Ground sign generation method and system based on airport regulations AIP map

Abstract

The present invention belongs to the field of image data processing technology, and relates to a method and system for generating ground identification based on the airport regulations AIP map, aiming to solve the problems of incomplete information, poor adaptability, and insufficient compliance in existing identification generation methods. The present invention comprises: obtaining an ICAO rule vector; constructing a knowledge graph based on the identification information in the airport regulations AIP map, and obtaining the symbol features of the AIP map; performing cross-modal feature alignment on the satellite image and the symbol features of the AIP map, extracting the multi-scale features of the satellite image, combining the feature-enhanced AIP map symbol features, constructing a three-layer feature fusion network, and obtaining weighted comprehensive features; inputting the comprehensive features and the ICAO rule vector into a conditional GAN ​​generator to generate an identification image. The present invention deeply fuses three types of heterogeneous data: satellite images, AIP maps, and ICAO rules, with higher accuracy and adaptability to different standards and environments.

Description

Technical Field

[0001] The present invention belongs to the technical field of image data processing, and in particular relates to a method and system for generating ground markings based on an airport specification AIP diagram. Background Art

[0002] In the fields of aviation and transportation, airport ground signs, as a key component of the visual navigation aids system, provide precise guidance, risk warnings and information transmission for aircraft taxiing and ground vehicle operations through standardized symbols, color systems and text codes. They are the infrastructure to ensure the safety of airport ground operations.

[0003] In aviation flight simulation training, millimeter-level accurate reproduction of ground markings is a core prerequisite for creating immersive training scenarios. However, intelligent recognition and generation technologies face multiple technical bottlenecks. Conventional mainstream methods for generating ground markings suffer from the following issues: First, reliance on single-modal data (e.g., using only satellite remote sensing imagery or AIP text data) results in incomplete information acquisition, making it difficult to balance the semantic content of the markings with their spatial form. This often leads to issues such as semantic misalignment of runway numbers and geometric deviations of markings. Second, traditional algorithms lack adaptability to complex scenarios. The stability of marking feature extraction significantly decreases in environments with undulating terrain, vegetation obstruction, and changing light and shadow. For example, the accuracy of runway edge detection at mountainous airports is lower than that at plain airports. Furthermore, existing technologies lack deep integration with industry standards such as those of the International Civil Aviation Organization (ICAO), resulting in potential compliance risks for generated ground markings, such as color coding violations (e.g., runway numbers not being white) and substandard dimensions (e.g., taxiway widths being less than the standard).

[0004] Therefore, there is an urgent need for an intelligent ground sign generation technology that integrates multi-source heterogeneous data fusion, adapts to complex scenarios and meets industry regulatory constraints to meet the actual needs of aviation safety operations and digital traffic management. Summary of the Invention

[0005] In order to solve the above-mentioned problems in the prior art, namely, the problems that the existing identification generation methods lack information integrity, have poor adaptability to complex scenarios, and do not consider industry standards, the first aspect of the present invention proposes a ground identification generation method based on the airport detailed information map (AIP), the method comprising the following steps:

[0006] S1. Obtain ICAO rule vector based on pre-built dynamic semantic knowledge base;

[0007] S2. Based on the identification information in the airport regulations AIP graph and the ICAO rule vector, a knowledge graph is constructed and semantically embedded to obtain the AIP graph symbol features. The node attributes of the knowledge graph include ICAO standard codes, geometric parameters, and topological relationships.

[0008] S3. Obtain satellite images, perform cross-modal feature alignment between the satellite images and AIP map symbol features, and enhance the features based on a bidirectional attention mechanism;

[0009] S4. Extract the multi-scale features of the satellite image after feature enhancement, combine them with the symbol features of the AIP map after feature enhancement, build a three-layer feature fusion network, determine the dynamic weight of each layer of features according to the identification type, and obtain the weighted comprehensive features;

[0010] S5. Splicing the comprehensive features with the ICAO rule vector as input, generating a logo image based on a conditional GAN ​​generator, and performing geometric and topological checks on the generated results.

[0011] In some preferred embodiments, the ICAO rule vector is obtained by:

[0012] According to the current airport type, the corresponding ICAO rule vector is queried in real time in a pre-built dynamic semantic knowledge base; wherein, the dynamic semantic knowledge base is constructed by:

[0013] The ICAO rules are obtained and semantically parsed to obtain the semantic information and logical relationships corresponding to each semantic rule. The semantic rules are vector-encoded based on the knowledge graph embedding algorithm to obtain the ICAO rule vectors and form a dynamic semantic knowledge base.

[0014] In some preferred embodiments, the AIP diagram symbol features are obtained by:

[0015] Extract the identification information from the airport regulations AIP diagram, transform the identification information into a triple knowledge graph through named entity recognition and relationship extraction technology, use the TransE algorithm to map entities and relationships into low-dimensional vectors, and obtain a geometric-topological feature vector containing geometric positions and topological relationships. The geometric-topological feature vector is then combined with the ICAO rule vector to form the AIP diagram symbol feature. .

[0016] In some preferred embodiments, cross-modal feature alignment is performed on satellite images and AIP map symbol features by:

[0017] A spatial transformer network is used to geometrically align satellite images with AIP maps:

[0018] Constructing a cross-modal projection matrix 、 , using contrastive learning loss function, the AIP graph symbol features Satellite image features Map to a shared semantic space for semantic alignment:

[0019] ;

[0020] Contrastive Learning Loss Function for:

[0021] ;

[0022] in, 、 is the eigenvector after projection matrix mapping, The first i Satellite image features of negative sample feature vectors, N is the number of samples, is the temperature parameter, They are shared space dimension, AIP map symbol feature original dimension, and satellite feature original dimension respectively; is a function that calculates the similarity between two feature vectors.

[0023] In some preferred embodiments, the feature is enhanced based on the bidirectional attention mechanism, and the method is as follows:

[0024] Use AIP map symbol features as the query and satellite image features as the key and value:

[0025] ;

[0026] Using satellite image features as queries and AIP map symbol features as keys and values:

[0027] ;

[0028] in, is the feature dimension; Calculate the function for the attention weight between modal features;

[0029] The bidirectional attention weights are fused with the original features to generate enhanced features:

[0030] ;

[0031] ;

[0032] in, is the scaling factor, which is dynamically adjusted through back-propagation.

[0033] In some preferred embodiments, after obtaining the attention-enhanced features, the offset is learned using deformable convolution to dynamically adjust the sampling point position, and the method is as follows:

[0034] ;

[0035] in, represents feature splicing, is the offset of the deformable convolution, is the modulation weight;

[0036] Integrate the learning objectives of feature alignment, bidirectional attention, and deformable convolution into a unified loss function L middle:

[0037] ;

[0038] ;

[0039] ;

[0040] in, is the classification loss; is the alignment loss; is the offset loss, is the actual marked offset, is the weight of the loss.

[0041] In some preferred embodiments, the comprehensive features are generated by:

[0042] ResNet-50 extracts multi-scale features of satellite images after feature enhancement, including bottom-level detail features, mid-level shape features, and high-level semantic features;

[0043] Based on the enhanced AIP symbol features and ICAO rule vectors, multi-scale features are hierarchically fused to obtain bottom-level features, middle-level features, and high-level features.

[0044] The bottom-level features, middle-level features and high-level features are fused from top to bottom to generate fused features of each layer, and the weights of the fused features of each layer are determined according to the identification type, and the weighted comprehensive features are obtained.

[0045] In some preferred embodiments, the comprehensive features are obtained by:

[0046] After upsampling the high-level features, add them to the middle-level features, and strengthen cross-layer interactions through element-by-element multiplication. Then, upsample the fusion result and fuse it with the bottom-level features:

[0047] ;

[0048] in, is the upsampling operation function, is the interaction coefficient, Represents element-wise multiplication; For the i The lateral input features of the layer, For the current i The fusion features of the layers, is the fusion feature of the previous layer;

[0049] Determine dynamic weight based on identification type :

[0050] ;

[0051] Perform weighted fusion on the fused features to obtain comprehensive features :

[0052] ;

[0053] in, is the weight of different identification types, which is dynamically adjusted according to importance. MLP is a multi-layer perceptron network. is an exponential function, and N is the number of layers.

[0054] In some preferred embodiments, a logo image is generated based on a conditional GAN ​​generator by:

[0055] The comprehensive features are concatenated with the ICAO rule vector as the input of the conditional GAN ​​generator. x :

[0056] ;

[0057] in, r ICAO rule vectors are queried from the dynamic semantic knowledge base in real time according to the current airport type;

[0058] The semantic loss for generating the logo is:

[0059] ;

[0060] in, is the L2 norm, which is used to measure the Euclidean distance between two vectors; For the conditional GAN ​​generator, the input is mapped to the generated identity, A semantic embedding function for generating logos;

[0061] The loss function is defined as:

[0062] ;

[0063] in, To combat losses, is the weight coefficient.

[0064] A second aspect of the present invention provides a ground sign generation system based on an airport information plan (AIP) diagram, the system comprising:

[0065] The data preprocessing module is configured to acquire satellite images and perform preprocessing; construct a knowledge graph based on the identification information in the airport specification AIP diagram and the ICAO rule vector and perform semantic embedding to obtain the AIP diagram symbol features; and obtain the ICAO rule vector;

[0066] A feature alignment module is configured to perform cross-modal feature alignment on the pre-processed satellite imagery and the AIP map symbol features, and enhance the features based on a bidirectional attention mechanism;

[0067] The feature fusion module is configured to extract multi-scale features of the satellite image after feature enhancement, combine them with the symbolic features of the AIP map after feature enhancement, and generate comprehensive features through dynamic weight aggregation to generate comprehensive features;

[0068] The logo generation and verification module is configured to use the combined comprehensive features and the ICAO rule vector as input, generate a logo image based on a conditional GAN ​​generator, and perform geometric verification and topological verification on the generated result.

[0069] Beneficial effects of the present invention:

[0070] 1. This paper deeply fuses three types of heterogeneous data: satellite imagery, AIP maps, and ICAO rules. It constructs a knowledge graph through structured processing of the AIP map, combines it with STN geometric alignment of satellite images, and uses bidirectional attention to achieve cross-modal feature alignment. After hierarchical fusion and ICAO rule vector constraints, it finally generates ICAO-compliant ground signs through cGAN. This forms a complete closed loop from data preprocessing to semantic verification, breaking through the limitations of traditional single data sources and adapting to different airport layouts and signage standards.

[0071] 2. To address the core challenge of balancing semantic consistency and geometric accuracy in ground sign generation, we use ICAO rules and AIP text for semantic guidance, combine the STN network, contrastive learning and bidirectional attention mechanism, achieve geometric compensation through deformable convolution, and use three-stage feature fusion to process information at different scales, achieving dual alignment of geometry and semantics, thereby improving recognition accuracy.

[0072] 3. Construct a three-layer feature fusion network. The bottom layer focuses on pixel-level details, the middle layer extracts shape features, and the high layer captures global semantics. Through top-down fusion and a dynamic weight mechanism, it effectively integrates multi-scale information. The dynamic weight can be adaptively adjusted according to the logo type to optimize computational efficiency and generation quality. Compared with traditional architectures with fixed weights, it has higher flexibility and adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Other features, objects and advantages of the present application will become more apparent upon reading the detailed description of non-limiting embodiments made with reference to the following drawings:

[0074] Figure 1 The present invention is a flowchart of a method for generating ground signs based on an airport specification AIP diagram in an embodiment of the present invention. DETAILED DESCRIPTION

[0075] The present application will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are intended only to illustrate the relevant invention and are not intended to limit the invention. It should also be noted that, for ease of description, only portions relevant to the relevant invention are shown in the accompanying drawings.

[0076] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0077] The present invention provides a method for generating ground signs based on airport regulations (AIP) charts. This method constructs a three-stage fusion network by fusing multi-source data such as satellite images, airport regulations (AIP) charts, and ICAO rules. Combined with top-down pyramid fusion, this method achieves multi-scale information integration of "details-shape-semantics." Based on deep feature interaction and semantic association mechanisms, the modal barriers between AIP data and satellite images are broken to achieve deep fusion of information, thereby realizing high-precision, automated, and industry-standard-compliant ground sign generation, solving the problems of incomplete information, poor adaptability, and insufficient compliance in the existing technology.

[0078] The present invention provides a method for generating ground markings based on an airport information plan (AIP) diagram, comprising the following steps:

[0079] S1. Obtain ICAO rule vector based on pre-built dynamic semantic knowledge base;

[0080] S2. Based on the identification information in the airport regulations AIP diagram and the ICAO rule vector, a knowledge graph is constructed and semantically embedded to obtain AIP features. The node attributes of the knowledge graph include ICAO standard codes, geometric parameters, and topological relationships.

[0081] S3. Obtain satellite images, perform cross-modal feature alignment between satellite images and AIP features, and enhance features based on a bidirectional attention mechanism.

[0082] S4. Extract the multi-scale features of the satellite image after feature enhancement, combine them with the symbol features of the AIP map after feature enhancement, build a three-layer feature fusion network, determine the dynamic weight of each layer of features according to the identification type, and obtain the weighted comprehensive features;

[0083] S5. Splicing the comprehensive features with the ICAO rule vector as input, generating a logo image based on a conditional GAN ​​generator, and performing geometric and topological checks on the generated results.

[0084] In order to more clearly illustrate the method for generating ground signs based on the airport regulations AIP map of the present invention, the following is combined with Figure 1 Each step in the embodiment of the present invention is described in detail.

[0085] The method for generating ground signs based on an airport information plan (AIP) diagram according to the first embodiment of the present invention includes steps S1 to S5, each of which is described in detail as follows:

[0086] S1. Based on the pre-built dynamic semantic knowledge base, the ICAO rule vector is obtained as follows:

[0087] Based on the current airport type, the corresponding ICAO rule vector is queried in real time in the pre-built dynamic semantic knowledge base;

[0088] Based on the airport type (hub / feeder / general) and sign type (runway / taxiway / holding line), the most matching rule vector set is retrieved through cosine similarity, and a weighted fusion of multiple rule vectors is performed using an attention mechanism. For example, runway sign generation must satisfy both width and clearance rules.

[0089] The dynamic semantic knowledge base is constructed as follows:

[0090] The ICAO rules are obtained and semantically parsed to obtain the semantic information and logical relationships corresponding to each semantic rule. The semantic rules are vector-encoded based on the knowledge graph embedding algorithm to obtain the ICAO rule vectors and form a dynamic semantic knowledge base.

[0091] S2. Based on the identification information in the airport regulations AIP diagram and the ICAO rule vector, a knowledge graph is constructed and semantic embedding is performed to obtain the AIP graph symbol features. The node attributes of the knowledge graph include ICAO standard codes, geometric parameters, and topological relationships.

[0092] Preferably, the AIP diagram symbol features are obtained by:

[0093] Extract identification information from the airport specification (AIP) diagram and convert it into a triplet knowledge graph using named entity recognition and relationship extraction techniques. Use the TransE algorithm to map entities and relationships into low-dimensional vectors, generating a geometric-topological feature vector containing geometric positions and topological relationships (existing technology).

[0094] The geometric-topological feature vector and the ICAO rule vector are combined to form the AIP diagram symbol feature. .

[0095] Further preferably, in this embodiment, the node attributes of the triple knowledge graph include ICAO standard codes, geometric parameters (e.g., length, width), and topological relationships (e.g., connectivity, subordination). The graphical information of the AIP graph provides the foundation for constructing the knowledge graph. The lines and identifiers in the graph are converted into node-edge relationships and stored in a graph database (e.g., Neo4j) for easy subsequent processing and analysis.

[0096] S3. Obtain satellite images, perform cross-modal feature alignment between satellite images and AIP map symbol features, and enhance features based on a bidirectional attention mechanism.

[0097] Preferably, the satellite image is obtained and preprocessed by:

[0098] Based on the characteristics of airport scenes, visible light and near-infrared satellite imagery are integrated, and band differences are used to enhance the distinction between artificial facilities such as runways and aprons and natural features;

[0099] Generative adversarial networks (GANs) are used for image restoration to eliminate cloud occlusions and shadows in satellite images and improve the integrity of identification features.

[0100] Preferably, cross-modal feature alignment is performed on satellite images and AIP map symbol features. By learning the geometric and semantic transformation relationship between cross-modal features, the feature position and expression form are dynamically adjusted. The method is as follows:

[0101] The spatial transformer network STN (Spatial Transformer Network) is used to geometrically align the visual features of satellite images with the symbol features of AIP maps:

[0102] ;

[0103] in, is the pixel coordinate of the satellite image, To align the target coordinates to the AIP coordinate system, is the perspective transformation parameter (learned by STN network);

[0104] Constructing a cross-modal projection matrix 、 , the visual features of the aligned satellite images are extracted through the convolutional neural network CNN network , using contrastive learning loss function, the AIP graph symbol features Satellite image features Map to a shared semantic space for semantic alignment:

[0105] ;

[0106] ;

[0107] Contrastive Learning Loss Function for:

[0108] ;

[0109] in, 、 is the eigenvector after projection matrix mapping, The first i Satellite image features of negative sample feature vectors, They are shared space dimension, AIPIP map symbol feature original dimension, and satellite feature original dimension respectively; is a function for calculating the similarity of two feature vectors; is the temperature parameter, which controls the feature clustering density so that the cross-modal features of the same type are closer in the shared space than the heterogeneous features; the cross-modal projection matrix 、 is the linear transformation matrix, The AIP graph features (dimension ) are projected into a shared space (dimension D), Satellite image features (dimension ) are projected into a shared space (dimension D).

[0110] Furthermore, if the dimension is too low, semantic information will be lost (e.g., D=32), while if it is too high, computational cost will be increased (e.g., D=1024). Usually, D=256-512 is selected according to task complexity. In this embodiment, the vector dimension is The value is 256-dimensional, vector dimension The value is 512 dimensions, the vector dimension D is 256 dimensions, and the semantic rule dimension is 128 dimensions.

[0111] Adjustment through backpropagation and Parameters, so Minimize. When the loss converges, the similarity of positive sample pairs (such as AIP "runway" and satellite runway area) Maximize; minimize the similarity of negative sample pairs (such as AIP "runway" and satellite vegetation area), select N negative samples for each training to enhance the discrimination of contrastive learning.

[0112] In this embodiment, negative samples are obtained by existing technologies, such as hard example mining ( Hard Negative Mining , select negative samples that are semantically similar to positive samples but do not match their geometry, which increases the difficulty of contrastive learning.

[0113] The geographic coordinates of the satellite image are projected onto the AIP plane coordinates so that the two can be compared under the same spatial reference, solving the global spatial misalignment and providing a geometric reference for subsequent deformable convolution. Contrastive learning loss is introduced for semantic projection to establish a correspondence between abstract semantic symbols in the AIP image (such as the "waiting line" text) and specific visual features in the satellite image (the yellow line segment representing the waiting line). This allows cross-modal features of similar logos to be clustered in a shared space, providing a feature basis for semantic alignment for bidirectional attention.

[0114] Further preferably, the feature is enhanced based on the bidirectional attention mechanism, and the method is as follows:

[0115] AIP → Satellite Imagery: Guide the model to focus on areas in the satellite image that are semantically relevant to AIP, using the AIP image symbol features as queries and the satellite image features as keys and values:

[0116] ;

[0117] Satellite image → AIP: Feedback AIP semantic understanding through visual features, using satellite image features as queries and AIP symbol features as keys and values ​​to correct fuzzy parameters in AIP:

[0118] ;

[0119] in, is the feature dimension; Calculate the function for the attention weight between modal features;

[0120] The bidirectional attention weights are fused with the original features to generate enhanced features:

[0121] ;

[0122] ;

[0123] in, is the scaling factor, which is dynamically adjusted through back-propagation.

[0124] Preferably, in the bidirectional interaction between satellite image features and AIP semantic features, the ICAO rule vector is used as an additional query input and a dynamic control signal to guide the visual features to align to the compliance area.

[0125] Further preferably, after obtaining the attention-enhanced features, the offset is learned using deformable convolution to dynamically adjust the sampling point position of the satellite image, and the method is as follows:

[0126] Offset prediction:

[0127] ;

[0128] in, represents feature splicing, is the offset of the deformable convolution, is the modulation weight; Working in synergy with the attention mechanism, high attention areas correspond to Δm The value is larger when When it is close to 0, it suppresses invalid areas (such as grass), and when it is close to 1, it strengthens the marked area;

[0129] Integrate the learning objectives of feature alignment, bidirectional attention, and deformable convolution into a unified loss function L middle:

[0130] ;

[0131] ;

[0132] ;

[0133] in, is the classification loss; is the alignment loss; is the offset loss, is the actual marked offset, is the weight of the loss.

[0134] The bidirectional attention mechanism has achieved semantic feature enhancement through cross-modal feature interaction, and the generated features and Since semantic association information is already included, deformable convolution is introduced at this time to compensate for feature space misalignment from a geometric level and avoid interference from invalid areas. The offset-compensated features are closer to the true identification form, so that the loss function constrains the consistency of the offset and the true deformation, which can optimize the model more accurately.

[0135] In this embodiment, for example, the text "Runway 18" in the AIP is used as a query, and the corresponding runway area is located in the satellite image. The attention weight enhances the features of the area. However, the runway in the satellite image is trapezoidal due to the inclination of the aerial photography, which is geometrically different from the rectangular definition in the AIP. Based on the features after attention enhancement, the offset is calculated to shift the sampling points on the edge of the trapezoid toward the rectangular outline, and the weight is modulated at the same time. Δm The characteristic responses of non-runway areas (such as the grass on both sides of the runway) are suppressed, and finally a rectangular runway marking that meets the AIP standard is generated.

[0136] S4. Extract the multi-scale features of the satellite image after feature enhancement, combine them with the symbol features of the AIP map after feature enhancement, build a three-layer feature fusion network, determine the dynamic weight according to the identification type, calculate the feature weight of each layer, and obtain the comprehensive features.

[0137] Preferably, the comprehensive features are generated by:

[0138] ResNet-50 extracts multi-scale features of satellite images after feature enhancement, including bottom-level detail features, mid-level shape features, and high-level semantic features;

[0139] Based on the enhanced AIP symbol features and ICAO rule vectors, multi-scale features are hierarchically fused to obtain bottom-level features, middle-level features, and high-level features.

[0140] The bottom-level features, middle-level features and high-level features are fused from top to bottom to generate fused features of each layer, and the weights of the fused features of each layer are determined according to the identification type, and the weighted comprehensive features are obtained.

[0141] Preferably, the bottom-layer network (Conv2) extracts pixel-level details of the sign, such as edges and textures, and preserves the continuity of the sign lines; the middle-layer network (Conv4) extracts shape features of the sign, such as rectangles and curves, and identifies the geometric contours of runways and parking spaces; the high-layer network (Conv5) captures global features of the satellite image, such as sign category, scene context, and topological relationships, such as the connection between "runway number" and "taxiway."

[0142] Further preferably, based on the feature-enhanced AIP diagram symbol features and ICAO rule vectors, multi-scale features are hierarchically fused to obtain bottom-level features, middle-level features, and high-level features, and the method is as follows:

[0143] After converting the vectorized line information (such as the wait line coordinates) in the AIP map symbol features into a binary image, it is concatenated with the satellite image features along the channel dimension using the detail features extracted in the underlying network of the hierarchical feature fusion architecture. The concatenated features are then input into the convolutional layer to reduce the dimensionality of the concatenated features, generating a bottom-level feature map that contains the AIP map coordinate accuracy and the satellite image edge details as the bottom-level features.

[0144] The shape features identified by the middle-level network are fused with the shape priors in the symbolic features of the AIP graph. Through the Transformer multi-head attention mechanism, the AIP graph shape is used as the query and the satellite image features as the key value to strengthen the feature response of the corresponding shape and generate a shape feature map with semantic guidance as the middle-level feature.

[0145] The global semantic features extracted by the high-level network are compressed into a vector, and then concatenated with the topological relationship vector and symbolic semantic vector in the AIP graph symbolic features to generate a high-level feature graph containing global semantics and topological constraints as a preliminary fused feature vector. The preliminary fused feature vector is interacted with the ICAO rule vector through a fully connected layer, and the concatenated vector is mapped to the semantic space to obtain high-level features.

[0146] Among them, the ICAO rule vector serves as a dynamic fusion weight, adjusting the weights of other features through the fully connected layer. For example, when the rule vector emphasizes "prohibited obstacles", the feature response of the corresponding area in the satellite image is suppressed.

[0147] Further preferably, the dynamic weight is determined according to the identification type to calculate the feature weight of each layer to obtain the comprehensive feature, and the method is as follows:

[0148] After upsampling the high-level features, add them to the middle-level features, strengthen cross-layer interaction through element-by-element multiplication, and then upsample the fusion result and fuse it with the bottom-level features:

[0149] ;

[0150] in, is the upsampling operation function, is the interaction coefficient, Represents element-wise multiplication; is the lateral input feature of the i-th layer, is the fusion feature of the current i-th layer, is the fusion feature of the previous layer;

[0151] Determine dynamic weight based on identification type :

[0152] ;

[0153] in, The weights of different identification types (such as the feature preferences of runways and taxiways) are dynamically adjusted according to their importance. MLP is a multi-layer perceptron network. is an exponential function;

[0154] Perform weighted fusion on the fused features to obtain comprehensive features :

[0155] ;

[0156] Among them, N=3.

[0157] A three-layer feature fusion network is constructed, and three-stage feature fusion is used to process information of different scales. The bottom layer focuses on pixel-level details, the middle layer extracts shape features, and the high layer captures global semantics. Through top-down fusion and dynamic weight mechanism, effective integration of multi-scale information is achieved, breaking the single mode of traditional feature processing. Dynamic weights can be adaptively adjusted according to the identification type to optimize computing efficiency and generation quality. Compared with the traditional architecture with fixed weights, it has higher flexibility and adaptability.

[0158] S5. Splicing the comprehensive features with the ICAO rule vector as input, generating a logo image based on a conditional GAN ​​generator, and performing geometric and topological checks on the generated results.

[0159] Preferably, the identification image is generated based on the conditional GAN ​​generator, and the method is as follows:

[0160] The comprehensive features are concatenated with the ICAO rule vector as the input of the conditional GAN ​​generator. x :

[0161] ;

[0162] in, r ICAO rule vectors are used as mandatory constraints to ensure that the results do not violate the specifications. Based on the current airport type, ICAO rule vectors are queried from the dynamic semantic knowledge base in real time;

[0163] The semantic loss for generating the logo is:

[0164] ;

[0165] in, is the L2 norm, which is used to measure the Euclidean distance between two vectors; For the conditional GAN ​​generator, the input is mapped to the generated identity, A semantic embedding function for generating logos;

[0166] The loss function is defined as:

[0167] ;

[0168] in, To combat losses, is the weight coefficient.

[0169] Preferably, considering accuracy, a multi-scale reconstruction loss can be added to the loss. By using layered constraints, the generated image is aligned with the real data in the three dimensions of pixels, features, and semantics, addressing the "fuzzification" and "semantic drift" problems in traditional generative models. Specifically, in this embodiment, the multi-scale reconstruction loss includes pixel-level loss, feature-level loss, and rule-level loss. The combination of these three prevents the generative model from only memorizing local patterns in the training data, thereby ensuring pixel authenticity, semantic consistency, and structural rationality.

[0170] By building an ICAO dynamic knowledge base, the fusion features and ICAO rule vectors are spliced ​​and input into the cGAN generator. The semantic compliance of the constraint generation results is achieved to achieve dual protection of visual feature-driven generation and semantic verification of industry rules, ensuring that the generated ground signs comply with the visual characteristics of satellite images, reduce the computing resource consumption of unnecessary information, and eliminate invalid solutions that do not meet civil aviation standards.

[0171] The method of the present invention deeply integrates three types of heterogeneous data: satellite images, AIP texts and ICAO rules; the knowledge graph and semantic guidance mechanism adopted make the model decision-making process traceable. When local changes occur in the airport, the local features of the AIP data and satellite images in the changed area are updated to facilitate manual verification and rule updates, achieving significant improvements in accuracy, robustness and intelligence, and providing more reliable technical support for the digital construction of airports.

[0172] Although the various steps in the above embodiment are described in the above-mentioned order, those skilled in the art will understand that in order to achieve the effect of this embodiment, different steps do not have to be executed in such an order. They can be executed simultaneously (in parallel) or in a reverse order. These simple changes are within the scope of protection of the present invention.

[0173] A second embodiment of the present invention provides a ground sign generation system based on an airport information plan (AIP) diagram, the system comprising:

[0174] The data preprocessing module is configured to acquire satellite images and perform preprocessing; construct a knowledge graph based on the identification information in the airport specification AIP diagram and the ICAO rule vector and perform semantic embedding to obtain the AIP diagram symbol features; and obtain the ICAO rule vector;

[0175] A feature alignment module is configured to perform cross-modal feature alignment on the pre-processed satellite imagery and the AIP map symbol features, and enhance the features based on a bidirectional attention mechanism;

[0176] The feature fusion module is configured to extract multi-scale features of the satellite image after feature enhancement, combine them with the symbolic features of the AIP map after feature enhancement, and generate comprehensive features through dynamic weight aggregation to generate comprehensive features;

[0177] The logo generation and verification module is configured to use the combined comprehensive features and the ICAO rule vector as input, generate a logo image based on a conditional GAN ​​generator, and perform geometric verification and topological verification on the generated result.

[0178] It should be noted that the ground sign generation system based on the airport regulations AIP diagram provided in the above embodiment is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of the present invention can be decomposed or combined. For example, the modules in the above embodiment can be combined into one module, or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing the modules or steps and are not regarded as improper limitations on the present invention.

[0179] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working process and related instructions of the system described above can refer to the corresponding process in the aforementioned method embodiment and will not be repeated here.

[0180] An electronic device according to a third embodiment of the present invention includes:

[0181] at least one processor; and

[0182] a memory communicatively connected to at least one of the processors; wherein,

[0183] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the above-mentioned ground identification generation method based on the airport information plan AIP chart.

[0184] A fourth embodiment of the present invention provides a computer-readable storage medium storing computer instructions, wherein the computer instructions are used to be executed by a computer to implement the above-mentioned method for generating ground markings based on an airport specification AIP chart.

[0185] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes and related instructions of the electronic device and computer-readable storage medium described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0186] Those skilled in the art should be able to appreciate that the modules and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two, and the programs corresponding to the software modules and method steps can be placed in random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art. In order to clearly illustrate the interchangeability of electronic hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0187] Computer program code for performing the operations of the present application may be written in one or more programming languages, or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider).

[0188] The flow charts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present application. In this regard, each box in the flow chart or block diagram can represent a module, program segment or a part of code, and the module, program segment or a part of code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order than that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be implemented by a dedicated hardware-based system that performs the specified function or operation, or can be implemented by a combination of dedicated hardware and computer instructions.

[0189] The terms "first", "second", etc. are used to distinguish similar objects, rather than to describe or indicate a particular order or sequence.

[0190] The term "comprise" or any other similar term is intended to cover non-exclusive inclusion such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0191] Thus far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art may make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will fall within the scope of protection of the present invention.

Claims

1. A method for generating ground signs based on an airport information plan (AIP) diagram, characterized in that: The method comprises the following steps: S1. Obtain ICAO rule vector based on pre-built dynamic semantic knowledge base; S2. Based on the identification information in the airport regulations AIP graph and the ICAO rule vector, a knowledge graph is constructed and semantically embedded to obtain the symbolic features of the AIP graph. The node attributes of the knowledge graph include ICAO standard codes, geometric parameters, and topological relationships. The method to obtain the AIP diagram symbol features is: Extract the identification information from the airport regulations AIP diagram, transform the identification information into a triple knowledge graph through named entity recognition and relationship extraction technology, use the TransE algorithm to map entities and relationships into low-dimensional vectors, and obtain a geometric-topological feature vector containing geometric positions and topological relationships. The geometric-topological feature vector is then combined with the ICAO rule vector to form the AIP diagram symbol feature. ; S3. Obtain satellite images, perform cross-modal feature alignment between the satellite images and AIP map symbol features, and enhance the features based on a bidirectional attention mechanism; The cross-modal feature alignment between satellite images and AIP map symbol features is performed as follows: A spatial transformer network is used to geometrically align satellite images with AIP maps: Constructing a cross-modal projection matrix 、 , using contrastive learning loss function, the AIP graph symbol features Satellite image features Map to a shared semantic space for semantic alignment: ； Contrastive Learning Loss Function for: ； in, 、 is the eigenvector after projection matrix mapping, is the satellite image feature of the i-th negative sample feature vector in the satellite image feature set, N is the number of samples, is the temperature parameter, They are shared space dimension, AIP map symbol feature original dimension, and satellite feature original dimension respectively; is a function for calculating the similarity of two feature vectors; The method of enhancing features based on the bidirectional attention mechanism is as follows: Use AIP map symbol features as the query and satellite image features as the key and value: ； Using satellite image features as queries and AIP map symbol features as keys and values: ； in, is the feature dimension; Calculate the function for the attention weight between modal features; The bidirectional attention weights are fused with the original features to generate enhanced features: ； ； in, is the scaling factor, which is dynamically adjusted through back-propagation; S4. Extract the multi-scale features of the satellite image after feature enhancement, combine them with the AIP map symbol features after feature enhancement, build a three-layer feature fusion network, determine the dynamic weight of each layer of features according to the identification type, and obtain the weighted comprehensive features; S5. Splicing the comprehensive features with the ICAO rule vector as input, generating a logo image based on a conditional GAN ​​generator, and performing geometric and topological checks on the generated results.

2. The method for generating ground markings based on the airport detailed information AIP map according to claim 1, characterized in that: The method to obtain the ICAO rule vector is: According to the current airport type, the corresponding ICAO rule vector is queried in real time in a pre-built dynamic semantic knowledge base; wherein, the dynamic semantic knowledge base is constructed by: The ICAO rules are obtained and semantically parsed to obtain the semantic information and logical relationships corresponding to each semantic rule. The semantic rules are vector-encoded based on the knowledge graph embedding algorithm to obtain the ICAO rule vectors and form a dynamic semantic knowledge base.

3. The method for generating ground signs based on the airport detailed information AIP map according to claim 1, characterized in that: After obtaining the attention-enhanced features, the offset is learned using deformable convolution to dynamically adjust the sampling point position. The method is as follows: ； in, represents feature splicing, is the offset of the deformable convolution, is the modulation weight; The learning objectives of feature alignment, bidirectional attention, and deformable convolution are integrated into a unified loss function L: ； ； ； in, is the classification loss; is the alignment loss; is the offset loss, is the actual marked offset, is the weight of the loss.

4. The method for generating ground markings based on the airport detailed information AIP map according to claim 1, characterized in that: The method to obtain comprehensive features is: ResNet-50 extracts multi-scale features of satellite images after feature enhancement, including bottom-level detail features, mid-level shape features, and high-level semantic features; Based on the enhanced AIP symbol features and ICAO rule vectors, multi-scale features are hierarchically fused to obtain bottom-level features, middle-level features, and high-level features. The bottom-level features, middle-level features and high-level features are fused from top to bottom to generate fused features of each layer, and the weights of the fused features of each layer are determined according to the identification type, and the weighted comprehensive features are obtained.

5. The method for generating ground markings based on the airport detailed information AIP diagram according to claim 4, characterized in that: The method to obtain comprehensive features is: After upsampling the high-level features, add them to the middle-level features, and strengthen cross-layer interactions through element-by-element multiplication. Then, upsample the fusion result and fuse it with the bottom-level features: ； in, is the upsampling operation function, is the interaction coefficient, Represents element-wise multiplication; is the lateral input feature of the i-th layer, is the fusion feature of the current i-th layer, is the fusion feature of the previous layer; Determine dynamic weight based on identification type : ； Perform weighted fusion on the fused features to obtain comprehensive features : ； in, is the weight of different identification types, which is dynamically adjusted according to importance. MLP is a multi-layer perceptron network. is an exponential function, and N is the number of layers.

6. The method for generating ground markings based on the airport detailed information AIP diagram according to claim 5, characterized in that: The method for generating logo images based on the conditional GAN ​​generator is as follows: The comprehensive features are concatenated with the ICAO rule vector as the input x of the conditional GAN ​​generator: ； Among them, r is the ICAO rule vector. According to the current airport type, the ICAO rule vector is queried in real time from the dynamic semantic knowledge base; The semantic loss for generating the logo is: ； in, is the L2 norm, which is used to measure the Euclidean distance between two vectors; For the conditional GAN ​​generator, the input is mapped to the generated identity, A semantic embedding function for generating logos; The loss function is defined as: ； in, To combat losses, is the weight coefficient.

7. A ground sign generation system based on an airport specification AIP map, according to the ground sign generation method based on an airport specification AIP map according to any one of claims 1 to 6, characterized in that: include: a data preprocessing module configured to acquire satellite images and perform preprocessing; Based on the identification information in the airport regulations AIP diagram and the ICAO rule vector, a knowledge graph is constructed and semantic embedding is performed to obtain the symbolic features of the AIP diagram. Get ICAO rule vector; A feature alignment module is configured to perform cross-modal feature alignment on the pre-processed satellite imagery and the AIP map symbol features, and enhance the features based on a bidirectional attention mechanism; The feature fusion module is configured to extract multi-scale features of the satellite image after feature enhancement, combine them with the symbolic features of the AIP map after feature enhancement, and generate comprehensive features through dynamic weight aggregation to generate comprehensive features; The logo generation and verification module is configured to use the combined comprehensive features and the ICAO rule vector as input, generate a logo image based on a conditional GAN ​​generator, and perform geometric verification and topological verification on the generated result.

Images