Urban and rural planning surveying and mapping data analysis method based on neural network

By combining multi-scale convolutional neural network, graph convolutional network and other neural network technologies, the problems of low efficiency and insufficient recognition accuracy in urban and rural planning surveying and mapping data analysis are solved, efficient and automated data processing and analysis are realized, and planning is improved scientific and practical.

CN118968296AInactive Publication Date: 2024-11-15YUYAO PLANNING SURVEYING & DESIGNING INST
View PDF 0 Cites 19 Cited by

Patent Information

Application Number
CN202411008041.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-26
Publication Date
2024-11-15
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing urban and rural planning surveying and mapping data analysis methods are inefficient and have limited identification accuracy when processing large-scale multi-source data. Traditional methods are difficult to ensure data consistency and wide deployment of models. Optimization is difficult, computational complexity is high, and there is a lack of comprehensive utilization. The ability of source data leads to insufficient scientificity and practicality of planning suggestions.

Method used

Using multi-scale convolutional neural networks, graph convolutional networks, attention mechanisms, generative adversarial networks and reinforcement learning technologies to carry out data acquisition and preprocessing, feature extraction and classification, feature fusion and analysis, data enhancement and completion and dynamics Optimization and model updates, and high-precision geographic feature recognition and automated data processing are achieved through multi-level supervised learning, image pyramid method, self-supervised learning, multi-head self-attention mechanism, generative adversarial networks and reinforcement learning algorithms.

Benefits of technology

It significantly improves the processing efficiency and analysis accuracy of urban and rural planning surveying and mapping data, provides more scientific and reliable data support, improves the scientificity and practicality of the planning, and ensures the stability and adaptability of the model in different environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118968296A_ABST
    Figure CN118968296A_ABST
Patent Text Reader

Abstract

The invention discloses an urban and rural planning surveying and mapping data analysis method based on a neural network. The method comprises the following steps: acquiring and preprocessing a remote sensing image, a topographic map and vector data; constructing and training a multi-scale convolutional neural network model, and optimally extracting and classifying geographic features in the multi-source geographic information data; integrating a graph convolutional network module based on an attention mechanism, and carrying out weighted processing and fusion analysis on the geographic features; carrying out feature pre-training on the unlabeled data by utilizing self-supervised learning, and carrying out data enhancement and missing data complementation through a generative adversarial network; generating a high-precision geographic feature recognition graph according to the classification result and the spatial relationship, optimizing the recognition graph in combination with a reinforcement learning algorithm, and generating a geographic information data tag; the labels are comprehensively analyzed, planning suggestions and reports are generated in combination with urban and rural planning requirements, and the model is updated and optimized through continuous learning and transfer learning technologies.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and geographic information system, and in particular to a method for analyzing urban and rural planning surveying and mapping data based on a neural network. Background Art

[0002] In the field of urban and rural planning and surveying and mapping, geographic information system (GIS) has become an important tool. GIS technology can effectively collect, store, manage, analyze and display various geospatial data, which provides important support for many fields such as urban and rural planning, environmental protection, and resource management. However, with the development of remote sensing technology, drone technology, and sensor technology, the amount and types of data that GIS needs to process have increased significantly, and traditional manual processing methods can no longer meet the current needs for efficient and accurate data analysis.

[0003] Geographic information data processing technology plays an important role in urban and rural planning, but existing methods have some limitations. Remote sensing images, topographic maps, and vector data are the main sources of data acquisition, which provide rich surface information and terrain features. However, data preprocessing steps, such as normalization, denoising, and format conversion, although they can improve data quality, are inefficient and difficult to ensure consistency when processing large-scale data.

[0004] Multi-scale convolutional neural networks (MSCNN) have made progress in feature extraction and classification, but they still face challenges when it comes to multi-source data fusion and analysis. Traditional methods often ignore the complexity of spatial relationships, resulting in limited recognition accuracy and classification effects. In addition, the optimization difficulty and high computational complexity of the MSCNN model limit its widespread deployment in practical applications.

[0005] The combination of graph convolutional networks and attention mechanisms provides a new perspective for the identification and analysis of geographic features. The multi-head self-attention mechanism and graph attention network can perform weighted processing on geographic features and enhance the representation of spatial relationship features. However, this method has problems of optimization difficulty and computational complexity in the implementation process.

[0006] Generative adversarial networks (GANs) perform well in data augmentation and missing data completion, and can generate samples that are similar to the real data distribution, enhancing the diversity and completeness of the data set. However, the training process of GANs is unstable, and the quality and diversity of the generated samples are difficult to guarantee, which poses a challenge to the reliability of the data. Reinforcement learning algorithms have shown potential in the dynamic optimization and adaptive adjustment of geographic feature recognition maps, and can generate more accurate geographic information data labels based on classification results and spatial relationship analysis. However, its application in large-scale data processing is limited by computing resources and algorithm complexity.

[0007] It is a complex task to comprehensively analyze geographic information data labels and generate planning recommendations and reports based on urban and rural planning needs. Traditional methods lack the comprehensive use of multi-source data, resulting in insufficient scientificity and practicality of planning recommendations. Although continuous learning mechanisms and transfer learning technologies can improve the adaptability and robustness of models in regular updates and optimizations of models, it is still a challenge to maintain the consistency of model performance in large-scale applications.

[0008] Therefore, how to provide a method for analyzing urban and rural planning and surveying data based on neural networks is an urgent problem that technicians in this field need to solve. Summary of the invention

[0009] One purpose of the present invention is to propose a method for analyzing urban and rural planning surveying and mapping data based on a neural network. The present invention makes full use of technologies such as multi-scale convolutional neural networks, graph convolutional networks, attention mechanisms, generative adversarial networks, and reinforcement learning, and describes in detail the entire process of data acquisition and preprocessing, feature extraction and classification, feature fusion and analysis, data enhancement and completion, as well as dynamic optimization and model updating. It has the advantages of high-precision geographic feature recognition, automated data processing, comprehensive data analysis, and model adaptive optimization. Through this method, the processing efficiency and analysis accuracy of urban and rural planning surveying and mapping data can be greatly improved, providing more scientific and reliable data support for urban and rural planning.

[0010] A method for analyzing urban and rural planning surveying and mapping data based on a neural network according to an embodiment of the present invention comprises the following steps:

[0011] S1. Obtain remote sensing images, topographic maps and vector data, and perform preprocessing;

[0012] S2. Build and train a multi-scale convolutional neural network model, optimize it through multi-level supervised learning method and image pyramid method, extract geographic features of different scales and improve recognition ability, input the preprocessed multi-source geographic information data into the multi-scale convolutional neural network model for feature extraction and classification;

[0013] S3. Integrate the graph convolutional network module based on the attention mechanism, perform weighted processing on the extracted geographic features through the multi-head self-attention mechanism and the graph attention network, use the position encoding technology to enhance the representation ability of spatial relationship features, and use the multimodal fusion technology to fuse the features of different data sources, and use the cross-modal conversion network to achieve feature complementarity and joint analysis;

[0014] S4. Use self-supervised learning methods to pre-train features of unlabeled data, and use generative adversarial networks to perform data enhancement and missing data completion;

[0015] S5. Generate a high-precision geographic feature recognition map based on the classification results and spatial relationship analysis, and combine the optimization module of the reinforcement learning algorithm to dynamically optimize and adaptively adjust the recognition map to generate geographic information data labels;

[0016] S6. Comprehensively analyze geographic information data labels, combine urban and rural planning needs and actual conditions, generate planning recommendations and reports, and use continuous learning mechanisms and transfer learning techniques to regularly update and optimize multi-scale convolutional neural network models, graph convolutional network models, and cross-modal conversion network models.

[0017] Optionally, the step S1 specifically includes:

[0018] S11. Acquire remote sensing image data, topographic map data and vector data;

[0019] S12. Perform data normalization on the acquired data:

[0020]

[0021] Where x is the original data value, x' is the normalized data value, min(x) and max(x) are the minimum and maximum values ​​in the data set respectively;

[0022] S13. Perform denoising on the normalized data using a Gaussian filter:

[0023]

[0024] Among them, G(x,y) is the filtered value, σ is the standard deviation, and x and y are the coordinates of the pixel points;

[0025] S14. Perform geometric correction and radiation correction on the denoised data, where the geometric correction is achieved through the geometric transformation matrix T:

[0026]

[0027] Among them, (x, y) is the image coordinate, (x', y') is the geographic coordinate, and T is the geometric transformation matrix;

[0028] Radiometric correction:

[0029]

[0030] Among them, L is the corrected radiation brightness value, L max and L min are the maximum and minimum radiation brightness values, Q is the original radiation value, Q max and Q min are the maximum and minimum values ​​of the original radiation values ​​respectively;

[0031] S15. Convert the preprocessed data into a data format, including converting vector data into raster data:

[0032]

[0033] Among them, R(i,j) is the value of the grid data at position (i,j), V k is the attribute value of the vector data, f k (i,j) is the rasterization function, and n is the number of vector data.

[0034] Optionally, the step S2 specifically includes:

[0035] S21. Construct a multi-scale convolutional neural network model, wherein the model includes multiple convolutional layers, pooling layers and fully connected layers;

[0036] S22. Design a convolutional layer, where the size and number of convolution kernels are k×k and n respectively, and the output feature map of the convolutional layer is expressed as:

[0037]

[0038] Among them, F l (i, j) is the input feature map of the lth layer at position (i, j), W l (p,q) is the convolution kernel of the lth layer at position (p,q), b l is the bias, f is the activation function;

[0039] S23. Design the pooling layer and use the maximum pooling operation with a pooling window size of m×m:

[0040] P i,j =max 0≤p<m,0≤q<m F i+p,j+q ;

[0041] Among them, P i,j is the pooling result, F i+p,j+q is the feature value within the pooling window;

[0042] S24. The multi-scale convolutional neural network model is trained by a multi-level supervised learning method, wherein the training process includes forward propagation and back propagation, wherein the loss function is a cross entropy loss function:

[0043]

[0044] Where N is the number of samples, y i is the true label, is the predicted label;

[0045] S25. Use the image pyramid method to perform multi-scale processing on the input remote sensing image data and generate image pyramids of different resolutions:

[0046]

[0047] Among them, I s is the s-th layer pyramid image, I original is the original image, s is the number of pyramid layers;

[0048] S26. Input the generated multi-scale image pyramid into the multi-scale convolutional neural network model for feature extraction:

[0049] F s,l (i,j)=MSCNN(I s (i,j));

[0050] Among them, F s,l (i, j) is the value of the feature map of the s-th layer pyramid image at position (i, j) in the l-th layer convolutional neural network;

[0051] S27. Optimize the multi-scale convolutional neural network model using the Adam optimization algorithm. The parameter update process includes the following steps:

[0052]

[0053] Among them, m t and v t are the first-order and second-order momentum estimates at time t, β1 and β2 are the decay rates of the momentum term, is the loss function L t For the model parameters θ t The gradient of

[0054] Perform bias correction on the momentum and compute the bias-corrected momentum estimate:

[0055]

[0056] Update model parameters:

[0057]

[0058] Among them, η is the learning rate and is the smoothing term;

[0059] S28. Input the preprocessed multi-source geographic information data into the multi-scale convolutional neural network model for feature extraction and classification:

[0060] F output (i,j)=MSCNN(I input (i,j));

[0061] Among them, F output (i, j) is the output feature of the input data at position (i, j), I input (i,j) is the preprocessed input data.

[0062] Optionally, the step S3 specifically includes:

[0063] S31. Construct a graph convolution network module based on the attention mechanism, which includes a graph convolution layer and a multi-head self-attention mechanism, and processes the multi-scale features of the input through the graph convolution layer:

[0064] H (l+1) =σ(D -1 / 2 AD -1 / 2 H (l) W (l) );

[0065] Among them, H (l) is the input feature matrix of the lth layer, A is the adjacency matrix, D is the degree matrix, W (l) is the weight matrix of the lth layer, σ is the activation function;

[0066] S32. Use the multi-head self-attention mechanism to weight the output features of the graph convolution layer:

[0067]

[0068] Among them, Q is the query matrix, K is the key matrix, and V is the value matrix. and They are the weight matrices of the i-th attention head:

[0069]

[0070] Among them, d k is the dimension of the key matrix, h is the number of attention heads;

[0071] S33. Use position coding technology to enhance the representation ability of spatial relationship features. The calculation formula of position coding is:

[0072]

[0073] Combining the weighted features with the positional encoding, the enhanced features are:

[0074] F enhanced =Z+PE;

[0075] Among them, F enhanced is the enhanced feature, Z is the weighted feature, and Z is the position code;

[0076] S34. Use multimodal fusion technology to fuse the features of different data sources;

[0077] S35. Using cross-modal conversion networks to achieve feature complementarity and joint analysis;

[0078] Optionally, the step S4 specifically includes:

[0079] S41. Use the self-supervised learning method to pre-train the features of unlabeled data, adopt the contrastive learning strategy, construct positive and negative sample pairs, and the contrastive learning loss function is:

[0080]

[0081] Where N is the number of samples, x i is the input sample, is a positive sample, is a negative sample, y i is the sample pair label, m is the boundary distance, and f is the feature extraction function;

[0082] S42. Data enhancement and missing data completion are performed through generative adversarial networks. The generative adversarial network consists of a generator G and a discriminator D. The objective function of the generative adversarial network is:

[0083]

[0084] Among them, G is the generator, D is the discriminator, x is the real data, z is the noise vector, p data (x) is the real data distribution, p z (z) is the noise distribution;

[0085] S43. Train the generator G to generate samples similar to the real data distribution:

[0086]

[0087] Among them, θ G is the parameter of the generator, η G is the learning rate, m is the batch size, z i is the noise vector;

[0088] S44. Train the discriminator D to distinguish between real data and generated data:

[0089]

[0090] Among them, θ D is the parameter of the discriminator, η D is the learning rate, x i is the real data sample, z i is the noise vector;

[0091] S45. Use data generated by generative adversarial networks for data augmentation and missing data completion.

[0092] Optionally, the step S5 specifically includes:

[0093] S51. Based on the classification results and spatial relationship analysis, a preliminary geographic feature recognition map is generated, and a convolutional neural network is used to perform spatial relationship analysis on the classification results:

[0094]

[0095] Among them, R(i,j) is the spatial relationship feature value of position (i,j), W(k,l) is the convolution kernel weight, C(i+k,j+l) is the classification result of position (i+k,j+l), and m and n are the height and width of the convolution kernel respectively;

[0096] S52. Use the reinforcement learning algorithm to dynamically optimize and adaptively adjust the preliminary geographic feature recognition map, define the state s, action a and reward r, and the optimization goal is to maximize the cumulative reward:

[0097]

[0098] Among them, π * (s) is the optimal strategy, γ is the discount factor, r t is the reward at time t, s is the state, π is the strategy, and T is the number of time steps;

[0099] S53. Using deep networks for strategy optimization in the framework of reinforcement learning:

[0100] Q(s t ,a t )←Q(s t ,a t )+α[r t +γmax a' Q(s t+1 ,a')-Q(s t ,a t )];

[0101] Among them, Q(s t ,a t ) is state s t and action a t Q value, α is the learning rate, γ is the discount factor, r t is the reward value, s t+1 is the next state, a' is the next action;

[0102] S54. Combined with the generative adversarial network, the recognition graph is further optimized. The objective function of the generative adversarial network is:

[0103]

[0104] Among them, G is the generator, D is the discriminator, x is the real data, z is the noise vector, p data (x) is the real data distribution, p z (z) is the noise distribution;

[0105] S55. Combine the optimized recognition map with multi-source geographic information data to generate a high-precision geographic feature recognition map, and use a comprehensive analysis algorithm to evaluate and correct the results:

[0106] F final (i,j)=α·F opt (i,j)+(1-α)·G info (i,j);

[0107] Among them, F final (i, j) is the final eigenvalue at position (i, j), F opt (i,j) is the optimized eigenvalue, G info (i, j) is the characteristic value of multi-source geographic information data, and α is the weight coefficient.

[0108] Optionally, the step S6 specifically includes:

[0109] S61. Comprehensively analyze geographic information data labels, conduct joint analysis on fused features, and use a multi-task learning framework:

[0110] L total =αL classification +βL regression +γL segmentation ;

[0111] Among them, L classification is the classification loss, L regression is the regression loss, L segmentation is the image segmentation loss, α, β and γ are weight coefficients;

[0112] S62. Combine the urban and rural planning needs and actual conditions to comprehensively process the analysis results and generate preliminary planning recommendations and reports;

[0113] S63. Use the reinforcement learning algorithm to optimize the preliminary planning suggestions, define the state s, action a and reward r, and the optimization goal is to maximize the cumulative reward:

[0114]

[0115] Among them, π* (s) is the optimal strategy, γ is the discount factor, r t is the reward at time t, s is the state, π is the strategy, and T is the number of time steps;

[0116] S64. Generate a final planning report based on the optimized planning suggestions, and provide specific planning solutions and implementation suggestions in combination with urban and rural planning needs and actual conditions;

[0117] S65. Use continuous learning mechanisms and transfer learning techniques to regularly update and optimize multi-scale convolutional neural network models, graph convolutional network models, and cross-modal conversion network models:

[0118]

[0119] Among them, L new is the loss of the new task, λ is the regularization parameter, F i is the Fisher information matrix, θ i are the current model parameters, are the model parameters of the old task;

[0120] S66. In the transfer learning framework, adversarial training is used for feature alignment. The feature extraction of the source domain and the target domain are f s (x s ) and f t (x t ):

[0121]

[0122] Among them, Dist is the feature distribution distance function, Loss t is the target domain loss function, α and β are weight coefficients.

[0123] The beneficial effects of the present invention are:

[0124] The present invention significantly improves the analysis accuracy and processing efficiency of urban and rural planning and mapping data by comprehensively applying a variety of advanced neural network technologies and machine learning methods. The multi-scale convolutional neural network model is used to extract and classify geographical features such as buildings, roads and rivers with high precision. Combined with the graph convolutional network module and the multi-head self-attention mechanism, the geographical features are weighted and fused to improve the accuracy of recognition and the ability to express details. Self-supervised learning and generative adversarial networks realize automated data processing, perform feature pre-training, data enhancement and missing data completion on unlabeled data, reduce manual intervention, and improve the consistency and efficiency of data processing. The automated preprocessing steps include data normalization, denoising and data format conversion, which effectively improves data quality. Multimodal fusion technology realizes the complementary and joint analysis of features from different data sources, making data analysis more comprehensive and accurate. Comprehensively analyze geographic information data labels, combine with actual urban and rural planning needs, and generate more scientific and feasible planning suggestions and reports. Reinforcement learning algorithms are used to dynamically optimize and adaptively adjust geographic feature recognition maps, and continuously optimize the accuracy and applicability of recognition maps based on classification results and spatial relationship analysis. Through continuous learning and transfer learning techniques, the model can be regularly updated and optimized to maintain stability and adaptability in different environments and conditions. Generative adversarial networks (GANs) effectively improve the diversity and completeness of data sets, and enhance the generalization and robustness of models. The high efficiency of data augmentation and missing data completion enables the model to handle more diverse and complex geographic information data. The continuous learning mechanism ensures that the model continuously optimizes itself as new data and new environments change, avoiding model degradation and maintaining high performance. Transfer learning technology enables the model to quickly adapt to new data distributions and features, improving flexibility and adaptability in practical applications.

[0125] The present invention realizes high-precision automatic analysis and processing of urban and rural planning surveying and mapping data through the comprehensive application of multiple advanced technologies. Compared with traditional methods, the present invention has significantly improved data processing efficiency, feature recognition accuracy, data fusion and analysis capabilities, dynamic optimization and adaptive adjustment, model updating and optimization, etc., providing more scientific and reliable data support for urban and rural planning, and significantly improving the scientificity, practicality and accuracy of planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0126] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0127] Figure 1 This is a general flow chart of a method for analyzing urban and rural planning surveying and mapping data based on a neural network proposed by the present invention;

[0128] Figure 2A flowchart for constructing and training a multi-scale convolutional neural network model in the present invention;

[0129] Figure 3 It is a structural diagram of the graph convolutional network module based on the attention mechanism in the present invention;

[0130] Figure 4 Flowchart for data enhancement and missing data completion for self-supervised learning and generative adversarial networks in the present invention. DETAILED DESCRIPTION

[0131] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0132] refer to Figure 1 , a method for analyzing urban and rural planning surveying and mapping data based on neural network, comprising the following steps:

[0133] S1. Obtain remote sensing images, topographic maps and vector data, and perform preprocessing;

[0134] In this implementation, step S1 specifically includes:

[0135] S11. Acquire remote sensing image data, topographic map data and vector data;

[0136] S12. Perform data normalization on the acquired data:

[0137]

[0138] Where x is the original data value, x' is the normalized data value, min(x) and max(x) are the minimum and maximum values ​​in the data set respectively;

[0139] S13. Perform denoising on the normalized data using a Gaussian filter:

[0140]

[0141] Among them, G(x,y) is the filtered value, σ is the standard deviation, and x and y are the coordinates of the pixel points;

[0142] S14. Perform geometric correction and radiation correction on the denoised data, where the geometric correction is achieved through the geometric transformation matrix T:

[0143]

[0144] Among them, (x, y) is the image coordinate, (x', y') is the geographic coordinate, and T is the geometric transformation matrix;

[0145] Radiometric correction:

[0146]

[0147] Among them, L is the corrected radiation brightness value, L max and L min are the maximum and minimum radiation brightness values, Q is the original radiation value, Q max and Q min are the maximum and minimum values ​​of the original radiation values ​​respectively;

[0148] S15. Convert the preprocessed data into a data format, including converting vector data into raster data:

[0149]

[0150] Among them, R(i,j) is the value of the grid data at position (i,j), V k is the attribute value of the vector data, f k (i,j) is the rasterization function, and n is the number of vector data.

[0151] S2. Build and train a multi-scale convolutional neural network model, optimize it through multi-level supervised learning method and image pyramid method, extract geographic features of different scales and improve recognition ability, input the preprocessed multi-source geographic information data into the multi-scale convolutional neural network model for feature extraction and classification;

[0152] In this implementation, step S2 specifically includes:

[0153] S21. Construct a multi-scale convolutional neural network model, wherein the model includes multiple convolutional layers, pooling layers and fully connected layers;

[0154] S22. Design a convolutional layer, where the size and number of convolution kernels are k×k and n respectively, and the output feature map of the convolutional layer is expressed as:

[0155]

[0156] Among them, F l (i, j) is the input feature map of the lth layer at position (i, j), W l (p,q) is the convolution kernel of the lth layer at position (p,q), b l is the bias, f is the activation function;

[0157] S23. Design the pooling layer and use the maximum pooling operation with a pooling window size of m×m:

[0158] P i,j =max 0≤p<m,0≤q<m F i+p,j+q ;

[0159] Among them, P i,j is the pooling result, F i+p,j+q is the feature value within the pooling window;

[0160] S24. The multi-scale convolutional neural network model is trained by a multi-level supervised learning method, wherein the training process includes forward propagation and back propagation, wherein the loss function is a cross entropy loss function:

[0161]

[0162] Where N is the number of samples, y i is the true label, is the predicted label;

[0163] S25. Use the image pyramid method to perform multi-scale processing on the input remote sensing image data and generate image pyramids of different resolutions:

[0164]

[0165] Among them, I s is the s-th layer pyramid image, I original is the original image, s is the number of pyramid layers;

[0166] S26. Input the generated multi-scale image pyramid into the multi-scale convolutional neural network model for feature extraction:

[0167] F s,l (i,j)=MSCNN(I s (i,j));

[0168] Among them, F s,l (i, j) is the value of the feature map of the s-th layer pyramid image at position (i, j) in the l-th layer convolutional neural network;

[0169] S27. Optimize the multi-scale convolutional neural network model using the Adam optimization algorithm. The parameter update process includes the following steps:

[0170]

[0171] Among them, m t and v t are the first-order and second-order momentum estimates at time t, β1 and β2 are the decay rates of the momentum term, is the loss function L t For the model parameters θ t The gradient of

[0172] Perform bias correction on the momentum and compute the bias-corrected momentum estimate:

[0173]

[0174] Update model parameters:

[0175]

[0176] Among them, η is the learning rate and is the smoothing term;

[0177] S28. Input the preprocessed multi-source geographic information data into the multi-scale convolutional neural network model for feature extraction and classification:

[0178] F output (i,j)=MSCNN(I input (i,j));

[0179] Among them, F output (i, j) is the output feature of the input data at position (i, j), I input (i,j) is the preprocessed input data.

[0180] S3. Integrate the graph convolutional network module based on the attention mechanism, perform weighted processing on the extracted geographic features through the multi-head self-attention mechanism and the graph attention network, use the position encoding technology to enhance the representation ability of spatial relationship features, and use the multimodal fusion technology to fuse the features of different data sources, and use the cross-modal conversion network to achieve feature complementarity and joint analysis;

[0181] In this implementation, step S3 specifically includes:

[0182] S31. Construct a graph convolution network module based on the attention mechanism, which includes a graph convolution layer and a multi-head self-attention mechanism, and processes the multi-scale features of the input through the graph convolution layer:

[0183] H (l+1) =σ(D -1 / 2 AD -1 / 2 H (l) W (l) );

[0184] Among them, H (l) is the input feature matrix of the lth layer, A is the adjacency matrix, D is the degree matrix, W (l) is the weight matrix of the lth layer, σ is the activation function;

[0185] S32. Use the multi-head self-attention mechanism to weight the output features of the graph convolution layer:

[0186]

[0187] Among them, Q is the query matrix, K is the key matrix, and V is the value matrix. and They are the weight matrices of the i-th attention head:

[0188]

[0189] Among them, d k is the dimension of the key matrix, h is the number of attention heads;

[0190] S33. Use position coding technology to enhance the representation ability of spatial relationship features. The calculation formula of position coding is:

[0191]

[0192] Combining the weighted features with the positional encoding, the enhanced features are:

[0193] F enhanced =Z+PE;

[0194] Among them, F enhanced is the enhanced feature, Z is the weighted feature, and Z is the position code;

[0195] S34. Use multimodal fusion technology to fuse the features of different data sources;

[0196] S35. Using cross-modal conversion networks to achieve feature complementarity and joint analysis;

[0197] S4. Use self-supervised learning methods to pre-train features of unlabeled data, and use generative adversarial networks to perform data enhancement and missing data completion;

[0198] In this implementation, step S4 specifically includes:

[0199] S41. Use the self-supervised learning method to pre-train the features of unlabeled data, adopt the contrastive learning strategy, construct positive and negative sample pairs, and the contrastive learning loss function is:

[0200]

[0201] Where N is the number of samples, x i is the input sample, is a positive sample, is a negative sample, y i is the sample pair label, m is the boundary distance, and f is the feature extraction function;

[0202] S42. Data enhancement and missing data completion are performed through generative adversarial networks. The generative adversarial network consists of a generator G and a discriminator D. The objective function of the generative adversarial network is:

[0203]

[0204] Among them, G is the generator, D is the discriminator, x is the real data, z is the noise vector, p data (x) is the real data distribution, p z (z) is the noise distribution;

[0205] S43. Train the generator G to generate samples similar to the real data distribution:

[0206]

[0207] Among them, θ G is the parameter of the generator, η G is the learning rate, m is the batch size, z i is the noise vector;

[0208] S44. Train the discriminator D to distinguish between real data and generated data:

[0209]

[0210] Among them, θ D is the parameter of the discriminator, η D is the learning rate, x i is the real data sample, z i is the noise vector;

[0211] S45. Use data generated by generative adversarial networks for data augmentation and missing data completion.

[0212] S5. Generate a high-precision geographic feature recognition map based on the classification results and spatial relationship analysis, and combine the optimization module of the reinforcement learning algorithm to dynamically optimize and adaptively adjust the recognition map to generate geographic information data labels;

[0213] In this implementation, step S5 specifically includes:

[0214] S51. Based on the classification results and spatial relationship analysis, a preliminary geographic feature recognition map is generated, and a convolutional neural network is used to perform spatial relationship analysis on the classification results:

[0215]

[0216] Among them, R(i,j) is the spatial relationship feature value of position (i,j), W(k,l) is the convolution kernel weight, C(i+k,j+l) is the classification result of position (i+k,j+l), and m and n are the height and width of the convolution kernel respectively;

[0217] S52. Use the reinforcement learning algorithm to dynamically optimize and adaptively adjust the preliminary geographic feature recognition map, define the state s, action a and reward r, and the optimization goal is to maximize the cumulative reward:

[0218]

[0219] Among them, π * (s) is the optimal strategy, γ is the discount factor, r t is the reward at time t, s is the state, π is the strategy, and T is the number of time steps;

[0220] S53. Using deep networks for strategy optimization in the framework of reinforcement learning:

[0221] Q(s t ,a t )←Q(s t ,a t )+α[r t +γmax a' Q(s t+1 ,a')-Q(s t ,a t )];

[0222] Among them, Q(s t ,a t ) is state s t and action a t Q value, α is the learning rate, γ is the discount factor, r t is the reward value, s t+1 is the next state, a' is the next action;

[0223] S54. Combined with the generative adversarial network, the recognition graph is further optimized. The objective function of the generative adversarial network is:

[0224]

[0225] Among them, G is the generator, D is the discriminator, x is the real data, z is the noise vector, p data (x) is the real data distribution, p z (z) is the noise distribution;

[0226] S55. Combine the optimized recognition map with multi-source geographic information data to generate a high-precision geographic feature recognition map, and use a comprehensive analysis algorithm to evaluate and correct the results:

[0227] F final (i,j)=α·F opt (i,j)+(1-α)·G info (i,j);

[0228] Among them, F final (i, j) is the final eigenvalue at position (i, j), F opt (i,j) is the optimized eigenvalue, G info (i, j) is the characteristic value of multi-source geographic information data, and α is the weight coefficient.

[0229] S6. Comprehensively analyze geographic information data labels, combine urban and rural planning needs and actual conditions, generate planning recommendations and reports, and use continuous learning mechanisms and transfer learning techniques to regularly update and optimize multi-scale convolutional neural network models, graph convolutional network models, and cross-modal conversion network models.

[0230] In this implementation, step S6 specifically includes:

[0231] S61. Comprehensively analyze geographic information data labels, conduct joint analysis on fused features, and use a multi-task learning framework:

[0232] L total =αL classification +βL regression +γL segmentation ;

[0233] Among them, L classification is the classification loss, L regression is the regression loss, L segmentation is the image segmentation loss, α, β and γ are weight coefficients;

[0234] S62. Combine the urban and rural planning needs and actual conditions to comprehensively process the analysis results and generate preliminary planning recommendations and reports;

[0235] S63. Use the reinforcement learning algorithm to optimize the preliminary planning suggestions, define the state s, action a and reward r, and the optimization goal is to maximize the cumulative reward:

[0236]

[0237] Among them, π * (s) is the optimal strategy, γ is the discount factor, r t is the reward at time t, s is the state, π is the strategy, and T is the number of time steps;

[0238] S64. Generate a final planning report based on the optimized planning suggestions, and provide specific planning solutions and implementation suggestions in combination with urban and rural planning needs and actual conditions;

[0239] S65. Use continuous learning mechanisms and transfer learning techniques to regularly update and optimize multi-scale convolutional neural network models, graph convolutional network models, and cross-modal conversion network models:

[0240]

[0241] Among them, L new is the loss of the new task, λ is the regularization parameter, F i is the Fisher information matrix, θ i are the current model parameters, are the model parameters of the old task;

[0242] S66. In the transfer learning framework, adversarial training is used for feature alignment. The feature extraction of the source domain and the target domain are f s (x s ) and f t (x t ):

[0243]

[0244] Among them, Dist is the feature distribution distance function, Loss t is the target domain loss function, α and β are weight coefficients.

[0245] Embodiment 1:

[0246] In a coastal city in eastern China, the government plans to conduct a comprehensive urban and rural planning to cope with the rapid urbanization process and changing population distribution. The city has a complex geographical environment, including rivers, mountains and dense buildings. At the same time, it is also necessary to consider the existing transportation network, public facilities and ecological environment. Traditional urban and rural planning methods can no longer meet the needs of efficient and high-precision data analysis in such a complex environment. Therefore, the municipal government decided to adopt a neural network-based urban and rural planning surveying and mapping data analysis method of the present invention.

[0247] First, we obtained comprehensive geographic information about the city and its surrounding areas through high-resolution remote sensing images, detailed topographic maps, and precise vector data. These data included key elements such as buildings, roads, rivers, and green spaces. To ensure the quality and consistency of the data, we preprocessed the data, including data normalization, denoising, and data format conversion. Through these steps, we were able to ensure the accuracy and usability of the data in the subsequent analysis process.

[0248] After preprocessing, we built and trained a multi-scale convolutional neural network (MSCNN) model. The model is optimized through multi-level supervised learning methods and image pyramid methods, and can extract and classify geographic features of different scales. After training, the MSCNN model can accurately identify geographic features such as buildings, roads, and rivers. In practical applications, we input the preprocessed multi-source geographic information data into the MSCNN model for feature extraction and classification. The results show that the accuracy of the model in identifying geographic features such as buildings, roads, and rivers is significantly improved, and the error rate is reduced to less than 5%.

[0249] In order to further improve the accuracy of feature recognition and the ability to express details, we integrated a graph convolutional network module based on the attention mechanism. The extracted geographic features are weighted through the multi-head self-attention mechanism and the graph convolutional network, and the position encoding technology is used to enhance the representation of spatial relationship features. This step enables us to more accurately identify subtle changes in complex terrain. At the same time, the features of different data sources are fused through multimodal fusion technology, and feature complementarity and joint analysis are achieved using a cross-modal conversion network. The results show that the graph convolutional network module based on the attention mechanism has improved its feature representation ability by 15% in complex terrain areas.

[0250] To address the problem of unlabeled data and missing data, we use self-supervised learning methods to pre-train features on unlabeled data, and use generative adversarial networks (GANs) for data augmentation and missing data completion. Through these methods, we are able to generate high-quality data that is similar to the distribution of real data, thereby improving the robustness and recognition accuracy of the model. In practical applications, the use of generative adversarial networks has increased the diversity of the dataset by 25% and the missing data completion rate has reached 95%.

[0251] After data processing and feature extraction, we generated a high-precision geographic feature recognition map based on the classification results and spatial relationship analysis. Combined with the optimization module of the reinforcement learning algorithm, the recognition map is dynamically optimized and adaptively adjusted. Through the reinforcement learning algorithm, we can continuously optimize the accuracy and applicability of the recognition map based on the classification results and spatial relationship analysis. In actual applications, the accuracy of the reinforcement learning algorithm in optimizing the recognition map has increased by 12%, and the response time of the model adaptive adjustment has been shortened to a few minutes.

[0252] By comprehensively analyzing these geographic information data tags and combining them with the actual needs and planning goals of the city, we generated detailed planning recommendations and reports. These recommendations include traffic optimization plans, ecological protection zone demarcation, and site selection for new public facilities, which significantly improve scientificity and practicality. Through continuous learning mechanisms and transfer learning techniques, our model can be regularly updated and optimized to maintain high performance and adaptability. Throughout the project cycle, the performance of the model continued to improve, ensuring stability and adaptability in different environments and conditions.

[0253] Table 1 Data acquisition, preprocessing, feature extraction and classification, and data enhancement and completion effects

[0254]

[0255] Table 2 Application and dynamic optimization effect of graph convolutional network based on attention mechanism

[0256]

[0257] Table 3 Comprehensive analysis and model updating effects

[0258]

[0259] Through the method of the present invention, the time for data acquisition and preprocessing is significantly reduced, the classification accuracy and data quality are significantly improved, and the data enhancement and missing data completion effects are obvious; the graph convolutional network based on the attention mechanism significantly improves the feature representation ability and fusion accuracy in complex terrain areas, the accuracy after optimization is greatly improved, and the dynamic adjustment response time is significantly shortened; the frequency and efficiency of comprehensive analysis and model updating are greatly improved, the practicality is significantly improved, and the planning suggestions provided are more comprehensive and scientific.

[0260] This embodiment significantly improves the analysis accuracy and processing efficiency of urban and rural planning surveying and mapping data by introducing a variety of advanced neural network and machine learning technologies. The specific method includes: obtaining and preprocessing remote sensing images, topographic maps and vector data, constructing and training a multi-scale convolutional neural network model for feature extraction and classification, integrating a graph convolutional network module based on an attention mechanism for weighted processing and fusion analysis, and using self-supervised learning and generative adversarial networks for data enhancement and missing data completion. Through these technical means, not only the accuracy of geographic feature recognition and the consistency of data processing are improved, but also the processing time is significantly reduced and the data quality is improved. The reinforcement learning algorithm further optimizes the dynamic adjustment of the recognition graph and improves the adaptability and real-time performance of the model. Comprehensively analyze the geographic information data labels, and generate more scientific and practical planning suggestions and reports in combination with actual urban and rural planning needs. Continuous learning and transfer learning technology ensure that the model can be regularly updated and optimized to maintain high performance and adaptability. In short, the present invention has significantly improved data processing efficiency, feature recognition accuracy, data fusion and analysis capabilities, dynamic optimization and adaptive adjustment, model update and optimization, etc., providing scientific and reliable decision support for urban and rural planning.

[0261] In summary, the present invention has significant improvements in data processing efficiency, feature recognition accuracy, data fusion and analysis capabilities, dynamic optimization and adaptive adjustment, model updating and optimization, etc., providing more scientific and reliable data support for urban and rural planning, and significantly improving the scientificity and practicality of planning and the accuracy of decision-making.

[0262] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for analyzing urban and rural planning surveying and mapping data based on neural network, characterized in that: The following steps are involved: S1. Obtain remote sensing images, topographic maps and vector data, and perform preprocessing; S2. Build and train a multi-scale convolutional neural network model, optimize it through multi-level supervised learning method and image pyramid method, extract geographic features of different scales and improve recognition ability, input the preprocessed multi-source geographic information data into the multi-scale convolutional neural network model for feature extraction and classification; S3. Integrate the graph convolutional network module based on the attention mechanism, perform weighted processing on the extracted geographic features through the multi-head self-attention mechanism and the graph attention network, use the position encoding technology to enhance the representation ability of spatial relationship features, and use the multimodal fusion technology to fuse the features of different data sources, and use the cross-modal conversion network to achieve feature complementarity and joint analysis; S4. Use self-supervised learning methods to pre-train features of unlabeled data, and use generative adversarial networks to perform data enhancement and missing data completion; S5. Generate a high-precision geographic feature recognition map based on the classification results and spatial relationship analysis, and combine the optimization module of the reinforcement learning algorithm to dynamically optimize and adaptively adjust the recognition map to generate geographic information data labels; S6. Comprehensively analyze geographic information data labels, combine urban and rural planning needs and actual conditions, generate planning recommendations and reports, and use continuous learning mechanisms and transfer learning techniques to regularly update and optimize multi-scale convolutional neural network models, graph convolutional network models, and cross-modal conversion network models.

2. The method for analyzing urban and rural planning surveying and mapping data based on a neural network according to claim 1, characterized in that: The step S1 comprises the following steps: S11. Acquire remote sensing image data, topographic map data and vector data; S12. Perform data normalization on the acquired data: Where x is the original data value, x' is the normalized data value, min(x) and max(x) are the minimum and maximum values ​​in the data set respectively; S13. Perform denoising on the normalized data using a Gaussian filter: Among them, G(x,y) is the filtered value, σ is the standard deviation, and x and y are the coordinates of the pixel points; S14. Perform geometric correction and radiation correction on the denoised data, where the geometric correction is achieved through the geometric transformation matrix T: Among them, (x, y) is the image coordinate, (x', y') is the geographic coordinate, and T is the geometric transformation matrix; Radiometric correction: Among them, L is the corrected radiation brightness value, L max and L min are the maximum and minimum radiation brightness values, Q is the original radiation value, Q max and Q min are the maximum and minimum values ​​of the original radiation values ​​respectively; S15. Convert the preprocessed data into a data format, including converting vector data into raster data: Among them, R(i,j) is the value of the grid data at position (i,j), V k is the attribute value of the vector data, f k (i,j) is the rasterization function, and n is the number of vector data.

3. The method for analyzing urban and rural planning surveying and mapping data based on a neural network according to claim 1, characterized in that: The step S2 comprises the following steps: S21. Construct a multi-scale convolutional neural network model, wherein the model includes multiple convolutional layers, pooling layers and fully connected layers; S22. Design a convolutional layer, where the size and number of convolution kernels are k×k and n respectively, and the output feature map of the convolutional layer is expressed as: Among them, F l (i, j) is the input feature map of the lth layer at position (i, j), W l (p,q) is the convolution kernel of the lth layer at position (p,q), b l is the bias, f is the activation function; S23. Design the pooling layer and use the maximum pooling operation with a pooling window size of m×m: P i,j =max 0≤p<m,0≤q<m F i+p,j+q ; Among them, P i,j is the pooling result, F i+p,j+q is the feature value within the pooling window; S24. The multi-scale convolutional neural network model is trained by a multi-level supervised learning method, wherein the training process includes forward propagation and back propagation, wherein the loss function is a cross entropy loss function: Where N is the number of samples, y i is the true label, is the predicted label; S25. Use the image pyramid method to perform multi-scale processing on the input remote sensing image data and generate image pyramids of different resolutions: Among them, I s is the s-th layer pyramid image, I original is the original image, s is the number of pyramid layers; S26. Input the generated multi-scale image pyramid into the multi-scale convolutional neural network model for feature extraction: F s,l (i,j)=MSCNN(I s (i,j)); Among them, F s,l (i, j) is the value of the feature map of the s-th layer pyramid image at position (i, j) in the l-th layer convolutional neural network; S27. Optimize the multi-scale convolutional neural network model using the Adam optimization algorithm. The parameter update process includes the following steps: Among them, m t and v t are the first-order and second-order momentum estimates at time t, β1 and β2 are the decay rates of the momentum term, is the loss function L t For the model parameters θ t The gradient of Perform bias correction on the momentum and compute the bias-corrected momentum estimate: Update model parameters: Among them, η is the learning rate and is the smoothing term; S28. Input the preprocessed multi-source geographic information data into the multi-scale convolutional neural network model for feature extraction and classification: F output (i,j)=MSCNN(I input (i,j)); Among them, F output (i, j) is the output feature of the input data at position (i, j), I input (i,j) is the preprocessed input data.

4. The method for analyzing urban and rural planning surveying and mapping data based on a neural network according to claim 1, characterized in that: The step S3 comprises the following steps: S31. Construct a graph convolution network module based on the attention mechanism, which includes a graph convolution layer and a multi-head self-attention mechanism, and processes the multi-scale features of the input through the graph convolution layer: H (l+1) =σ(D -1 / 2 AD -1 / 2 H (l) W (l) ); Among them, H (l) is the input feature matrix of the lth layer, A is the adjacency matrix, D is the degree matrix, W (l) is the weight matrix of the lth layer, σ is the activation function; S32. Use the multi-head self-attention mechanism to weight the output features of the graph convolution layer: Among them, Q is the query matrix, K is the key matrix, and V is the value matrix. and They are the weight matrices of the i-th attention head: Among them, d k is the dimension of the key matrix, h is the number of attention heads; S33. Use position coding technology to enhance the representation ability of spatial relationship features. The calculation formula of position coding is: Combining the weighted features with the positional encoding, the enhanced features are: F enhanced =Z+PE; Among them, F enhanced is the enhanced feature, Z is the weighted feature, and Z is the position code; S34. Use multimodal fusion technology to fuse the features of different data sources; S35. Using cross-modal conversion networks to achieve feature complementarity and joint analysis.

5. The method for analyzing urban and rural planning surveying and mapping data based on a neural network according to claim 1, characterized in that: The step S4 comprises the following steps: S41. Use the self-supervised learning method to pre-train the features of unlabeled data, adopt the contrastive learning strategy, construct positive and negative sample pairs, and the contrastive learning loss function is: Where N is the number of samples, x i is the input sample, is a positive sample, is a negative sample, y i is the sample pair label, m is the boundary distance, and f is the feature extraction function; S42. Data enhancement and missing data completion are performed through generative adversarial networks. The generative adversarial network consists of a generator G and a discriminator D. The objective function of the generative adversarial network is: Among them, G is the generator, D is the discriminator, x is the real data, z is the noise vector, p data (x) is the real data distribution, p z (z) is the noise distribution; S43. Train the generator G to generate samples similar to the real data distribution: Among them, θ G is the parameter of the generator, η G is the learning rate, m is the batch size, z i is the noise vector; S44. Train the discriminator D to distinguish between real data and generated data: Among them, θ D is the parameter of the discriminator, η D is the learning rate, x i is the real data sample, z i is the noise vector; S45. Use data generated by generative adversarial networks for data augmentation and missing data completion.

6. The method for analyzing urban and rural planning surveying and mapping data based on a neural network according to claim 1, characterized in that: The step S5 comprises the following steps: S51. Based on the classification results and spatial relationship analysis, a preliminary geographic feature recognition map is generated, and a convolutional neural network is used to perform spatial relationship analysis on the classification results: Among them, R(i,j) is the spatial relationship feature value of position (i,j), W(k,l) is the convolution kernel weight, C(i+k,j+l) is the classification result of position (i+k,j+l), and m and n are the height and width of the convolution kernel respectively; S52. Use the reinforcement learning algorithm to dynamically optimize and adaptively adjust the preliminary geographic feature recognition map, define the state s, action a and reward r, and the optimization goal is to maximize the cumulative reward: Among them, π * (s) is the optimal strategy, γ is the discount factor, r t is the reward at time t, s is the state, π is the strategy, and T is the number of time steps; S53. Using deep networks for strategy optimization in the framework of reinforcement learning: Q(s t ,a t )←Q(s t ,a t )+α[r t +γmax a' Q(s t+1 ,a')-Q(s t ,a t )]; Among them, Q(s t ,a t ) is state s t and action a t Q value, α is the learning rate, γ is the discount factor, r t is the reward value, s t+1 is the next state, a' is the next action; S54. Combined with the generative adversarial network, the recognition graph is further optimized. The objective function of the generative adversarial network is: Among them, G is the generator, D is the discriminator, x is the real data, z is the noise vector, p data (x) is the real data distribution, p z (z) is the noise distribution; S55. Combine the optimized recognition map with multi-source geographic information data to generate a high-precision geographic feature recognition map, and use a comprehensive analysis algorithm to evaluate and correct the results: F final (i,j)=α·F opt (i,j)+(1-α)·G info (i,j); Among them, F final (i, j) is the final eigenvalue at position (i, j), F opt (i,j) is the optimized eigenvalue, G info (i, j) is the characteristic value of multi-source geographic information data, and α is the weight coefficient.

7. The method for analyzing urban and rural planning surveying and mapping data based on a neural network according to claim 1, characterized in that: The step S6 comprises the following steps: S61. Comprehensively analyze geographic information data labels, conduct joint analysis on fused features, and use a multi-task learning framework: L total =αL classification +βL regression +γL segmentation ; Among them, L classification is the classification loss, L regression is the regression loss, L segmentation is the image segmentation loss, α, β and γ are weight coefficients; S62. Combine the urban and rural planning needs and actual conditions to comprehensively process the analysis results and generate preliminary planning recommendations and reports; S63. Use the reinforcement learning algorithm to optimize the preliminary planning suggestions, define the state s, action a and reward r, and the optimization goal is to maximize the cumulative reward: Among them, π * (s) is the optimal strategy, γ is the discount factor, r t is the reward at time t, s is the state, π is the strategy, and T is the number of time steps; S64. Generate a final planning report based on the optimized planning suggestions, and provide specific planning solutions and implementation suggestions in combination with urban and rural planning needs and actual conditions; S65. Use continuous learning mechanisms and transfer learning techniques to regularly update and optimize multi-scale convolutional neural network models, graph convolutional network models, and cross-modal conversion network models: Among them, L new is the loss of the new task, λ is the regularization parameter, F i is the Fisher information matrix, θ i are the current model parameters, are the model parameters of the old task; S66. In the transfer learning framework, adversarial training is used for feature alignment. The feature extraction of the source domain and the target domain are f s (x s ) and f t (x t ): Among them, Dist is the feature distribution distance function, Loss t is the target domain loss function, α and β are weight coefficients.

Citation Information

Cited By

  • Alteration zone detection and analysis method and system

    CN119312705A

  • AI-based surveying and mapping data processing system

    CN119377867A

  • Intelligent geographic surveying and mapping data processing and analyzing method and system

    CN119397482A

  • X-ray security inspection machine picture object detection method

    CN119600389A

  • Camera pollutant detection system and method for automobile

    CN119992505A