 1. Mask

[0153] The road surface mask is a key tool for identifying the road surface area in the image. Each image will generate a corresponding mask image through any pre-trained road surface area recognition model applicable to the framework of the method. The mask image is used to identify the road surface and non-road surface areas in the image, and provides a basis for the subgraph feature extraction and road surface proportion calculation in the model, which helps the model pay more attention to the weather conditions directly related to road traffic. The schematic diagram of generating the mask image is shown in Figure 8

[0154] For the collected road image F with a road surface, the road surface mask m can be represented as:

[0155] m = RoadSegModel (F)

[0156] Wherein, RoadSegModel is a road surface area recognition model, m is a generated binary mask image, and satisfies:

[0157]

[0158] The mask image m not only directly affects the calculation of the road surface proportion of the subgraph, but also indirectly affects the subsequent feature extraction and weight generation process through the road surface proportion, and is an important basis for the whole model.

[0159] 2. Subgraph division and road surface proportion calculation

[0160] The mask image is divided into N subgraph areas, denoted as f1 to f N ​For each sub-image, by calculating its intersection with the road mask, we can get the proportion of the road area in the sub-image r i , i=1~N, recorded as r1 to r N The flow chart is as follows: Figure 9 These road surface proportions are static parameters for the current input image and serve as additional input features of the model.

[0161] Calculating the road surface ratio of each subgraph can be expressed as:

[0162]

[0163] Where m(x,y) represents the pixel value at the coordinate (x,y) on the mask image, 1 represents road surface, and 0 represents non-road surface.

[0164] The flowchart for obtaining the proportion of road area in each sub-graph is as follows Figure 9 shown.

[0165] 3. Multi-region weather feature extraction

[0166] The improved Vision Encoder model is based on the Visual Transformer (ViT), which captures the semantic correlation between different regions in the image through a multi-head self-attention mechanism, and combines a dynamic feature fusion module to achieve multi-granularity feature extraction and enhance the ability to perceive local details. Figure 10 As shown. The dynamic feature fusion module consists of a spatial branch and a channel branch. The spatial branch captures local spatial details through 3×3 depth-separable convolution, with the output channel number C=256 and stride=2, achieving downsampling while retaining the key spatial structure; the channel branch compresses the channel dimension (compression ratio r=16) through 1×1 convolution + SE channel attention mechanism to extract global semantic information. The number of channels after compression is C / r=16, reducing redundant information. The two branches extract features in parallel and then fuse them to obtain the weather feature vector. The model supports multi-query attention mechanism, converts the image plane data (32×32 resolution image block) into a linear token sequence, and predicts the next token through autoregression to obtain each sub-image f i Unified embedding representation of weather features - weather feature vector e i , and then calculate the probability distribution y of the meteorological conditions of each subgraph based on these features i , i=1~N;

[0167] y i =[p i,1 ,p i,2 ,……,p i,K ], where p i,j, represents the probability of the i-th subgraph corresponding to the j-th weather condition, j = 1 ~ K; K is the total number of weather categories (including 18 typical weather conditions in plateau mountainous areas, such as no disaster weather, light rain, moderate rain, heavy rain, thunderstorm, light fog, moderate fog, heavy fog, thick fog, light snow, moderate snow, heavy snow, blizzard, hail, sandstorm, halo, icy road surface). For example, y1 represents the weather condition probability distribution obtained by processing the subgraph f1 by the improved vision encoder model, which is a K-dimensional vector, y i = [p i,1 , p i,2 , …, p i,K ], where p 1,j represents the probability of the i-th subgraph corresponding to the j-th weather condition, K is the total number of weather categories; y2 represents the weather condition probability distribution obtained by processing the subgraph f2 by the improved vision encoder model, which is also a K-dimensional vector, y i = [p i,1 , p i,2 , …, p i,K ]; and so on to y N .

[0168] The parameters of the improved vision encoder model are shown in Table 2:

[0169] Table 2

[0170]

[0171] 4. Dynamic weight generation and fusion mechanism

[0172] A core innovation of the present application is the dynamic weight generation mechanism, which is not a complex independent network, but a learnable calculation link of an end-to-end model. The weight parameter generation method corresponding to each subgraph is realized by using a lightweight, learnable dynamic weight calculation module. The input of the module is only the calculated road area ratio r i and the weather condition probability distribution y i .

[0173] The specific calculation process is as follows:

[0174] First, for the i-th subgraph, the road area ratio (a scalar) r i and the K-dimensional weather condition probability distribution vector y i are concatenated to form a (K+1)-dimensional input feature vector x i :

[0175] x i = concat(r i , y i )

[0176] Subsequently, the feature vector x i is sent into a light-weight fully connected neural network (FCNN) which usually contains two hidden layers (with dimension 32) to learn the non-linear relationship between input features and weight importance, and outputs a weight score s i .

[0177] Finally, to ensure the sum of all subgraph weights is 1, forming an effective attention distribution mechanism, the model will collect the weight scores s i of all N subgraphs, and then normalize them by a Softmax function (which will assign the highest weight to the subgraph with the highest score), obtaining the final weight parameter w i :

[0178]

[0179] In this way, the model can automatically and learnably assign reasonable weights to each subgraph according to the preliminary judgment of road surface information and weather, so as to highlight the contribution of key areas in the final fusion step.

[0180] 5. Adaptive fusion

[0181] The final classification result is obtained by weighted fusion of the weather condition probability distribution of each subgraph:

[0182]

[0183] where y is also a K-dimensional probability vector representing the probability distribution of each weather condition in the whole image. y = [p1, p2,..., p K ], where p j represents the probability of the whole image corresponding to the jth weather condition, and K is the total number of weather categories; then take y as the final classification result of the whole image of the mountainous highway weather condition.

[0184] This fusion method enables the method proposed by the present application to adaptively focus on the most representative and traffic safety related areas according to different scenarios, thereby improving the overall recognition accuracy.

[0185] As shown in Figure 11 , a highland mountainous highway weather classification system using image partition feature dynamic fusion, which adopts the above-mentioned highland mountainous highway weather classification method using image partition feature dynamic fusion, comprises:

[0186] The mask generation module is configured to identify a road surface area in the collected road image with a road surface, and generate a corresponding road surface area mask image.

[0187] The partition and proportion calculation module is connected to the mask generation module, and is configured to divide the road image into a plurality of sub-images, and calculate a proportion of the road surface area in each sub-image according to the road surface area mask image.

[0188] The weather feature extraction module is connected to the partition and proportion calculation module, and is configured to extract weather features of each sub-image, and calculate a probability distribution of each type of meteorological condition of each sub-image.

[0189] The weight calculation module is connected to the partition and proportion calculation module and the weather feature extraction module, and is configured to generate corresponding weight parameters based on the probability distribution of each type of meteorological condition and the road surface proportion information of each sub-image.

[0190] The weighted fusion module is connected to the weight calculation module and the weather feature extraction module, and is configured to obtain the probability distribution of each type of meteorological condition of each sub-image by weighting the weight parameters obtained by the weight calculation module, and obtain a final meteorological condition classification result of the whole image.

[0191] Application example

[0192] 1. Mask

[0193] First, the collected road image F with a road surface will pass through any pre-trained road surface area recognition model applicable to the framework of the present method to identify the road surface area in the image and generate a corresponding mask image m. The pixel value on the mask image is 1 for the road surface area and 0 for the non-road surface area.

[0194] 2. Sub-image division and road surface proportion calculation

[0195] In this example, the road image F is divided into nine sub-images f1 to f9 in the form of a nine-square grid. For each sub-image, the intersection of the sub-image and the road surface mask image m is calculated to obtain the proportion r1 to r9 of the road surface area in the sub-image. These road surface proportion parameters will be input into the model as part of the features to participate in the subsequent calculation process.

[0196] The calculation formula is:

[0197] 3. Multi-region feature extraction and processing

[0198] For each sub-image, the improved vision encoder model (Vision Encoder) extracts its image features and calculates the corresponding meteorological condition probability distribution (y1, y2,..., y9) of the region. Each distribution y iis a K-dimensional vector, which contains the probability estimates of various possible weather categories in the corresponding subgraph. In this embodiment, K = 18, and these weather categories include 18 typical plateau mountain meteorological conditions such as no disaster weather, light rain, moderate rain, heavy rain, heavy rain, lightning, light fog, moderate fog, heavy fog, thick fog, light snow, moderate snow, heavy snow, blizzard, hail, sandstorm, halo, and icy road surface.

[0199] 4. Dynamic weight fusion

[0200] The learnable dynamic weight generation module can generate a set of dynamic weight parameters (w1, w2,..., w9) based on the meteorological condition probability distribution of each subgraph and the area ratio information of the road surface. These weight parameters are used to fuse the meteorological condition probability distribution of each subgraph to obtain the final classification result of the overall image:

[0201]

[0202] In this way, the model can adaptively adjust the importance of different subgraphs in the final decision, especially focusing more on the areas that have a direct impact on traffic safety.

[0203] 5. Model training

[0204] In the training phase, the method proposed in the present application learns feature extraction, subgraph meteorological condition probability distribution calculation, and dynamic weight generation through an end-to-end manner. By minimizing the classification loss function, the model can automatically learn the appropriate weight distribution strategy, so that the final fusion result can accurately reflect the meteorological conditions of the overall image, especially the key meteorological factors that have a practical impact on traffic safety.

[0205] The training target can be represented as:

[0206] where θ represents the model parameters, represents the classification loss function (such as binary cross-entropy loss), y is the final prediction output of the model, i.e., the meteorological condition probability vector of the entire image obtained by weighted fusion of the probability distribution of each subgraph (each bit is a decimal between 0 and 1), which is a "multi-label probability distribution vector", where each element corresponds to the probability of occurrence of a specific weather phenomenon, allowing the judgment of multiple weather phenomena. y true is the true label corresponding to the image, which is a vector with the same dimension as y, and each bit in the vector corresponds to a "standard answer" (0 or 1) of a weather category.

[0207] The main parameters of the end-to-end model of the present application are shown in Table 3:

[0208] Table 3

[0209]

[0210]

[0211] Figures 2 to 5 The whole process of the example of the present invention is shown.

[0212] The collected images are as follows Figure 2 As shown, then mask it and get Figure 3 The mask shown;

[0213] The highway image is then divided into multiple sub-images, and the proportion of the road area in each sub-image is calculated based on the road area mask image; weather features are extracted from each sub-image, and the probability distribution of various meteorological conditions in each sub-image is calculated; based on the probability distribution of various meteorological conditions and road area proportion information of each sub-image, the corresponding weight parameters are generated, such as Figure 4 As shown;

[0214] The obtained weight parameters are used to weight the probability distribution of various meteorological conditions of each sub-image to obtain the final meteorological condition classification result of the entire image, such as Figure 5 As shown in the figure, the recognition results (visible light: no disaster weather (confidence: 0.901), moderate fog (visibility 500-1000 meters) (confidence: 0.847)) accurately eliminate the interference of snowy weather characteristics outside the road surface area of ​​the highway image. The confidence level is the probability of the weather condition.

[0215] Figure 6 and Figure 7 These are the meteorological classification results for the same highway image under the non-partitioned global image weather recognition model. It can be noticed that the recognition results of both global models are snowy weather, which is inconsistent with the application of actual traffic scenarios (there is no snowfall in the road area).

[0216] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A plateau mountain highway weather classification method based on dynamic fusion of image partition features, characterized by: The steps include: Step (1), performing road surface area recognition on the collected highway image with the road surface, and generating a corresponding road surface area mask map; Step (2), dividing the highway image into multiple sub-images, and calculating the proportion of the road area in each sub-image based on the road area mask image; Step (3), extract weather features from each sub-graph and calculate the probability distribution of various meteorological conditions for each sub-graph; Step (4): Generate corresponding weight parameters based on the probability distribution of various meteorological conditions and road surface ratio information of each subgraph; Step (5) is to weight and fuse the probability distributions of various meteorological conditions of each sub-image using the weight parameters obtained in step (4) to obtain the final meteorological condition classification result of the entire image.

2. The plateau mountainous highway weather classification method based on dynamic fusion of image partition features according to claim 1 is characterized in that: In step (1), the collected highway image with a road surface is subjected to road surface area recognition. The specific method for generating the corresponding mask map is as follows: the collected highway image with a road surface is recognized, the road surface area and the non-road surface area are identified, and then the pixel value of the road surface area is set to 1, and the pixel value of the non-road surface area is set to 0, thereby obtaining the mask map.

3. The plateau mountainous highway weather classification method based on dynamic fusion of image partition features according to claim 2 is characterized in that: For a highway image F with a road surface, the road surface mask m is expressed as: m=RoadSegModel(F) Among them, RoadSegModel is the road area recognition model, and m is the generated binary mask image, which satisfies: m(x,y) represents the pixel value at coordinate (x,y) on the mask image.

4. The plateau mountainous highway weather classification method based on dynamic fusion of image partition features according to claim 1 is characterized in that: The specific method of step (2) is: Divide the highway image into N sub-images, denoted as f1 to f N ; For each sub-image, calculate its intersection with the road mask image m to obtain the proportion r of the road area in the sub-image i , i = 1 to N; the calculation formula is: Where m(x,y) represents the pixel value at the coordinate (x,y) on the mask image, 1 represents road surface, and 0 represents non-road surface; f i is the i-th subgraph.

5. The plateau mountainous highway weather classification method based on dynamic fusion of image partition features according to claim 1 is characterized in that: The specific method of step (3) is: using the improved visual encoder model, for each sub-image f i The weather characteristics are extracted to obtain the weather feature vector e i , and then calculate the probability distribution y of the meteorological conditions of each subgraph based on these features i , i=1~N; y i =[p i,1 ,p i,2 ,……,p i,K ], where p i,j represents the probability that the i-th subgraph corresponds to the j-th type of weather condition, j = 1 to K; K is the total number of weather categories; y i The calculation method is as follows: the weather characteristic vector e i Map it to the weather category through the fully connected layer to obtain the unnormalized category score Z i , and then normalized by the Softmax function to obtain the meteorological condition probability distribution y i ; Z i =w T e i +b y i =Softmax(Z i ) Where w is the initialized learning weight matrix, T represents the transpose operation, b is the initialization bias vector, and Softmax represents the normalized exponential function; The improved visual encoder model includes a visual Transformer module, a dynamic feature fusion module, and an output module; the visual Transformer module, the dynamic feature fusion module, and the output module are connected sequentially; The dynamic feature fusion module includes a spatial branch and a channel branch; the spatial branch captures local spatial details through 3×3 depthwise separable convolution; the channel branch compresses the channel dimension through sequentially connected 1×1 convolutions and the SE channel attention mechanism to extract global semantic information. The number of channels after compression is C / r=16, where C is the number of channels before compression and r is the compression ratio; the two branches extract features in parallel, fuse them, and output them to obtain a weather feature vector.

6. The plateau mountainous highway weather classification method based on dynamic fusion of image partition features according to claim 5 is characterized in that: K=18, weather categories include non-hazardous weather, light rain, moderate rain, heavy rain, rainstorm, lightning, light fog, moderate fog, heavy fog, dense fog, light snow, moderate snow, heavy snow, blizzard, hail, sandstorm, halo, and icy and snowy road surface; compression ratio r=16.

7. The plateau mountainous highway weather classification method based on dynamic fusion of image partition features according to claim 1 is characterized in that: In step (4), the specific method of generating weight parameters is as follows: (a) For each sub-graph i, its road area ratio r i and the probability distribution of weather conditions y i Splice to form an input feature vector x i ; (b) The input feature vector x i Input into the fully connected neural network, and the weight score s corresponding to the subgraph is calculated through the network i ; (c) Weight scores s1 to s1 for all N subgraphs N Apply the Softmax function for normalization to obtain the final weight parameter of each subgraph, where the final weight parameter of subgraph i is w i .

8. The plateau mountainous highway weather classification method based on dynamic fusion of image partition features according to claim 1 is characterized in that: The specific method of step (5) is: The weight parameters obtained in step (4) are weighted to fuse the probability distribution of meteorological conditions of each sub-graph and calculated by the following formula: Where y is a K-dimensional probability vector, y=[p1,p2,...,p K ], p j represents the probability that the entire image corresponds to the jth type of weather condition, and K is the total number of weather categories; Then y is used as the final classification result of the mountain road meteorological conditions of the entire image.

9. A plateau mountainous highway meteorological classification system with dynamic fusion of image partition features, which adopts the plateau mountainous highway meteorological classification method with dynamic fusion of image partition features according to any one of claims 1 to 8, characterized in that: include: The mask generation module is used to identify the road surface area of ​​the collected highway image with the road surface and generate a corresponding road surface area mask map; The partitioning and proportion calculation module is connected to the mask generation module and is used to divide the highway image into multiple sub-images and calculate the proportion of the road area in each sub-image based on the road area mask map; The weather feature extraction module is connected to the partition and proportion calculation module, and is used to extract weather features from each sub-graph and calculate the probability distribution of various meteorological conditions in each sub-graph; The weight calculation module is connected to the partition and proportion calculation module and the weather feature extraction module respectively, and is used to generate corresponding weight parameters based on the probability distribution of various meteorological conditions and road surface proportion information of each sub-graph; The weighted fusion module is connected to the weight calculation module and the weather feature extraction module respectively. It is used to weightedly fuse the probability distribution of various meteorological conditions of each sub-image through the weight parameters obtained by the weight calculation module to obtain the final meteorological condition classification result of the entire image.

Images