Intelligent manufacturing system and method for text travel digital cultural product based on large model

Abstract

The invention discloses an intelligent manufacturing system and method for a text travel digital cultural product based on a large model, and relates to the technical field of text travel digitization, and the system comprises a data graph module which receives and processes multi-source data, and constructs a text travel knowledge graph; the semantic extraction module is used for finely tuning the model to extract semantic tags, and calculating and verifying the tag reliability; the gene pool matching module is used for retrieving a gene pool and matching visual features with cultural relic parameters; the double-drive dynamic generation module is used for generating a static model and simulating dynamic evolution in combination with semantic vision; the achievement verification output module is used for converting the data into multimedia, verifying the similarity and judging the achievement; according to the invention, through semantic-vision dual-drive dynamic generation and multi-dimensional verification of a closed loop, it is ensured that digital products fit historical logic of cultural relics; meanwhile, through multi-source data integration, semantic optimization and gene pool accurate matching, full-process data and semantic optimization is achieved, and the problems of semantic vision disjunction, data fragmentation and the like in traditional text travel digitization are solved.

Description

Technical Field

[0001] This invention relates to the field of digital cultural tourism technology, specifically to an intelligent production system and method for digital cultural tourism products based on a large model. Background Technology

[0002] With the accelerated digital transformation of the cultural and tourism industry, the demand for high-quality digital products in cultural relic protection, cultural display, and intangible cultural heritage transmission is becoming increasingly urgent, especially requiring technical solutions that balance "cultural accuracy" and "dynamic interactivity." Currently, a large amount of multi-source cultural and tourism data has accumulated, and the maturity of artificial intelligence technology has made data transformation possible—technology such as large-scale models, semantic extraction, and generative modeling are gradually being applied in the cultural and tourism field. However, the industry has not yet formed a complete system that can achieve "deep synergy between semantic and visual features, and dynamic evolution that conforms to historical logic." The core demand for cultural and tourism digital products has shifted from simple visual presentation to "dynamic generation based on authentic cultural semantics." This requires technological integration to solve the key problem of "semantic guidance of vision and dynamic alignment with history," which has become the core direction for promoting the upgrade of cultural and tourism digitalization from "static display" to "dynamic and living inheritance."

[0003] The core flaw in traditional cultural tourism digital product generation technologies lies in their inability to achieve "semantic-visual synergy" and "dynamic evolution compliance," resulting in insufficient cultural accuracy. Firstly, traditional generative models focus solely on visual effect optimization, neglecting the directional guidance of semantic features in visual generation. Static models often exhibit a disconnect between visual and cultural semantics. Dynamic evolution simulations fail to incorporate factors such as artifact materials and historical environment to construct a scientific evolutionary logic, relying solely on simple timeline animations, which cannot accurately recreate the true historical changes of artifacts, violating the principle of cultural authenticity. Secondly, there is a lack of quantitative semantic verification and a closed-loop optimization mechanism for results. Semantic extraction lacks a reliable confidence assessment system, making it prone to label bias. Result verification relies heavily on subjective human judgment, lacking standardized multi-dimensional verification rules encompassing "semantics + visuals + historical logic," making it difficult to guarantee stable digital product quality and meet the stringent requirements for cultural accuracy in cultural tourism display and education scenarios. This is a core bottleneck that traditional technologies cannot overcome. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an intelligent production system and method for digital cultural products in the cultural tourism industry based on a large model. This invention integrates multi-source data through a data graph module to construct a knowledge graph; a semantic extraction module uses a fine-tuning model to accurately extract and verify semantic tags; a gene pool matching module achieves accurate association between visual features and semantic tags; a dual-drive generation module integrates semantic and visual features to generate static and dynamic models that conform to historical logic; and a results verification module ensures product quality through multi-factor verification. This invention solves the problems of semantic-visual disconnect and data fragmentation in traditional digital cultural tourism.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: On one hand, a smart production system for digital cultural products for cultural tourism based on a large model, the system comprising: Data Graph Module: Receives archaeological reports on cultural relics, oral recordings of intangible cultural heritage, high-definition photos of cultural relics, and historical environmental data. It standardizes the format of reports, transcribes and cleans audio recordings, denoises and corrects photos, and uses a graph database to construct a cultural tourism knowledge graph containing cultural relics, cultural symbols, historical scene entities, and their relationships. Semantic extraction module: Receives the above knowledge graph, extracts attribute-class and parameter-class semantic tags by fine-tuning the semantic extraction model, calculates the tag reliability using the cultural tourism semantic tag confidence calculation formula, manually verifies tags that do not meet the reliability standard, and outputs tags after they pass the verification. Gene library matching module: Receives the above semantic tags, retrieves the constructed visual gene library through word segmentation index, and matches the corresponding visual feature vectors and physical attribute parameters of cultural relics; Dual-drive generation module: Receives the above feature vectors and physical parameters, constructs a grid rule matrix using the cultural tourism semantic cell state transition formula and updates the cell state, inputs the improved generative large model, uses the cultural tourism semantic-visual cross-attention weight calculation formula to adjust the semantic and visual association strength, and generates an initial static model; then uses the cultural tourism cultural relics dynamic evolution rate calculation formula to determine the evolution strength, simulates dynamic evolution, and saves keyframes at intervals to obtain dynamic evolution data. Results Verification Output Module: Receives the above dynamic data, converts it into playable videos, interactive 3D models, and AR materials, uses the similarity verification formula for cultural and tourism dynamic results to calculate the degree of consistency between keyframes and archaeological reference data, judges the qualification of the results, and if it is not qualified, adjusts the parameters and returns to re-simulate.

[0006] Furthermore, the fine-tuned semantic extraction model in the semantic extraction module is configured as follows: BERT-Base is used as the training backbone network, with two layers of bidirectional LSTM and one layer of CRF connected downstream; the fine-tuning dataset is labeled data in the cultural tourism field, including semantically labeled samples of cultural relics, with a total data volume of ≥100,000 records; the fine-tuning uses the AdamW optimizer, with the learning rate initially set to 5e-5, decreasing by 10% every 500 iterations for dynamic adjustment, and the number of iterations is 800-1200. After fine-tuning, the model achieves an entity recognition accuracy of ≥97% on the cultural tourism semantic extraction validation set. The target semantic label probability and the second highest probability output by the model are used as inputs to the formula for calculating the confidence of cultural tourism semantic labels.

[0007] Furthermore, the formula for calculating the confidence level of the cultural tourism semantic tags in the semantic extraction module is as follows: ,in, The final reliability of semantic tags is used to measure the credibility of the semantic tag extraction results; To fine-tune the target semantic label probability output by the semantic extraction model, that is, the probability value that the model determines the current semantic label as the correct result; To fine-tune the second-highest probability semantic label output by the semantic extraction model, that is, the probability value of the semantic label that the model determines is most likely to be correct besides the target label; This is the second-highest probability correction coefficient, used to reduce the interference of the second-highest probability on the reliability calculation. For bronze ware and ancient building artifacts, it is taken as 0.8-0.9, and for modern and contemporary cultural and creative products, it is taken as 0.6-0.7. This is the semantic importance weight, used to distinguish the priority of different semantics. Core historical semantics have a value of ≥0.8, while secondary semantics have a value of ≤0.5. This is an adaptation factor for the cultural tourism field, used to adapt the semantic complexity of different types of cultural relics. For ancient cultural relics, a factor of 0.9-1.0 is used, and for modern and contemporary cultural relics, a factor of 0.7-0.8 is used. The reliability not meeting the standard indicates... If the value is less than 0.95, incremental model tuning is triggered. Each incremental tuning adds 100-200 semantic label samples that failed the validation.

[0008] Furthermore, the visual feature vector extraction process of the visual gene pool in the gene pool matching module is as follows: the ViT-L / 16 model is used to process high-definition photos of cultural relics. First, the photos are resized to a uniform size. The 12-layer Transformer encoder of the model outputs 1024-dimensional original features, which are then reduced to 512 dimensions by PCA and L2 normalized. The feature vectors and semantic labels are associated through a semantic word segmentation-feature vector mapping table. This mapping table is updated once a quarter in conjunction with the newly added cultural relics data. When updating, the semantic extraction model of the semantic extraction module is used to annotate the semantics of the newly added cultural relics.

[0009] Furthermore, the formula for calculating the state transition of the cultural tourism semantic cell in the dual-drive state generation module is as follows: ,in, For cells ( )exist The updated state at any given moment includes three dimensions: color, texture, and transparency, reflecting the attributes of the cell after dynamic changes. For cells ( The initial state at time t is taken from the visual feature vector mapping value output by the gene pool matching module, which is the starting state of dynamic evolution; The semantic driving coefficient is used to control the influence of semantic reliability on cell state updates. It is set to 0.1-0.2 for mural artifacts and 0.05-0.1 for bronze artifacts. The final reliability of the semantic labels calculated for the semantic extraction module is used to ensure that the cell state update conforms to the semantic logic; The time decay factor is used to simulate the weakening of the influence of semantics on the cell state over time, and is calculated as "1 - current evolution time / total evolution period". The weights are for 3×3 neighboring cells, used to measure the influence of neighboring cells on the state of the central cell. The weights are 0.1-0.2 for central neighboring cells and 0.025-0.05 for diagonal cells. The visual feature matching degree output by the gene pool matching module is used to ensure the visual consistency of the state update of neighboring cells and is calculated from the cosine similarity of the visual feature vectors. The neighbor cell index, ranging from 1 to 8, corresponds to the 8 neighbor cells in a 3×3 neighborhood excluding the center cell. When calculating the color dimension of the cell state S(t+1,i,j), if the result exceeds the RGB value range of 0-255, linear normalization is applied. The normalization formula is: ,in, The normalized cell color value is ensured to be within the valid range of RGB values; The original calculation result of the cell color dimension is the color value obtained through the cultural tourism semantic cell state transition formula; This is the minimum value among the cell color calculation results for this batch; This is the maximum value among the cell color calculation results for this batch.

[0010] Furthermore, the formula for calculating the cultural tourism semantic-visual cross-attention weight in the dual-drive generation module is as follows: ,in, is the semantic-visual cross-attention weight, used to quantify the degree of attention of semantic features to the visual feature generation process in the improved generative large model, with a value range of [0,1]. The activation function normalizes the raw scores obtained from attention calculations to the [0,1] interval, ensuring that the sum of weights conforms to the calculation logic. Its calculation method is as follows: ,in, For the first One original attention score; The sum of the exponents of all the original attention scores; To represent the semantic feature query vector, the semantic label output by the semantic extraction module is input into the fine-tuned BERT-Base model, and the output vector at the [CLS] position of the model is extracted. The vector dimension is 512. To represent the visual feature key vector, which is the visual feature vector obtained from the visual gene pool in the semantic extraction module, the vector dimension is 512; To represent visual feature key vectors The transpose matrix is ​​used to interact with the semantic feature query vector. Perform matrix multiplication to calculate the similarity between the two. To represent the vector dimension parameter, a value of 512 is used, and... , The dimension is consistent, and its function is to avoid gradient instability caused by excessively large matrix multiplication results by normalizing the denominator; The attention enhancement coefficient, ranging from [0.5, 1.0], is used to adjust the weighting of visual feature matching degree and semantic timeliness. The magnification effect; To represent the visual feature matching degree, it is calculated by the cosine similarity between the visual feature vector and the corresponding feature vector of the semantic label in the gene pool matching module, with a value range of [0,1], which is used to reflect the degree of matching between visual features and semantic labels; To represent the semantic timeliness weight, "T" is the identifier for "Time", which is used to reflect the historical time priority of the semantic tag. The semantic weight for ancient cultural and tourism elements is 1.0, and the semantic weight for modern cultural and tourism elements is 0.7-0.9.

[0011] Furthermore, the improved generative large model in the dual-drive generation module is based on the StyleGAN3 architecture, and its core configuration includes: Module Embedding: The logic for calculating the cross-attention weights of cultural tourism semantics and vision is integrated into layers 5-7 of the model generator, thus embedding the cross-attention weights of cultural tourism semantics and vision. As a directional control parameter for visual feature generation, it links semantic features with the visual generation process; Training data: A high-definition image dataset containing 100,000 to 200,000 cultural and tourism elements is used. The image annotation information consists of attribute-class and parameter-class semantic labels output by the semantic extraction module. Training strategy: Mixed precision training (FP16+FP32) is adopted, the batch size is set to 32-64, the optimizer is Adam, the initial learning rate is 5e-4, the learning rate decays by 50% every 20,000 iterations, and the total number of iterations is ≥100,000. Performance metrics: Structural similarity (SSIM) between the generated initial static model and the corresponding cultural relic reference image ≥ 0.92, semantic tag matching accuracy ≥ 95%, and time taken to generate a single 512×512 resolution model ≤ 2 seconds.

[0012] Furthermore, the formula for calculating the dynamic evolution rate of cultural relics in the dual-drive generation module is as follows: , where is the dynamic evolution rate of cultural relics, used to measure the intensity of changes in cultural relics over time, such as corrosion depth / year and pigment fading rate / year; It is a pre-exponential factor, a fundamental rate parameter for the evolutionary response of cultural relics, and is determined by the material properties of the cultural relics; The activation energy is the energy barrier required for the evolutionary reaction of cultural relics, and different values ​​correspond to different evolutionary reactions. is the gas constant, a fixed physical constant, with a value of 8.314 J / (mol・K); The absolute ambient temperature is calculated from historical environmental data output by the data mapping module, using the formula " =The average annual temperature of the site is +273.15”; The material attenuation coefficient is used to measure the stability of the material of cultural relics. The more stable the material, the larger the coefficient. For tortoise shell cultural relics, the coefficient is 0.9-1.0, and for wooden cultural relics, the coefficient is 0.6-0.8. The cultural scene weight is used to distinguish the impact of different historical use scenarios of cultural relics on their evolution. The weight is 1.0 for sacrificial scenarios and 0.7-0.9 for daily use scenarios. The time interval for dynamic evolution is the time span for saving keyframes each time. For evolution cycles ≤ 100 years, take 1 month; for evolution cycles < 100 years < 500 years, take 3 months; and for evolution cycles > 500 years, take 6 months.

[0013] Furthermore, the formula for calculating the similarity of cultural and tourism dynamic results in the results verification output module is as follows: ,in, The final fit is used to measure the degree of consistency between the dynamic evolution results and the archaeological reference data; The visual feature vectors of the keyframes in the dynamic evolution data are extracted by the result verification output module using the ViT-L / 16 model from the keyframe images and then processed by PCA dimensionality reduction and L2 normalization. The visual feature vectors for archaeological reference data are taken from verified artifact feature data in archaeological research, and the processing method is similar to... Consistent; Cosine similarity of visual features is used to measure... and Spatial visual similarity; The time weight is used to distinguish the contribution of time series similarity and semantic reliability to the final match. The weight is 0.8 for time interval ≤ 1 year, 0.6 for 1 year < time interval ≤ 3 years, and 0.5 for time interval > 3 years. For time series similarity, it is used to measure the continuity of dynamically evolving sequences; The final reliability of the semantic tags calculated by the semantic extraction module is used to ensure that the consistency judgment conforms to the cultural semantic logic. When the consistency is ≥0.85, the result is deemed qualified. When the consistency is 0.8≤consistency<0.85, the result is sent to experts for review. When the consistency is <0.8, the parameters are adjusted and the result is returned to the dual-drive state generation module for re-simulation.

[0014] On the other hand, a method for intelligent production of digital cultural products for cultural tourism based on a large model includes: Data preprocessing and graph construction: Receive archaeological reports on cultural relics, oral recordings of intangible cultural heritage, high-definition photos of cultural relics, and historical environmental data. Standardize the format of the reports, transcribe and clean the audio recordings, denoise and correct the photos, and use a graph database to construct a cultural tourism knowledge graph containing cultural relics, cultural symbols, historical scene entities, and related relationships. Semantic extraction: Receive the above knowledge graph, extract attribute-class and parameter-class semantic tags by fine-tuning the semantic extraction model, calculate the tag reliability using the cultural tourism semantic tag confidence calculation formula, manually verify tags that do not meet the reliability standard, and output tags after they pass the test. Gene database matching: Receive the above semantic tags, retrieve the constructed visual gene database through word segmentation index, and match the corresponding visual feature vectors and physical attribute parameters of cultural relics; Dual-drive generation: Receive the above feature vectors and physical parameters, construct a grid rule matrix using the cultural tourism semantic cell state transition formula and update the cell state, input the improved generative large model, use the cultural tourism semantic-visual cross-attention weight calculation formula to adjust the semantic and visual association strength, and generate an initial static model; then use the cultural tourism cultural relics dynamic evolution rate calculation formula to determine the evolution strength, simulate dynamic evolution, and save key frames at intervals to obtain dynamic evolution data. Output verification: Receive the above dynamic data, convert it into playable videos, interactive 3D models and AR materials, use the similarity calculation formula for cultural and tourism dynamic results to calculate the degree of consistency between key frames and archaeological reference data, judge the qualification of the results, and if it is not qualified, adjust the parameters and return to re-simulate.

[0015] Compared with existing technologies, this intelligent production system and method for digital cultural products for cultural tourism based on a large model has the following beneficial effects: I. This invention integrates semantic and visual features through a dual-drive generation module to construct a generation model that simulates the dynamic evolution of cultural relics over time, ensuring that both the generated static model and dynamic results conform to the historical attributes and cultural logic of the cultural relics. The result verification and output module evaluates the qualification of the results through a multi-factor fusion similarity judgment method, and adjusts and re-optimizes parameters for unqualified results, forming a complete quality control closed loop. This design ensures that the final generated videos, interactive 3D models, and AR materials not only conform to the real historical characteristics of cultural relics but also have good interactive usability, effectively meeting the high-quality requirements of digital products in core scenarios such as cultural tourism display and cultural dissemination. By constructing a "semantic-visual dual-drive" dynamic generation mechanism and a multi-dimensional verification closed loop, it solves the key pain points of semantic and visual disconnect and dynamic results not conforming to historical logic in the generation of traditional cultural tourism digital products.

[0016] Second, this invention standardizes and integrates multi-source cultural tourism data through a data graph module, constructing a knowledge graph containing entity relationships to break down data silos; the semantic extraction module combines an optimized semantic extraction model and confidence calculation method with manual verification and continuous optimization mechanisms to significantly improve the reliability of semantic tags; the gene pool matching module achieves accurate association between visual features and semantic tags, ensuring that the source data and semantic information of digital product generation are consistent with the true attributes of cultural relics, further enhancing the accuracy of subsequent generation processes, and providing a stable and high-quality input foundation for the core dynamic generation mechanism. Through the linkage of multiple modules, the entire process of cultural tourism data from input to semantic transformation is optimized, solving the problems of data fragmentation and insufficient reliability of semantic tags in traditional cultural tourism digitization.

[0017] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0019] Figure 1 A flowchart of an intelligent production system for digital cultural products for cultural tourism based on a large model; Figure 2 This is a framework diagram of an intelligent production system for digital cultural products in the cultural tourism industry based on a large model. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] Example 1: Digital Exhibit Production of Shang Dynasty Bronze Ding for Museums This embodiment is applied to the construction of an online digital exhibition hall for a Shang Dynasty bronze ding (a type of ancient Chinese cooking vessel) at a provincial museum. The goal is to generate an interactive 3D model and a video showcasing the historical evolution of the bronze ding, meeting the audience's need for an immersive online understanding of the artifact. The specific implementation process is as follows: Figure 1 As shown: First, the data graph module is activated to receive the archaeological report of the Shang Dynasty bronze ding, high-resolution photos of the archaeological site, detailed photos of the ding's decorative patterns, historical environmental data of the Shang Dynasty, and oral recordings of the ding's restoration process. The archaeological report is standardized in format, unifying document structure and field naming. This avoids data confusion caused by different report formats and lays the foundation for subsequent multi-source data integration. The oral recordings are transcribed, removing redundant words and repetitive content, enabling accurate extraction of key information during the restoration process and reducing interference from invalid data. All high-resolution photos undergo noise reduction and color correction to ensure clear image details, providing high-quality image sources for subsequent visual feature extraction. Then, a cultural tourism knowledge graph is constructed using a graph database, establishing associations between the bronze ding and cultural symbols and historical scenes such as the late Shang Dynasty taotie-patterned ritual vessels and their uses, forming a complete multi-source data association system. This system breaks down data silos, allowing subsequent semantic extraction and visual matching to be conducted based on a unified data logic.

[0022] Next, the semantic extraction module is run, receiving the constructed knowledge graph. A fine-tuned semantic extraction model, using BERT-Base as the training backbone network and a downstream cascaded bidirectional LSTM and a 1-layer CRF, is employed to extract attribute-type and parameter-type semantic labels for the bronze tripod. The training dataset for this fine-tuned model is labeled data from the cultural tourism field, containing semantically labeled samples of cultural relics, with a total data volume ≥100,000 entries. The fine-tuning process uses the AdamW optimizer, with iterations controlled between 800-1200. After fine-tuning, the model achieves entity recognition accuracy ≥97% on the cultural tourism semantic extraction validation set. This specialized fine-tuning design for the cultural tourism field significantly improves the accuracy of semantic label extraction, avoiding misjudgments of cultural tourism terminology by general models. After label extraction, the reliability of each label is calculated using the cultural tourism semantic label confidence calculation formula. This formula quantifies the credibility of the labels, providing an objective basis for label quality screening. The cultural tourism semantic label confidence calculation formula is: ,in, For the final reliability of semantic tags; To fine-tune the target semantic label probability output by the semantic extraction model; To fine-tune the second-highest probability semantic label output by the semantic extraction model; This is the second highest probability correction coefficient; This is the semantic importance weight, used to distinguish the priority of different semantics. Core historical semantics have a value of ≥0.8, while secondary semantics have a value of ≤0.5. It is an adaptation factor for the cultural tourism field; for tags with a reliability of <0.95, manual verification is triggered. After review and correction by archaeological experts, qualified semantic tags are output. The manual verification process can further correct model deviations and ensure that the semantic tags fully conform to the true cultural attributes of the bronze tripod.

[0023] The process then proceeds to the gene pool matching module, where qualified semantic tags are received and the constructed visual gene pool is retrieved via word segmentation indexing. The visual feature vectors of this visual gene pool are extracted using the ViT-L / 16 model. First, high-resolution images of cultural relics are resized to a uniform size. The model's 12-layer Transformer encoder outputs 1024-dimensional original features, which are then reduced to 512 dimensions using PCA and L2 normalized. This feature extraction method preserves the core visual information of the images while reducing data dimensionality to improve subsequent matching efficiency. Feature vectors and semantic tags are associated through a semantic word segmentation-feature vector mapping table. This mapping table is updated quarterly with newly added cultural relic data, ensuring the timeliness of the gene pool and guaranteeing the matching of the latest and most accurate visual features. After retrieval, visual feature vectors and physical attribute parameters corresponding to a Shang Dynasty bronze ding (a type of ancient Chinese cooking vessel) are matched. This matching result provides accurate visual and physical input for the subsequent generation model, ensuring that the generated digital product accurately reflects the actual characteristics of the bronze ding.

[0024] Then, the dual-drive state generation module is activated to receive the matched feature vectors and physical parameters. First, a grid rule matrix is ​​constructed using the cultural tourism semantic cell state transition formula. The cultural tourism semantic cell state transition calculation formula is as follows: ,in, For cells ( )exist The updated state at any given moment includes three dimensions: color, texture, and transparency, reflecting the attributes of the cell after dynamic changes. For cells ( The initial state at time t is taken from the visual feature vector mapping value output by the gene pool matching module, which is the starting state of dynamic evolution; These are semantically driven coefficients; The final reliability of the semantic tags calculated for the semantic extraction module; This is the time decay factor; The weights are for 3×3 neighborhood cells; The visual feature matching degree output by the gene pool matching module; This is the index of the neighboring cells; when calculating the color dimension of the cell state S(t+1,i,j), if the result exceeds the RGB value range of 0-255, linear normalization is applied. The normalization formula is: ,in, The normalized cell color value; This is the original calculation result for the cell color dimension; This is the minimum value among the cell color calculation results for this batch; This represents the maximum value among the cell color calculation results. Using the visual feature vector mapping value as the initial state of the cell, the color, texture, and transparency dimensions of the cell are updated by combining semantic label reliability and visual feature matching degree. This formula deeply integrates semantic and visual information, providing a basic framework for static model generation that conforms to cultural logic and visual rules. The updated matrix is ​​input into an improved generative large-scale model based on StyleGAN3 architecture. Layers 5-7 of this model generator incorporate the logic for calculating the cultural tourism semantic-visual cross-attention weights. The formula for calculating the cultural tourism semantic-visual cross-attention weights is: ,in, For semantic-visual cross-attention weights; The activation function is calculated as follows: ,in, For the first One original attention score; The sum of the exponents of all the original attention scores; This represents a semantic feature query vector; To represent the visual feature key vector; To represent visual feature key vectors The transpose of the matrix; To represent the vector dimension parameter; To represent the attention enhancement coefficient; To represent the degree of visual feature matching; To represent the semantic timeliness weight, this formula adjusts the correlation strength between semantic and visual features, avoiding the visual-semantic disconnect in the generated model. This ensures the initial static 3D model closely matches the authentic attributes of the bronze tripod. The structural similarity between the generated static model and the bronze tripod reference image is ≥0.92, the semantic label matching accuracy is ≥95%, and the time to generate a single 512×512 resolution model is ≤2 seconds. These performance indicators guarantee the visual realism and generation efficiency of the static model. Next, the dynamic evolution rate calculation formula for cultural relics is used, combined with Shang Dynasty historical environment data and the material characteristics of the bronze tripod to determine the evolution intensity. This formula simulates the state change patterns of cultural relics in a real historical environment, ensuring the dynamic evolution process conforms to historical logic. The dynamic evolution rate calculation formula for cultural relics is as follows: Among them, represents the dynamic evolution rate of cultural relics; Pre-exponential factors; Activation energy; It is the gas constant; The absolute temperature of the environment; The material attenuation coefficient; Weighting based on cultural context; The time interval for dynamic evolution is defined as follows: keyframes are saved at 6-month intervals for evolution cycles greater than 500 years to obtain dynamic evolution data. Regularly saving keyframes can form a continuous record of the dynamic process, providing data support for the subsequent generation of evolution videos.

[0025] Finally, the results verification and output module is run to receive dynamic evolution data and convert it into playable historical evolution videos and interactive 3D models. The format conversion operation is adapted to the playback and interaction requirements of the online exhibition hall, allowing viewers to conveniently view the data on their devices. A similarity verification calculation formula for cultural tourism dynamic results is used to calculate the degree of agreement between the keyframe visual feature vectors and the archaeological reference data. This formula can quantify the quality of results from multiple dimensions, including visual, temporal, and semantic dimensions, avoiding the one-sidedness of single-dimensional verification. The similarity verification calculation formula for cultural tourism dynamic results is as follows: ,in, For the final fit; Visual feature vectors for keyframes of dynamically evolving data; The visual feature vectors for archaeological reference data are taken from verified artifact feature data in archaeological research, and the processing method is similar to... Consistent; Cosine similarity of visual features is used to measure... and Spatial visual similarity; Time weighting; For time series similarity; The semantic tag final reliability is calculated for the semantic extraction module; when the time interval is set to >3 years, the time weight is 0.5, and the final consistency calculation result is 0.91≥0.85, which is considered qualified. The qualified judgment standard can ensure that the output digital products meet the requirements of cultural accuracy; the video and 3D model are deployed to the museum's online exhibition hall. Viewers can rotate the 3D model to view the details and watch dynamic videos to understand the historical changes of the bronze tripod. This deployment can break the time and space limitations of offline exhibitions, expand the coverage of bronze tripod culture dissemination, and enhance the cultural experience of the audience.

[0026] In summary, this embodiment addresses the needs of creating online digital exhibits of Shang Dynasty bronze ding vessels for museums. It integrates multi-source cultural relic data and constructs a related knowledge graph through a data graph module, breaking down data silos. The semantic extraction module utilizes a cultural tourism-specific fine-tuning model and confidence calculations, combined with manual verification, to ensure semantic tags accurately reflect the bronze ding's authentic attributes. The gene pool matching module efficiently retrieves precise visual features and physical parameters. The dual-drive generation module achieves deep semantic-visual fusion through multi-formula collaboration, generating a highly similar static model and dynamic evolution data consistent with historical logic. The results verification module quantitatively evaluates and outputs qualified digital exhibits. This progressive process ensures both the cultural accuracy and visual authenticity of the digital exhibits and breaks down the spatial and temporal limitations of offline exhibitions through online deployment, effectively meeting the audience's need for an immersive understanding of bronze ding culture.

[0027] Example 2: Production of AR teaching materials for Suzhou embroidery peony patterns (an intangible cultural heritage) This embodiment is applied to an online teaching project of Suzhou embroidery techniques at a certain intangible cultural heritage protection center. The goal is to generate AR interactive materials featuring Suzhou embroidery peony patterns to help learners intuitively grasp the embroidery steps and cultural connotations. The specific implementation process is as follows: Figure 2 As shown: First, the data mapping module is activated to receive audio recordings of intangible cultural heritage inheritors demonstrating Suzhou embroidery peony patterns, high-definition photos of different embroidery stages, historical records of surviving Qing Dynasty Suzhou embroidery peony patterns, and environmental data from modern Suzhou embroidery teaching. The historical records are standardized in format, and cultural information such as the peony pattern symbolizing wealth and the traditional color scheme dominated by red and green is systematically organized, providing a complete cultural background for subsequent semantic extraction. The audio recordings are transcribed to extract technical descriptions such as the stitching sequence, accurately preserving the core technical experience accumulated by the inheritors and preventing the loss of technical information. The high-definition photos are denoised and corrected to highlight the details and textures of the embroidery stitches, providing a clear image foundation for subsequent visual feature extraction and matching, ensuring that stitch details can be accurately identified. By using a graph database to construct a cultural tourism knowledge graph, the entity of Suzhou embroidery peony pattern is linked with entities such as the Qing Dynasty's inherited patterns, stitching techniques, the cultural connotations of wealth and nobility, and the skills of the inheritors, forming a data system for intangible cultural heritage skills. This system can connect cultural connotations, skill steps, and historical background, providing data support for the subsequent generation of teaching materials that are both culturally relevant and practical.

[0028] Next, the semantic extraction module is run, receiving the knowledge graph. Using a fine-tuned semantic extraction model with the same architecture as in Example 1, attribute-type semantic tags and parameter-type semantic tags for the Suzhou embroidery peony pattern are extracted. The model training dataset also meets the requirement of a total data volume ≥ 100,000 entries, and the entity recognition accuracy after fine-tuning is ≥ 97%. This mature model architecture ensures accurate extraction of professional terms in the Suzhou embroidery field, avoiding errors in teaching material parameters due to terminology misjudgment. After tag extraction, reliability is calculated using the cultural tourism semantic tag confidence calculation formula. This formula objectively assesses tag credibility and filters out tags that may have biases. For tags with a reliability < 0.95, such as 7 stitches per centimeter, intangible cultural heritage inheritors are invited to manually review them. After correcting them to 8 stitches per centimeter, qualified semantic tags are output. The inheritor's review ensures that the tag parameters are completely consistent with the actual embroidery techniques, laying the foundation for generating accurate teaching demonstration materials later.

[0029] Then, the module enters the gene bank matching module, receives qualified semantic tags, and retrieves the visual gene bank through word segmentation indexing. This gene bank has stored a large number of visual feature vectors of Suzhou embroidery patterns. The feature vector extraction method is consistent with that in Example 1. The unified extraction method can ensure the compatibility and accuracy of feature vectors, facilitating cross-scene matching. The mapping table is updated quarterly and has included the latest collected Suzhou embroidery intangible cultural heritage materials. Regular updates can ensure that the gene bank contains the latest Suzhou embroidery technique features, avoiding matching outdated or inaccurate visual information. After retrieval, the visual feature vector and physical attribute parameters corresponding to the overlay stitch of the Suzhou embroidery peony pattern are matched. This matching result can provide accurate visual and physical basis for the subsequent generation model, ensuring that the stitch details and color matching effects presented in the AR material are consistent with the real Suzhou embroidery techniques.

[0030] Subsequently, the dual-drive generation module is activated, receiving feature vectors and physical parameters. First, a grid rule matrix is ​​constructed using the cultural tourism semantic cell state transition formula. The visual feature vector mapping value serves as the initial cell state, and the cell state is updated by combining semantic label reliability and visual feature matching degree. This formula deeply integrates the cultural semantics and visual elements of Suzhou embroidery, ensuring that the generated static model conforms to visual rules while also containing cultural connotations. The matrix is ​​then input into an improved generative large-scale model. The model adjusts the correlation strength between the symbolism of the peony pattern and the visual presentation of the needlework through the cultural tourism semantic-visual cross-attention weight calculation logic. This avoids the problem of AR materials only displaying needlework without conveying meaning, ensuring that the generated initial static model of the Suzhou embroidery peony pattern possesses both technical accuracy and cultural expressiveness. The model structure similarity is ≥0.92, and the semantic matching accuracy is ≥95%. These indicators guarantee the visual realism and cultural fit of the static model. Then, the dynamic evolution rate calculation formula of cultural relics is adopted, and the evolution intensity is determined by combining the environmental data of the embroidery process and the material characteristics of the embroidery thread. This formula can simulate the state changes in the real embroidery process and ensure that the dynamic demonstration conforms to the actual embroidery logic. The time interval is 1 month according to the evolution cycle ≤ 100 years. Here, it is analogous to the embroidery step interval, which is 1 hour as the simulation interval to save key frames and obtain dynamic embroidery process data. The reasonable time interval can clearly present the details of each step of embroidery, which makes it easier for learners to gradually understand the skill steps.

[0031] Finally, the output verification module is run to convert the dynamic data into AR teaching materials. Learners can trigger a dynamic embroidery demonstration by scanning a specified pattern. The AR format conversion enables an interactive scanning demonstration mode, enhancing the intuitiveness and fun of the teaching. A similarity verification formula for cultural tourism dynamic results is used to compare the visual features of keyframes with the actual embroidery steps of intangible cultural heritage inheritors in videos. This formula verifies the consistency between the dynamic demonstration and the real skills from both visual and temporal dimensions, ensuring the accuracy of the teaching materials' techniques. When the time interval is ≤1 year, the time weight is 0.8, and the final consistency is ≥0.85, indicating the result is qualified. This qualification standard ensures that the AR materials meet the strict requirements for the accuracy of techniques in teaching scenarios. By integrating AR materials into the online teaching platform, learners can scan the embroidery fabric with their mobile phones to watch a live dynamic embroidery demonstration. Clicking on any step will allow them to view the corresponding semantic explanation. For example, this step uses the nested stitch technique, and it is important to ensure that the direction of the stitches matches the texture of the petals to reflect the three-dimensionality of the peony. This integration allows AR materials to be deeply integrated with the teaching scenario, effectively solving the problem of abstract and difficult-to-understand needlework techniques in traditional Suzhou embroidery teaching. It helps learners quickly master the embroidery skills and cultural connotations of the peony pattern in Suzhou embroidery.

[0032] In summary, this embodiment focuses on the creation of AR teaching materials for Suzhou embroidery peony patterns, an intangible cultural heritage. It is based on a data mapping module that organizes historical and cultural data and techniques of Suzhou embroidery and constructs a relational system; a semantic extraction module accurately extracts semantic tags such as stitches and color schemes, ensuring parameter accuracy through confidence calculation and verification by inheritors; a gene bank matching module obtains visual and physical parameters that match the stitching techniques; a dual-drive generation module integrates semantic and visual data to generate static models and dynamic embroidery data that combine technical accuracy and cultural expressiveness; and a results verification module outputs AR materials after verification. This process transforms Suzhou embroidery techniques into intuitive and interactive materials, addressing the pain point of abstract stitches in traditional teaching. Deployed through an online teaching platform, it helps learners quickly master embroidery techniques and cultural connotations, providing effective technical support for the living transmission of intangible cultural heritage.

[0033] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A smart production system for digital cultural products for cultural tourism based on a large model, characterized in that, The system includes: Data Graph Module: Receives archaeological reports on cultural relics, oral recordings of intangible cultural heritage, high-definition photos of cultural relics, and historical environmental data. It standardizes the format of reports, transcribes and cleans audio recordings, denoises and corrects photos, and uses a graph database to construct a cultural tourism knowledge graph containing cultural relics, cultural symbols, historical scene entities, and their relationships. Semantic extraction module: Receives the above knowledge graph, extracts attribute-class and parameter-class semantic tags by fine-tuning the semantic extraction model, calculates the tag reliability using the cultural tourism semantic tag confidence calculation formula, manually verifies tags that do not meet the reliability standard, and outputs tags after they pass the verification. Gene library matching module: Receives the above semantic tags, retrieves the constructed visual gene library through word segmentation index, and matches the corresponding visual feature vectors and physical attribute parameters of cultural relics; Dual-drive generation module: Receives the above feature vectors and physical parameters, constructs a grid rule matrix using the cultural tourism semantic cell state transition formula and updates the cell state, inputs the improved generative large model, uses the cultural tourism semantic-visual cross-attention weight calculation formula to adjust the semantic and visual association strength, and generates an initial static model; then uses the cultural tourism cultural relics dynamic evolution rate calculation formula to determine the evolution strength, simulates dynamic evolution, and saves keyframes at intervals to obtain dynamic evolution data. Results Verification Output Module: Receives the above dynamic data, converts it into playable videos, interactive 3D models, and AR materials, uses the similarity verification formula for cultural and tourism dynamic results to calculate the degree of consistency between keyframes and archaeological reference data, judges the qualification of the results, and if it is not qualified, adjusts the parameters and returns to re-simulate.

2. The intelligent production system for digital cultural products for cultural tourism based on a large model as described in claim 1, characterized in that, The fine-tuned semantic extraction model in the semantic extraction module is configured as follows: BERT-Base is used as the training backbone network, with a downstream cascade of two layers of bidirectional LSTM and one layer of CRF; the fine-tuning dataset is labeled data in the cultural tourism field, including semantically labeled samples of cultural relics, with a total data volume of ≥100,000 records; the fine-tuning uses the AdamW optimizer, with the learning rate initially set to 5e-5, decreasing by 10% every 500 iterations for dynamic adjustment, and the number of iterations is 800-1200. After fine-tuning, the model achieves an entity recognition accuracy of ≥97% on the cultural tourism semantic extraction validation set. The target semantic label probability and the second-highest probability output by the model are used as inputs to the formula for calculating the confidence of cultural tourism semantic labels.

3. The intelligent production system for digital cultural products for cultural tourism based on a large model as described in claim 1, characterized in that, The formula for calculating the confidence level of cultural and tourism semantic tags in the semantic extraction module is as follows: ,in, For the final reliability of semantic tags; To fine-tune the target semantic label probability output by the semantic extraction model; To fine-tune the second-highest probability semantic label output by the semantic extraction model; This is the second highest probability correction coefficient; This is the semantic importance weight, used to distinguish the priority of different semantics. Core historical semantics have a value of ≥0.8, while secondary semantics have a value of ≤0.

5. Adaptation factors for the cultural and tourism sector.

4. The intelligent production system for digital cultural products for cultural tourism based on a large model as described in claim 1, characterized in that, The visual feature vector extraction process of the visual gene bank in the gene bank matching module is as follows: The ViT-L / 16 model is used to process high-definition photos of cultural relics. First, the photos are resized to a uniform size. The 12-layer Transformer encoder of the model outputs 1024-dimensional original features, which are then reduced to 512 dimensions by PCA and L2 normalized. The feature vectors and semantic labels are associated through a semantic word segmentation-feature vector mapping table. This mapping table is updated once a quarter in combination with the data of newly added cultural relics. When updating, the semantic extraction model of the semantic extraction module is used to annotate the semantics of the newly added cultural relics.

5. The intelligent production system for digital cultural products for cultural tourism based on a large model according to claim 1, characterized in that, The formula for calculating the state transition of the cultural tourism semantic cell in the dual-drive generation module is as follows: ,in, For cells ( )exist The updated state at any given moment includes three dimensions: color, texture, and transparency, reflecting the attributes of the cell after dynamic changes. For cells ( The initial state at time t is taken from the visual feature vector mapping value output by the gene pool matching module, which is the starting state of dynamic evolution; These are semantically driven coefficients; The final reliability of the semantic tags calculated for the semantic extraction module; This is the time decay factor; The weights are for 3×3 neighborhood cells; The visual feature matching degree output by the gene pool matching module; This is the index of the neighboring cells; when calculating the color dimension of the cell state S(t+1,i,j), if the result exceeds the RGB value range of 0-255, linear normalization is applied. The normalization formula is: ,in, The normalized cell color value; This is the original calculation result for the cell color dimension; This is the minimum value among the cell color calculation results for this batch; This is the maximum value among the cell color calculation results for this batch.

6. The intelligent production system for digital cultural products for cultural tourism based on a large model according to claim 1, characterized in that, The formula for calculating the cultural tourism semantic-visual cross-attention weight in the dual-drive generation module is as follows: ,in, For semantic-visual cross-attention weights; The activation function is calculated as follows: ,in, For the first One original attention score; The sum of the exponents of all the original attention scores; This represents a semantic feature query vector; To represent the visual feature key vector; To represent visual feature key vectors The transpose of the matrix; To represent the vector dimension parameter; To represent the attention enhancement coefficient; To represent the degree of visual feature matching; This represents the semantic timeliness weight.

7. The intelligent production system for digital cultural products for cultural tourism based on a large model according to claim 1, characterized in that, The improved generative large model in the dual-drive generation module is based on StyleGAN3 architecture, and its core configuration includes: Module Embedding: The logic for calculating the cross-attention weights of cultural tourism semantics and vision is integrated into layers 5-7 of the model generator, thus embedding the cross-attention weights of cultural tourism semantics and vision. As a directional control parameter for visual feature generation, it links semantic features with the visual generation process; Training data: A high-definition image dataset containing 100,000 to 200,000 cultural and tourism elements is used. The image annotation information consists of attribute-class and parameter-class semantic labels output by the semantic extraction module. Training strategy: Mixed precision training is adopted, the batch size is set to 32-64, the optimizer is Adam, the initial learning rate is 5e-4, the learning rate decays by 50% every 20,000 iterations, and the total number of iterations is ≥100,000. Performance metrics: The structural similarity between the generated initial static model and the corresponding cultural relic reference image is ≥0.92, the semantic tag matching accuracy is ≥95%, and the time to generate a single 512×512 resolution model is ≤2 seconds.

8. The intelligent production system for digital cultural products for cultural tourism based on a large model according to claim 1, characterized in that, The formula for calculating the dynamic evolution rate of cultural relics in the dual-drive generation module is as follows: , where is the rate of dynamic evolution of cultural relics; Pre-exponential factors; It is the activation energy; It is the gas constant; The absolute temperature of the environment; The material attenuation coefficient; Weighting based on cultural context; This represents the time interval for dynamic evolution.

9. The intelligent production system for digital cultural products for cultural tourism based on a large model according to claim 1, characterized in that, The formula for calculating the similarity of cultural and tourism dynamic results in the results verification output module is as follows: ,in, For the final fit; Visual feature vectors for keyframes of dynamically evolving data; The visual feature vectors for archaeological reference data are taken from verified artifact feature data in archaeological research, and the processing method is similar to... Consistent; Cosine similarity of visual features is used to measure... and Spatial visual similarity; Time weighting; For time series similarity; The final reliability of the semantic tags calculated for the semantic extraction module.

10. A method for intelligent production of digital cultural products for cultural tourism based on a large model, the method being applicable to the intelligent production system for digital cultural products for cultural tourism based on a large model as described in any one of claims 1-9, characterized in that, The method includes: Data preprocessing and graph construction: Receive archaeological reports on cultural relics, oral recordings of intangible cultural heritage, high-definition photos of cultural relics, and historical environmental data. Standardize the format of the reports, transcribe and clean the audio recordings, denoise and correct the photos, and use a graph database to construct a cultural tourism knowledge graph containing cultural relics, cultural symbols, historical scene entities, and related relationships. Semantic extraction: Receive the above knowledge graph, extract attribute-class and parameter-class semantic tags by fine-tuning the semantic extraction model, calculate the tag reliability using the cultural tourism semantic tag confidence calculation formula, manually verify tags that do not meet the reliability standard, and output tags after they pass the test. Gene database matching: Receive the above semantic tags, retrieve the constructed visual gene database through word segmentation index, and match the corresponding visual feature vectors and physical attribute parameters of cultural relics; Dual-drive generation: Receive the above feature vectors and physical parameters, construct a grid rule matrix using the cultural tourism semantic cell state transition formula and update the cell state, input the improved generative large model, use the cultural tourism semantic-visual cross-attention weight calculation formula to adjust the semantic and visual association strength, and generate an initial static model; then use the cultural tourism cultural relics dynamic evolution rate calculation formula to determine the evolution strength, simulate dynamic evolution, and save key frames at intervals to obtain dynamic evolution data. Output verification: Receive the above dynamic data, convert it into playable videos, interactive 3D models and AR materials, use the similarity calculation formula for cultural and tourism dynamic results to calculate the degree of consistency between key frames and archaeological reference data, judge the qualification of the results, and if it is not qualified, adjust the parameters and return to re-simulate.

Images