An online automatic thematic mapping method and system based on artificial intelligence

Abstract

The present application relates to the technical field of automatic mapping, in particular to an online automatic thematic mapping method and system based on artificial intelligence; the method comprises the following steps: synchronously collecting multi-modal input parameters of a user and performing time alignment, respectively extracting geometric features and semantic features, and generating unified semantic geometric vectors through adaptive confidence mechanism fusion; through double-channel parallel analysis of semantic projection and geometric projection, candidate mapping schemes of the concept layer and the graphic layer are respectively generated, and the candidates are calibrated and fused in terms of semantic consistency and spatial consistency; the candidate set is subjected to local geometric optimization, semantic adjustment and global layout optimization, and micro-variation is introduced to generate diversity, and the output graphics can be directly rendered online through feature weighting fusion and global consistency optimization; a self-learning knowledge base is constructed by collecting historical samples, and continuous accumulation and personalized output optimization of mapping experience are completed. The present application realizes a complete intelligent closed loop from multi-modal input to automatic mapping.

Description

Technical Field

[0001] This invention relates to the field of automatic mapping technology, specifically to an online automatic thematic map creation method and system based on artificial intelligence. Background Technology

[0002] With the deepening application of artificial intelligence technology in the field of graphics generation, online automatic mapping technology is developing from single instruction processing to multimodal intelligent understanding. Current mapping systems usually rely on text descriptions or pre-made templates for graphics generation. Although they can achieve basic functions, they still face integration challenges when dealing with diverse interaction methods such as hand-drawn input and voice commands.

[0003] Chinese invention patent application CN120318354A discloses an interactive painting generation method, system, and device based on artificial intelligence, relating to the field of artificial intelligence technology. It achieves high-quality image generation through multimodal data fusion, semantic alignment, conflict detection, and iterative correction. The scheme includes the following core steps: acquiring multimodal data such as text, speech, and sketches and processing them into a formatted feature dataset; aligning and weighting the multimodal features to generate a unified semantic space feature; using a visual-text attention mechanism to detect semantic inconsistencies between the image and text descriptions, generating a conflict index matrix and triggering an alarm marker; and iteratively correcting and generating the image based on the unified semantic space feature, the conflict index matrix, and the alarm marker.

[0004] However, in the specific field of thematic map creation, systems typically rely on text descriptions or predefined templates to generate graphics with specific themes and symbolic expressions, achieving basic information visualization functions. Meanwhile, researchers are actively exploring the integration of more natural and diverse interaction methods, such as hand-drawn input and voice commands, into the mapping process to improve the efficiency and richness of intent expression. This is particularly beneficial when dealing with thematic maps containing complex semantics, symbolic annotations, and spatial layouts. However, effectively achieving temporal alignment and unified understanding of multimodal inputs, and further ensuring the coordination and professionalism of generated thematic maps in terms of geometric structure, semantic information, and visual aesthetics, has become a noteworthy technical direction in this field. There is an urgent need to establish a technical framework that can connect multimodal input understanding with high-quality thematic map generation. Summary of the Invention

[0005] The purpose of this invention is to address the problems existing in the background technology by proposing an online automatic thematic map creation method and system based on artificial intelligence.

[0006] The technical solution of this invention: an online automatic thematic map creation method based on artificial intelligence, comprising the following specific implementation steps: S1. Synchronously collect the user's hand-drawn trajectory, voice commands, text descriptions, and structured parameters to construct a unified time index to achieve temporal alignment of multimodal inputs; extract geometric features from the temporally aligned hand-drawn trajectory, and extract semantic features from the temporally aligned voice commands and text descriptions; fuse the geometric and semantic features through an adaptive confidence mechanism to generate a unified semantic geometric vector. S2. Based on semantic geometric vectors, candidate cartographic schemes for the conceptual layer and the graphical layer are generated through parallel parsing of semantic projection and geometric projection. The candidate cartographic schemes are calibrated and fused for semantic consistency and spatial consistency to generate an intermediate cartographic scheme set. S3. Perform local geometric optimization, semantic adjustment and global layout optimization on the intermediate mapping scheme set, and introduce micro-variation to generate diversity; S4. Through feature weighted fusion and global consistency optimization, output graphics that can be directly rendered online; S5. By collecting historical cartographic samples and extracting geometric, semantic, and visual features, these samples are structured and stored in a knowledge base. Cluster analysis is used to form template nodes, completing the summarization of cartographic patterns and the accumulation of experience. Semantic similarity is combined to perform template migration and adaptive parameter adjustment. At the same time, newly generated graphics are dynamically updated and optimized to build a closed-loop self-learning mechanism, continuously accumulating knowledge, optimizing output, and supporting personalized automatic cartography.

[0007] Preferably, the specific steps for timing alignment include: The location information, pen pressure, pen speed, and pause interval of the hand-drawn trajectory are collected at a millisecond sampling frequency on the global timeline. Simultaneously capture voice signals and receive real-time input text commands; Hand-drawn trajectory data, voice signal sequences, text-based instructions, and structured parameter inputs are uniformly encapsulated in the acquisition layer; The time difference between the timestamps of speech segments and text events and the timestamps of hand-drawn trajectory points is calculated using a time alignment function; When the time difference is less than the preset time difference threshold, the voice event and text event are determined to be associated with the hand-drawn trajectory action.

[0008] Preferably, the specific steps for extracting geometric features include: Curve fitting and dynamic feature analysis were performed on hand-drawn trajectories composed of continuous point sequences. Extract pen pressure information, pen speed change rate information, and trajectory curvature information to generate a stroke feature vector sequence; The stroke feature vector sequence is input into the attention-based Bi-LSTM temporal coding model; A global geometric intent vector is generated through a temporal coding model to capture the shape of the graphic, structural trends, and user drawing logic.

[0009] Preferably, the specific steps for extracting semantic features include: The speech signal was transcribed into a text sequence, and speech energy features and speech rate fluctuation features were extracted. Calculate speech confidence weights based on speech energy characteristics and speech rate fluctuation characteristics; Text instructions are embedded to obtain high-dimensional vector representations; The high-dimensional vector representation is weighted using speech confidence weights to generate an overall semantic vector.

[0010] Preferably, the specific steps of the adaptive confidence mechanism include: The confidence level of the geometric vector is obtained by evaluating the confidence level of the geometric intention vector. The evaluation factors include stroke integrity, curvature continuity and overall closure. The confidence of semantic vectors is obtained by evaluating the confidence of semantic vectors. The evaluation factors include speech energy, speech rate fluctuation and weighted text weight. The fusion coefficients of geometric and semantic modalities are dynamically calculated based on the confidence scores of geometric and semantic vectors. The geometric intent vector and semantic vector are mapped to the same high-dimensional fusion space through a trainable mapping matrix; The mapped geometric vector and semantic vector are weighted and fused according to the fusion coefficient to generate a unified semantic geometric vector.

[0011] Preferably, the specific steps of dual-channel parallel parsing include: Extract geometrically relevant sub-vectors and semantic sub-vectors from the unified semantic geometric vector; The geometrically related subvectors are mapped to a set of geometric candidates through a geometric projection mechanism. The geometric projection mechanism uses a geometric mapping matrix and a geometric bias vector, and introduces a local perturbation strategy based on Gaussian distribution during the projection process. The semantic subvectors are mapped to a set of semantic candidates through a semantic projection mechanism. The semantic projection mechanism uses a semantic mapping matrix and a semantic bias vector, and introduces a parameter perturbation mechanism based on Gaussian distribution during the projection process. Calculate the matching score between the geometric candidate set and the semantic candidate set. The matching score is a weighted sum of geometric similarity and semantic similarity. Based on the matching score, high consistency candidate pairs are fused to form a cross-modal fusion candidate set; The cross-modal fusion candidate set is sorted and filtered by confidence to output a high-quality candidate set.

[0012] Preferably, the specific steps of local geometry optimization include: The vertex set of the candidate graph is subjected to weighted Bézier curve smoothing, and the neighborhood vertex coordinate weighted average algorithm is used to ensure boundary smoothness and curve continuity. Constrain the proportions of the candidate graphics to ensure that the local proportions are coordinated and reasonable; The positions of symbols and annotation elements in the candidate graphics are fine-tuned by using a dynamic semantic adjustment function that combines the geometric center of the candidate graphics, adjacent elements, and user intent.

[0013] Preferably, the specific steps for global layout optimization include: Calculate the global energy function of all elements of the candidate graphic. The global energy function includes the distance deviation energy between elements, the overlap energy between elements, and the aesthetic energy. Aesthetic energy includes visual indicators such as symmetry, proportional balance, and layout balance. Heuristic optimization algorithms are used to adjust the positions of candidate elements in order to minimize the global energy function.

[0014] Preferably, the specific steps for introducing subtle variations to generate diversity include: A small random perturbation following a zero-mean normal distribution is introduced into the vertex positions of the optimized candidate graph; A small random perturbation following a zero-mean normal distribution is introduced into the sign position of the optimized candidate graphic; Several small-variant candidates are generated in parallel to form a small-variant candidate set.

[0015] The technical solution of this invention: An online automatic thematic map making system based on artificial intelligence, which is used to execute the above-mentioned online automatic thematic map making method based on artificial intelligence, including: The multimodal input acquisition and semantic vectorization module is used to simultaneously collect the user's hand-drawn trajectory, voice commands, text descriptions and structured parameters and perform temporal alignment, extract geometric features and semantic features respectively, and fuse them through an adaptive confidence mechanism to generate a unified semantic geometric vector; The dual-projection parallel candidate generation module is used to generate candidate mapping schemes for the conceptual layer and the graphical layer respectively through dual-channel parallel analysis of semantic and geometric projection based on semantic and geometric vectors, and to perform semantic consistency and spatial consistency calibration and fusion on the candidates. The cross-modal candidate optimization and micro-variation generation module is used to perform local geometric optimization, semantic adjustment and global layout optimization on high-quality candidate sets, and introduces micro-variation generation diversity. The final graphics rendering and online output module is used to output graphics that can be directly rendered online through feature weighted fusion and global consistency optimization. The self-learning knowledge base module is responsible for collecting and generating graphic features, building a template library, performing semantic-driven parameter migration, and dynamically updating the knowledge base to complete the system's continuous self-learning and experience accumulation.

[0016] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects: This invention designs an online automatic thematic map generation method and system based on artificial intelligence. By using temporal alignment of multimodal inputs and adaptive confidence fusion, it effectively improves the accuracy and robustness of capturing complex user mapping intentions. Its dual projection mechanism can generate diverse candidates in parallel at the semantic and geometric levels, and achieves high-quality matching between semantics and graphics through cross-modal calibration and optimization, significantly enhancing the rationality and completeness of automatic map generation. Subsequent geometric optimization, semantic adjustment, and global layout processing further ensure the comprehensive quality of the output graphics in terms of structural standardization, visual aesthetics, and clarity of information expression. Simultaneously, the micro-variation generation strategy introduced by the system effectively enriches the diversity of graphic outputs while maintaining the core design intent, meeting the personalized needs of different scenarios and users. The entire process of this invention achieves full automation from multimodal input to high-quality graphic generation, significantly reducing the technical threshold and time cost of professional mapping, and possessing good practicality and ease of use. Attached Figure Description

[0017] Figure 1 This is a flowchart of an online automatic thematic map creation method based on artificial intelligence proposed in this invention. Figure 2 This is a system architecture diagram of an online automatic thematic map making system based on artificial intelligence proposed in this invention. Detailed Implementation

[0018] Example 1, as Figure 1 As shown, the present invention proposes an online automatic thematic map creation method based on artificial intelligence, which includes the following specific implementation steps: S1. By synchronously collecting users' hand-drawn trajectories, voice commands, text descriptions, and structured parameters, a unified time index is constructed to achieve temporal alignment of multimodal inputs. Subsequently, geometric and semantic features are extracted separately, and geometric intent and semantic intent are fused through an adaptive confidence mechanism to generate a unified high-dimensional semantic-geometric vector. The specific implementation process is as follows: S11. Collect hand-drawn trajectories, voice signals, text input, and structured parameters through the front end, and align the data of each modality on the global timeline to form a unified multimodal input stream, specifically: During the user's drawing process, the stylus position information (x) is collected at a millisecond sampling frequency. t ,y t ), pen pressure pt Pen speed v t and the pause time interval τ t ; Simultaneously, it captures the microphone signal A(t) and receives commands or explanatory sentences input in real time; The above data is uniformly encapsulated at the acquisition layer as follows: ; During the time synchronization phase, a global time axis T={t1,t2,…,t} is constructed. n Using time alignment functions: ; when They believe that the voice / text event is associated with the stroke action. Where S(x,t) represents the hand-drawn trajectory data, including the position (x,t). t ,y t ), pen pressure p t Speed ​​v t and pause time τ t V(t) represents the speech signal sequence, including time-domain signal, spectral features, and energy; T(t) represents textual instructions or descriptions; P(t) represents structured parameter input, including but not limited to size, angle, scale, and level; C(t) represents context control information, including but not limited to reference templates and the current drawing layer state. The timestamp of point i in the hand-drawn trajectory; This represents the timestamp of audio segment j; This represents the time difference between two inputs; This represents the time difference threshold, used to determine input correlation. S12. Perform curve fitting and dynamic feature analysis on the hand-drawn trajectory, extract information such as pen pressure, pen speed change rate, and curvature, generate a stroke feature vector sequence, and generate a global geometric intent vector through a temporal coding model to capture graphic shape, structural trends, and user drawing logic, thereby achieving a high-dimensional representation of geometric intent. Specifically: Each stroke is composed of a continuous sequence of points {(x t ,y t ,p t ,v t The stroke vector is calculated through curve fitting and dynamic feature extraction. ; stroke sequence The input is fed into a lightweight temporal coding model (in this embodiment, an attention-based Bi-LSTM structure is used) to generate a global geometric intent vector: ; in, and This represents the difference in coordinates between trajectory point i and i+1, reflecting local directional changes; It indicates the rate of change in pen speed, depicting the transitions and smoothness of movements; It represents the curvature of a trajectory, measuring the shape of arcs, closures, or corners; G represents the encoding function, which maps local stroke features to a fixed-dimensional vector; G represents the global geometric intent vector, which is the high-dimensional geometric features of the entire stroke sequence. This represents the parameters of the trained temporal encoder; in this embodiment, a Bi-LSTM network is used. S13. Perform semantic parsing on the speech and text input, extract key instructions, calculate confidence weights based on speech energy and speech rate fluctuations, and generate a weighted semantic vector by weighting the text embedding representation. Specifically: The speech signal V(t) is transcribed into a text sequence, and speech energy and rhythm features are extracted. Define speech confidence weights: ; Embedding computation of semantic segments with confidence scores: ; Where K represents the total number of text or speech segments currently collected; E represents the speech confidence weight; j This represents the average energy of speech segment j; This represents the overall average speech energy. α represents the speech rate fluctuation rate; β are empirical weighting coefficients; S represents the overall semantic vector, reflecting a weighted combination of the user's oral or written intentions; This represents the high-dimensional vector representation of text segment j. S14. The geometric intent vector and semantic vector are fused using an adaptive confidence fusion mechanism to generate a unified semantic-geometric vector. The geometric and semantic weights are dynamically adjusted to adapt to incomplete or biased input, providing a robust core input representation for subsequent dual projection and automatic mapping. Specifically: Confidence assessments are performed on the geometric intent vector G and the semantic vector S, measuring input completeness, clarity, and reliability, respectively. ; ; in, The confidence level of the geometric vector is represented by stroke integrity, curvature continuity, and overall closure. The semantic vector confidence is represented by the speech energy E. j Speech rate fluctuation σ j With weighted text weight w j Comprehensive calculation; and These are the geometric and semantic confidence evaluation functions, respectively; To fuse feature vectors from different modalities, it is necessary to first map the geometric vector G and the semantic vector S to the same dimension d. h : ; ; Among them, W g W s b represents a trainable mapping matrix; g b s Represents the bias vector; Representation of geometric information in a high-dimensional fusion space; The representation of semantic information in a high-dimensional fusion space; The fusion coefficient γ of geometric and semantic modalities is dynamically calculated based on the input confidence level. ; The two are weighted and fused to obtain the final unified semantic geometry vector H: ; in, This represents the weight ratio of geometric modes in the final fusion, with a value range of [0,1].

[0019] S2. Based on the unified semantic geometric vector generated in step S1, candidate cartographic schemes for the conceptual layer and the graphic layer are generated through dual-channel parallel analysis of semantic projection and geometric projection, respectively. A dynamic mapping relationship is established between the two to achieve self-consistent derivation from abstract semantics to concrete graphics. The specific implementation process is as follows: S21. Through a semantic projection mechanism, the unified semantic geometric vector is mapped to a high-dimensional semantic latent space. The semantic distribution of the latent graphical intent is calculated using the semantic constraint matrix, generating a preliminary semantic candidate set, specifically: Extracting geometrically related sub-vectors from a unified vector H: ; Based on geometric correlation subvector H g Generating a set of morphological candidates using multi-channel projection operators: ; To avoid generating overly simplistic or rigid candidate pools, a local perturbation strategy is introduced during the geometric projection process: ; After completing the geometric projection, we obtain the geometric candidate set: ; Among them, H g Represents geometrically related subvectors, i.e. geometric feature vectors extracted from a unified vector, encoding curvature, closure, scale, and local morphological information; Represents the geometric mapping matrix, a weight matrix that maps a high-dimensional unified vector to the geometric latent space; Represents the geometric bias vector, a mapping bias used to adjust the centering and nonlinear smoothing of geometric subvectors; The geometric candidate figure represents the i-th generated geometric candidate, including vertex coordinates, curve parameters, scale, and closure information; N g This represents the number of geometric candidates, i.e., the total number of geometric candidates generated in parallel. This represents the geometric projection operator, i.e., the geometric subvector H. g The function that maps to the i-th geometric candidate figure comes from the structured geometry generation function defined inside the algorithm, combined with local curvature fitting, closure constraints and scaling control; The perturbation vector is generated using a Gaussian distribution for each geometric candidate's tiny random perturbation. Indicates a Gaussian distribution; Variance represents the magnitude of random perturbations. The larger the variance, the larger the perturbation magnitude and the greater the candidate diversity; the smaller the variance, the smaller the perturbation and the more stable the original geometric features are. This represents the geometric candidate vector generated after perturbation, combining the original geometric features with the random perturbation; C g Represents the geometric candidate set; S22. Input the semantic candidate set into the geometric projection mechanism, calculate the multidimensional geometric parameter distribution through the parameterized morphological mapping function, generate a geometric candidate group with attributes such as shape, proportion, and orientation, and retain the mapping index with the semantic label to realize the mapping from semantics to geometry. Specifically: Extracting semantic subvectors from the unified vector H: ; semantic vector H s The graph parameter space is mapped to generate diverse candidates, and semantic weight priority is considered during the projection process. ; To increase the diversity of semantic candidates, a parameter perturbation mechanism is introduced during the projection process: ; After semantic projection is completed, the semantic candidate set is obtained: ; in, This represents the semantic mapping matrix, which is the weight matrix that maps a unified vector to the semantic latent space; H represents the semantic bias vector, a mapping bias used to adjust the centering and non-linear expressive power of semantic sub-vectors; sIt represents a semantic feature vector extracted from a unified vector, encoding user language descriptions, text instructions, and semantic constraint information; The semantic candidate graphic, i.e., the j-th semantic candidate, includes parameters such as size, scale, symbol annotation, and positional constraints; N s w represents the number of semantic candidates, the total number of candidates generated by semantic projection; priority This indicates semantic priority weight, which is assigned based on the information or keywords emphasized by the user, and is used to strengthen key semantic constraints; The semantic projection operator represents the semantic vector H. s The function that maps to the parameter space of the j-th semantic candidate graph comes from the projection function defined inside the algorithm. Combined with semantic constraints and user instruction parsing, semantic candidates are generated to ensure that the candidate graphs conform to the user instructions. This represents the perturbation-resolved semantic candidate, i.e., the final semantic candidate vector, which combines the original semantic mapping with the perturbation. Represents a perturbation vector, a small random perturbation vector for the semantic candidate parameters; Represents the perturbation vector It follows a Gaussian (normal) distribution; C represents variance, controlling for the magnitude or intensity of the disturbance; s Represents a semantic candidate set; S23. Perform semantic and spatial consistency calibration on semantic and geometric candidates, calculate their matching scores using a confidence weighting function, and dynamically cluster and filter high-consistency combinations to form a set of intermediate mapping schemes for semantic-geometric dual-domain association, specifically: Computational geometry candidate and semantic candidates Similarity between them: ; Based on the matching score S ij High-scoring candidate pairs are selected for fusion: ; After fusion, the candidate set is managed in parallel using confidence level P. k Sort: ; Output cross-modal fusion candidate set: ; in, Geometric similarity is used to measure the degree of matching between geometric and semantic candidates in terms of shape, proportion, closure, etc. It represents semantic similarity and measures the degree to which candidates match semantic constraints (such as size, symbols, position, and user-emphasized information); and S represents the weighting coefficient, controlling the proportion of geometric similarity and semantic similarity in the total matching score; ij Represents the matching score, which is the combined matching score of geometric candidate i and semantic candidate j; Represents the fused candidate vector, the k-th fused candidate, and the multimodal candidate generated by combining geometric and semantic information; M represents the fusion coefficient, controlling the proportion of geometry and semantics in the fusion candidates; k P represents the candidate pair index set, used to generate the geometric and semantic candidate indices for the k-th fusion candidate; k Indicates the confidence level, i.e., the fusion candidate. The overall confidence score reflects the candidate quality and matching priority; The structural integrity score is used to evaluate the structural rationality of candidate graphics, including the continuity, closure, and proportional coordination of local curves, and is automatically calculated based on differential curvature analysis. C represents the weighting coefficient, balancing the proportion of the matching score and the structural integrity score in the confidence calculation. f N represents the cross-modal fusion candidate set; f Indicates the total number of cross-modal fusion candidates; S24. Using a cross-domain fusion module, the intermediate mapping scheme set is input into a dual optimization network to jointly evaluate morphological rationality and semantic integrity, dynamically adjust projection weights, and output the optimal candidate mapping result. This provides high-confidence input for subsequent structural optimization and layout adjustment, achieving a self-consistent semantic and geometric generation loop. Specifically: Confidence level P of fusion candidates k Normalization is performed: ; Normalized confidence Conduct multi-dimensional scoring: ; Based on the final optimal score R k Perform parallel screening to retain a high-quality candidate set C o : ; in, The innovation score quantifies the difference and creativity of a candidate relative to the existing candidate set. It calculates the difference or information gain by comparing with historical candidates or template library, including but not limited to form, proportion, symbol combination, and layout variation. This represents the weighting coefficient, which controls the proportion of innovation in the final score; This indicates the filtering threshold, which determines which candidates to retain. The minimum scoring standard; C o This represents the set of candidates after filtering.

[0020] S3. Further optimize and rearrange the high-quality candidate set output in step S2 to achieve final candidate graphics with geometric rationality, semantic consistency, visual aesthetics, and innovative diversity, providing directly renderable graphic output for online automatic mapping. The specific implementation process is as follows: S31. Perform neighborhood-weighted smoothing and proportional constraint processing on the vertex set and curve structure of the candidate graphs to ensure smooth boundaries, continuous curves, and reasonable local proportions. This provides a reliable geometric basis for subsequent semantic adjustments and global rearrangement, while maintaining the individualized form of the candidates. That is: for the candidate set C o For each candidate, perform local geometry optimization to ensure smooth boundaries, curve continuity, and proportional consistency. For each candidate The vertex set is smoothed using a weighted Bézier curve algorithm: ; in, V represents the optimized vertex coordinates; i This represents the coordinates of the i-th vertex of the candidate graphic; Indicates smoothing weights, N(i) represents the neighborhood set of vertex i; |N(i)| represents the number of neighborhood vertices, used for normalized average calculation. S32. Based on local geometric optimization, automatically adjust the positions of symbols, labels, and elements in the candidates to ensure clear expression of semantic information without obscuring key structures. Through a dynamic semantic adjustment function combined with the geometric center, adjacent elements, and user intent, achieve natural coordination between semantics and geometry. That is, based on geometric optimization, adjust the position, size, and symbol labels of each element in the candidates to ensure reasonable expression of semantic information. For candidate symbols or annotation elements S i Make minor adjustments to the position: ; Among them, S i Indicates the position of the original symbol or annotation element; This indicates the adjusted position of symbols or labels to ensure their semantic validity within the drawing; This represents the step size coefficient, which controls the adjustment range and is derived from system strategy or user preferences. This represents a semantic adjustment function that generates an adjustment vector based on information such as candidate geometric centers, adjacent elements, and user intent. S33. Optimize the global element layout within the candidate elements to achieve visual readability, proportional harmony, and innovative design, namely: For candidates Calculate the global energy function E for all elements total : ; Use heuristic optimization to adjust the positions of candidate elements and minimize E. total The final layout was obtained; Among them, E dist E represents the energy of the distance deviation between elements, used to maintain appropriate spacing between elements; overlap E represents the energy of element overlap, used to prevent graphic elements from overlapping; aesthetic It represents aesthetic energy, including visual indicators such as symmetry, proportion, and balanced layout; , and This represents the weighting coefficient, which controls the contribution of the three types of energy to the total energy and can be dynamically adjusted according to user preferences or task requirements. S34. Generate multiple minor variants for each optimization candidate to preserve candidate diversity and innovativeness, while managing the generated results in parallel, specifically as follows: Based on the optimized candidates, a small perturbation is introduced to slightly adjust the vertex and sign positions, generating M sets of slightly modified candidate sets in parallel. ; ; ; in, , These represent the vertex and sign positions after minor changes, respectively, used to generate diverse candidates; , These represent the optimized vertex and sign positions, respectively. , They represent small random disturbances, with a mean of 0 and a variance of σ. 2 It follows a normal distribution.

[0021] S4. Based on the candidate set optimized and slightly modified in step S3, a high-quality graphic that can be directly displayed online is generated through candidate feature weighted fusion, global consistency optimization, slightly modified variant generation, and final graphic rendering. The specific implementation process is as follows: S41. The candidate set of minor changes output in step S3 Feature-weighted fusion is performed to integrate the advantageous features of each candidate, generating preliminary fusion candidates: Extract candidate feature vector F k This includes geometric coordinates, semantic labels, local morphology, and scale information; Calculate the fused feature vector: ; ; in, F represents the fused candidate feature vector, providing a unified input for global consistency optimization; kThe feature vector representing the k-th candidate contains geometric coordinates, curvature, symbolic annotation, scale information, and semantic label; w k R represents the candidate weights; k The score represents the preferred score; M represents the total number of candidates for minor variations, which is determined by the set of minor variations output in step S3. S42. Based on the fused feature vector, define a three-dimensional energy function encompassing geometry, semantics, and visual aesthetics. Minimize the energy using a constraint optimization method, and adjust the candidate layout, proportion, and element position accordingly: Based on fusion feature F fused Define a globally consistent energy function E. consistency : ; Among them, E geom Represents geometric constraint energy, ensuring boundary, scale, curve continuity, and local symmetry; E semantic This represents semantic constraint energy, ensuring that symbol annotations, text information, and functional elements are laid out reasonably and do not obscure key structures; E visual It represents visual aesthetic energy, including symmetry, visual guidance, layout balance, and a sense of proportion among elements; , , This represents adjustable weighting coefficients that dynamically control the importance of geometry, semantics, and aesthetics. Minimize E using a heuristic constraint optimization algorithm. consistency The optimized fusion candidate C is obtained. opt ; S43. Based on the global optimization candidate, a small random perturbation is introduced to generate micro-variants. By generating several candidates in parallel, local feature micro-innovation is maintained, output diversity is increased, and diverse options are provided for the final candidate selection. Specifically: For each vertex V opt,i and symbol position S opt,i Introduce small random perturbations: ; ; Parallel generation of N v A set of candidate variants: Ensure diversity and innovation without compromising overall geometric and semantic harmony; Among them, V opt,i Indicates optimization of candidate C opt The coordinates of the i-th vertex; S opt,i Indicates optimization of candidate C opt The position of the i-th symbol; C v Represents the final set of candidate minor variations; This represents the k-th final differential candidate; S44. The variant candidate set C v It converts data into vector graphics or drawing instructions, applies suitable rendering styles and symbol templates, and enables online visual output. It also supports user interaction to fine-tune key parameters, ensuring that the output graphics are both reasonable and aesthetically pleasing, while also being innovative and selective, forming a complete online automatic drawing closed loop.

[0022] S5. By recording and analyzing users' historical cartographic behavior and system output results, the system automatically extracts cartographic patterns and parameter features to form a knowledge graph-style template library. A self-learning mechanism is also established, enabling the system to adaptively call upon experience templates in subsequent tasks, achieving continuous optimization and personalized evolution. The specific implementation process is as follows: S51. After a user completes an online mapping task, the generated graphic data, parameter configuration, and user interaction trajectory are automatically saved; the mapping result G is then extracted using the feature extraction module. t Represented as a high-dimensional feature vector F t : ; in, This indicates an operation to extract geometric structural features (such as shape proportions and node distribution); This indicates the operation of extracting semantic features (such as symbols, text, and label structures); This indicates the operation of extracting visual features (such as color scheme and line style); S52, Regarding the historical feature vector set {F t Cluster analysis is performed to identify high-frequency mapping patterns and form knowledge nodes; Clustering formula: ; Among them, K m This represents the m-th class of cartographic templates; This represents the central eigenvector of the cluster, signifying the core structure of this type of graph; This represents the clustering threshold, used to control the range of template similarity. Based on this, a multi-level template knowledge base is ultimately formed. Each node stores typical templates, parameter configurations, and semantic descriptions; Where M represents the total number of cartographic template clusters; S53. When a new mapping task arrives, based on the input semantic vector z in Similarity to template centers in the knowledge base (sim) m calculate: ; Select the best matching template and adaptively adjust the parameters according to the task characteristics; S54. After the mapping is completed, calculate the degree of difference between the newly generated result and the historical template. : ; If the difference is below the threshold, update the template center. : ; in, This represents the update coefficient, used to balance the contribution of historical experience with new data.

[0023] Example 2, as Figure 2 As shown, the present invention proposes an online automatic thematic map making system based on artificial intelligence, which is used to execute the online automatic thematic map making method based on artificial intelligence proposed in Embodiment 1. It includes: a multimodal input acquisition and semantic vectorization module, a dual projection parallel candidate generation module, a cross-modal candidate optimization and micro-variation generation module, a final graphics rendering and online output module, and a self-learning knowledge base module.

[0024] The multimodal input acquisition and semantic vectorization module is responsible for receiving multi-source inputs uploaded by users, including text descriptions, sketches, images or tabular data, and processing these inputs in a unified manner. It extracts key features of the inputs through artificial intelligence algorithms and maps geometric, semantic and visual information into high-dimensional vector representations to achieve a unified expression of semantics and geometry. The dual-projection parallel candidate generation module generates geometric projection candidates and semantic projection candidates simultaneously through a dual-projection strategy: geometric projection candidates preserve the spatial relationships and structural proportions of elements, while semantic projection candidates ensure the accuracy of functional symbols and labels; the candidates are managed in parallel, optimized, scored, and initially fused, and each candidate is assigned a weight to retain advantageous features; The cross-modal candidate optimization and micro-variation generation module performs cross-modal feature fusion and global consistency optimization on the candidates. Through multi-dimensional constraints of geometry, semantics and visual aesthetics, it optimizes the candidate layout and element relationships to ensure the overall harmony of the graphics. The optimized candidates are further generated into micro-variants, which increase diversity and innovation through small random perturbations, forming a rich candidate set. The final graphics rendering and online output module receives the optimized candidate and micro-variation set, transforms the feature vectors into renderable graphics output, including vector graphics, flowcharts, structural diagrams, and schematic diagrams, and provides online interactive functions. Users can select the final candidates or fine-tune key parameters to achieve adaptive output, and automatically match rendering templates and visual styles to ensure that the output graphics are both beautiful and meet professional standards. The self-learning knowledge base module is responsible for collecting and generating graphic features, building a template library, performing semantic-driven parameter migration, and dynamically updating the knowledge base to enable the system to continuously learn and accumulate experience.

[0025] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. An online automatic thematic map creation method based on artificial intelligence, characterized in that, The specific implementation steps include the following: S1. Synchronously collect users' hand-drawn trajectories, voice commands, text descriptions, and structured parameters to build a unified time index to achieve temporal alignment of multimodal inputs; Geometric features were extracted from the time-aligned hand-drawn trajectory, and semantic features were extracted from the time-aligned voice commands and text descriptions, respectively. Geometric and semantic features are fused through an adaptive confidence mechanism to generate a unified semantic-geometric vector; S2. Based on semantic geometric vectors, candidate cartographic schemes for the conceptual layer and the graphical layer are generated through dual-channel parallel analysis of semantic projection and geometric projection, respectively. The candidate mapping schemes are calibrated and fused for semantic and spatial consistency to generate an intermediate mapping scheme set. S3. Perform local geometric optimization, semantic adjustment and global layout optimization on the intermediate mapping scheme set, and introduce micro-variation to generate diversity; The specific steps of local geometry optimization include: The vertex set of the candidate graph is subjected to weighted Bézier curve smoothing, and the neighborhood vertex coordinate weighted average algorithm is used to ensure boundary smoothness and curve continuity. Constrain the proportions of the candidate graphics to ensure that the local proportions are coordinated and reasonable; The positions of symbols and annotation elements in the candidate graphics are fine-tuned by using a dynamic semantic adjustment function that combines the geometric center of the candidate graphics, adjacent elements, and user intent. The specific steps for global layout optimization include: Calculate the global energy function of all elements of the candidate graphic. The global energy function includes the distance deviation energy between elements, the overlap energy between elements, and the aesthetic energy. Aesthetic energy includes visual indicators such as symmetry, proportional balance, and layout balance. Use a heuristic optimization algorithm to adjust the positions of candidate elements in order to minimize the global energy function; The specific steps for introducing micro-variation to generate diversity include: A small random perturbation following a zero-mean normal distribution is introduced into the vertex positions of the optimized candidate graph; A small random perturbation following a zero-mean normal distribution is introduced into the sign position of the optimized candidate graphic; Several small-variant candidates are generated in parallel to form a small-variant candidate set; S4. Through feature weighted fusion and global consistency optimization, output graphics that can be directly rendered online; S5. By collecting historical cartographic samples and extracting geometric, semantic, and visual features, these samples are structured and stored in a knowledge base. Cluster analysis is used to form template nodes, completing the summarization of cartographic patterns and the accumulation of experience. Semantic similarity is combined to perform template migration and adaptive parameter adjustment. At the same time, newly generated graphics are dynamically updated and optimized to build a closed-loop self-learning mechanism, continuously accumulating knowledge, optimizing output, and supporting personalized automatic cartography.

2. The online automatic thematic map creation method based on artificial intelligence according to claim 1, characterized in that, The specific steps for timing alignment include: The location information, pen pressure, pen speed, and pause interval of the hand-drawn trajectory are collected at a millisecond sampling frequency on the global timeline. Simultaneously capture voice signals and receive real-time input text commands; Hand-drawn trajectory data, voice signal sequences, text-based instructions, and structured parameter inputs are uniformly encapsulated in the acquisition layer; The time difference between the timestamps of speech segments and text events and the timestamps of hand-drawn trajectory points is calculated using a time alignment function; When the time difference is less than the preset time difference threshold, the voice event and text event are determined to be associated with the hand-drawn trajectory action.

3. The online automatic thematic map creation method based on artificial intelligence according to claim 2, characterized in that, The specific steps for extracting geometric features include: Curve fitting and dynamic feature analysis were performed on hand-drawn trajectories composed of continuous point sequences. Extract pen pressure information, pen speed change rate information, and trajectory curvature information to generate a stroke feature vector sequence; The stroke feature vector sequence is input into the attention-based Bi-LSTM temporal coding model; A global geometric intent vector is generated through a temporal coding model to capture the shape of the graphic, structural trends, and user drawing logic.

4. The method for online automatic thematic map creation based on artificial intelligence according to claim 3, characterized in that, The specific steps for extracting semantic features include: The speech signal was transcribed into a text sequence, and speech energy features and speech rate fluctuation features were extracted. Calculate speech confidence weights based on speech energy characteristics and speech rate fluctuation characteristics; Text instructions are embedded to obtain high-dimensional vector representations; The high-dimensional vector representation is weighted using speech confidence weights to generate an overall semantic vector.

5. The online automatic thematic map creation method based on artificial intelligence according to claim 4, characterized in that, The specific steps of the adaptive confidence mechanism include: The confidence level of the geometric intention vector is obtained by evaluating the confidence level of the geometric vector. The evaluation factors include stroke integrity, curvature continuity and overall closure. The confidence score of the semantic vector is obtained by evaluating the confidence score of the semantic vector. The evaluation factors include speech energy, speech rate fluctuation and weighted text weight. The fusion coefficients of geometric and semantic modalities are dynamically calculated based on the confidence scores of geometric and semantic vectors. The geometric intent vector and semantic vector are mapped to the same high-dimensional fusion space through a trainable mapping matrix; The mapped geometric vector and semantic vector are weighted and fused according to the fusion coefficient to generate a unified semantic geometric vector.

6. The online automatic thematic map creation method based on artificial intelligence according to claim 5, characterized in that, The specific steps of dual-channel parallel parsing include: Extract geometrically relevant sub-vectors and semantic sub-vectors from the unified semantic geometric vector; The geometrically correlated subvectors are mapped to a set of geometric candidates through a geometric projection mechanism. The geometric projection mechanism uses a geometric mapping matrix and a geometric bias vector, and introduces a local perturbation strategy based on Gaussian distribution during the projection process. The semantic subvectors are mapped to a set of semantic candidates through a semantic projection mechanism. The semantic projection mechanism uses a semantic mapping matrix and a semantic bias vector, and introduces a parameter perturbation mechanism based on Gaussian distribution during the projection process. Calculate the matching score between the geometric candidate set and the semantic candidate set. The matching score is a weighted sum of geometric similarity and semantic similarity. Based on the matching score, high consistency candidate pairs are fused to form a cross-modal fusion candidate set; The cross-modal fusion candidate set is sorted and filtered by confidence to output a high-quality candidate set.

7. An online automatic thematic map making system based on artificial intelligence, used to execute the online automatic thematic map making method based on artificial intelligence as described in any one of claims 1 to 6, characterized in that, include: The multimodal input acquisition and semantic vectorization module is used to simultaneously collect the user's hand-drawn trajectory, voice commands, text descriptions and structured parameters and perform temporal alignment, extract geometric features and semantic features respectively, and fuse them through an adaptive confidence mechanism to generate a unified semantic geometric vector; The dual-projection parallel candidate generation module is used to generate candidate mapping schemes for the conceptual layer and the graphical layer respectively through dual-channel parallel analysis of semantic and geometric projection based on semantic and geometric vectors, and to perform semantic consistency and spatial consistency calibration and fusion on the candidates. The cross-modal candidate optimization and micro-variation generation module is used to perform local geometric optimization, semantic adjustment and global layout optimization on high-quality candidate sets, and introduces micro-variation generation diversity. The final graphics rendering and online output module is used to output graphics that can be directly rendered online through feature weighted fusion and global consistency optimization. The self-learning knowledge base module is responsible for collecting and generating graphic features, building a template library, performing semantic-driven parameter migration, and dynamically updating the knowledge base to complete the system's continuous self-learning and experience accumulation.

Images