Digital reconstruction system and method for cultural and tourism resources based on generative intelligent model

Abstract

The application provides a digital reconstruction system and method of cultural and tourism resources based on a generative intelligent model, relates to the technical field of intelligent model generation, analyzes the features of reconstruction requirements of a target historical scene, obtains constraint conditions of era features in a visual restoration process of the historical scene, and constructs a cultural knowledge graph; the target historical scene is reconstructed under constraints, a digital painting of the target historical scene is obtained, the digital painting is verified for consistency in eras through the cultural knowledge graph, and a consistency verification result of era features in the digital painting is obtained; when the consistency verification result is passed, the digital painting of the target historical scene is output; when the consistency verification result is not passed, era conflict features of the digital painting are identified; the digital painting is corrected for conflicts through the era conflict features, and a corrected digital painting is obtained. The application can inhibit the false generation of cross-era features of a cultural and tourism resource generation model, thereby reducing the generation probability of the generation model in an era mismatch.

Description

Technical Field

[0001] This application relates to the field of intelligent model generation technology, and more specifically, to a system and method for digital reconstruction of cultural and tourism resources based on generative intelligent models. Background Technology

[0002] With the deep integration of digital technology and the cultural industry, intelligent model generation technology, represented by generative intelligent models, has become the core engine driving the digital restoration and innovation of historical and cultural heritage. This technology can automatically generate high-resolution, high-fidelity visual images based on natural language descriptions or conceptual sketches, greatly improving the creative efficiency and expressive dimensions of the digital reconstruction of cultural and tourism resources such as historical scenes and cultural relics. It provides strong technical support for digital museums, immersive cultural and tourism experiences, and cultural heritage education, and is promoting the development of cultural formats towards visualization, interactivity, and intelligence.

[0003] However, in current technological practices, the use of generative models for historical scene reconstruction faces a serious problem of uncontrolled temporal accuracy. Generative models are typically trained on massive, multi-source, cross-era image data, and their internal representations are prone to mixing visual features from different historical periods. This often leads to temporal confusion and "time travel" in the generated results, such as architectural styles, clothing styles, artifact styles, and even lifestyle customs. This seriously damages the authenticity and cultural seriousness of historical restoration. Existing methods mostly focus on cue word engineering or single style transfer, lacking the ability to systematically model and utilize the complex and unstructured multi-era historical knowledge. They cannot effectively constrain temporal characteristics during the generation process, nor do they have a reliable mechanism for automated, fine-grained temporal consistency verification of the generated results. This makes the digital results of historical scenes output by generative models often have an imperceptible risk of temporal misalignment, restricting their high-quality application in professional cultural tourism and academic research fields. Therefore, how to suppress the misgeneration of cross-era features by cultural tourism resource generation models, thereby reducing the probability of temporal misalignment in the generation models, has become a difficult problem for the industry. Summary of the Invention

[0004] This application provides a system and method for digital reconstruction of cultural and tourism resources based on generative intelligent models, which can suppress the erroneous generation of cross-era features in the generation model of cultural and tourism resources, thereby reducing the probability of generation model misalignment.

[0005] Firstly, this application provides a method for digital reconstruction of cultural and tourism resources based on a generative intelligent model, the method comprising the following steps:

[0006] Obtain users' needs for reconstructing target historical scenarios and historical document data from different eras;

[0007] The reconstruction requirements are analyzed to obtain the constraints of the era characteristics in the process of visual restoration of historical scenes, and a cultural knowledge graph is constructed through all historical document data.

[0008] The target historical scene is reconstructed based on the constraints of the generative intelligent model and the characteristics of the era, resulting in a digital painting of the target historical scene. The digital painting is then verified for consistency with the era through the cultural knowledge graph, and the consistency verification result of the characteristics of the era within the digital painting is obtained.

[0009] When the consistency check result passes, the digital painting of the target historical scene is output; when the consistency check result fails, the period conflict characteristics of the digital painting are identified by the period style characteristics of the target historical scene.

[0010] The digital painting is corrected by applying the characteristics of the conflict of the era and the generative intelligent model to obtain the corrected digital painting.

[0011] In this embodiment, feature analysis is performed on the reconstruction requirements to obtain the constraints of era characteristics in the process of visual restoration of historical scenes, specifically including:

[0012] Keyword extraction of historical scene types is performed on the reconstruction requirements to obtain keywords of historical scene types in the process of visual restoration of historical scenes;

[0013] Semantic analysis of the era information is performed on the reconstruction requirements to obtain the semantic features of the era information in the process of visual restoration of historical scenes;

[0014] Based on the keywords of historical scene types and the semantic features of era information, the constraints of era characteristics in the process of visually restoring historical scenes are determined.

[0015] In this embodiment, constructing a cultural knowledge graph using all historical document data specifically includes:

[0016] Named entity recognition is performed on text-based historical document data to obtain different cultural knowledge entities;

[0017] Relationships are extracted from textual historical documents to obtain functional pairings and epochal affiliations among cultural knowledge entities.

[0018] Visual element annotation is performed on image-based historical document data to obtain the visual attribute relationships between cultural knowledge entities;

[0019] A cultural knowledge graph is constructed based on different cultural knowledge entities, the functional relationships and temporal affiliations between cultural knowledge entities, and the visual attribute relationships.

[0020] In this embodiment, the digital painting of the target historical scene is obtained by constraining and reconstructing the target historical scene based on the constraints of the generative intelligent model and the characteristics of the era. Specifically, this includes:

[0021] Initialize the generative intelligent model;

[0022] The constraints of the era's characteristics are transformed into positive and negative prompts that can be recognized by the generative intelligent model;

[0023] Based on a generative intelligent model, the target historical scene is reconstructed by combining the positive and negative prompts to obtain a digital painting of the target historical scene.

[0024] In this embodiment, the consistency verification of the digital painting's epochal characteristics using the cultural knowledge graph to obtain the consistency verification result of the epochal characteristics within the digital painting specifically includes:

[0025] The digital painting is deconstructed to obtain the constituent elements of different era scenes in the digital painting;

[0026] The cultural knowledge graph is then compared with all the elements that make up the historical scenes to obtain the consistency verification values ​​of the elements that make up the historical scenes.

[0027] The consistency check result of the epochal characteristics within the digital painting is determined based on all consistency check values.

[0028] In this embodiment, identifying the period conflict characteristics of the digital painting through the stylistic features of the target historical scene specifically includes:

[0029] Obtain the elements constituting the era scene that fail verification within the digital painting;

[0030] Acquiring cultural knowledge graphs;

[0031] Determine the stylistic characteristics of the target historical scene;

[0032] Based on the cultural knowledge graph and the stylistic features of the era, conflict identification is performed on the constituent elements of the era scene that fail the verification, and the era conflict features of the digital painting are obtained.

[0033] In this embodiment, the digital painting is modified by applying the aforementioned era conflict characteristics and generative intelligent model to obtain the modified digital painting, specifically including:

[0034] Based on the aforementioned characteristics of historical conflict, a conflict redraw instruction is generated for the elements constituting the historical scene in the digital painting that fail verification.

[0035] The digital painting is modified by using a generative intelligent model and conflict redrawing instructions for elements of the era scene that fail verification, resulting in a corrected digital painting.

[0036] In this embodiment, the consistency verification result represents the verification result of the degree to which the visual content of the era characteristics in the digital painting matches the historical era background.

[0037] In this embodiment, the "era conflict feature" refers to the conflict feature of digital painting that violates the norms of the era.

[0038] Secondly, this application provides a digital reconstruction system for cultural and tourism resources based on a generative intelligent model, used to execute a digital reconstruction method for cultural and tourism resources based on a generative intelligent model. The digital reconstruction system for cultural and tourism resources includes:

[0039] The data acquisition module is used to acquire users' reconstruction needs for target historical scenes and historical document data from different eras;

[0040] The feature extraction module is used to perform feature analysis on the reconstruction requirements, obtain the constraints of the era characteristics in the process of visual restoration of historical scenes, and construct a cultural knowledge graph through all historical document data;

[0041] The consistency verification module is used to reconstruct the target historical scene based on the constraints of the generative intelligent model and the characteristics of the era, to obtain a digital painting of the target historical scene. The digital painting is then verified for consistency with the era through the cultural knowledge graph to obtain the consistency verification result of the characteristics of the era within the digital painting.

[0042] The conflict identification module is used to output a digital painting of the target historical scene when the consistency verification result passes, and to identify the period conflict characteristics of the digital painting by the period style characteristics of the target historical scene when the consistency verification result fails.

[0043] The conflict correction module is used to correct the conflicts in the digital painting by using the conflict characteristics of the era and the generative intelligent model, so as to obtain the corrected digital painting.

[0044] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects:

[0045] The system acquires user requests for reconstructing a target historical scene and historical document data from different eras. It then performs feature analysis on the reconstruction requests to obtain constraints on the era characteristics during the visual restoration of the historical scene, and constructs a cultural knowledge graph using all historical document data. Based on the generative intelligent model and the constraints of the era characteristics, it performs constrained reconstruction of the target historical scene, resulting in a digital painting of the target historical scene. The cultural knowledge graph is used to verify the era consistency of the digital painting, yielding a consistency verification result for the era characteristics within the digital painting. If the consistency verification result passes, the digital painting of the target historical scene is output; if the consistency verification result fails, the era conflict characteristics of the digital painting are identified through the era style characteristics of the target historical scene. Finally, the era conflict characteristics and the generative intelligent model are used to correct the conflicts in the digital painting, resulting in a corrected digital painting.

[0046] Therefore, this application demonstrates that the digital painting can be corrected for conflicts using the aforementioned era conflict characteristics and generative intelligent model, resulting in a corrected digital painting. Firstly, by acquiring user reconstruction needs and historical document data from different eras, a clear target era and comprehensive knowledge base are established for the entire restoration process, defining the range of non-target era characteristics that need to be suppressed from the source. Secondly, feature analysis is performed on the reconstruction needs to extract the constraints of era characteristics, and a structured cultural knowledge graph is constructed using historical documents. This transforms the complex and unstructured multi-era knowledge into a computable and queryable feature rule base, providing precise guiding standards for the generation process and fundamentally solving the problems of mixed historical knowledge and its inability to be systematically utilized. Furthermore, during the generation stage, epochal constraints are directly applied to the generative intelligent model, guiding it to activate visual features of the target era. Simultaneously, a cultural knowledge graph is used to perform fine-grained epochal consistency verification on the output digital paintings. This achieves dual protection through proactive constraints before generation and precise detection after generation, effectively intercepting cross-epochal errors in elements such as architectural forms and clothing styles, and significantly reducing the risk of imperceptible epochal misalignment. Finally, when verification fails, the identified epochal conflict features are used to make targeted corrections using the generative model, enabling the system to adaptively reduce similar errors. This systematically improves the epochal accuracy of the generated model, suppresses cross-epochal feature confusion, and ensures the authenticity and reliability of the results of digital reconstruction of cultural and tourism resources.

[0047] In summary, the technical solution adopted in this application can suppress the erroneous generation of cross-era features in the cultural and tourism resource generation model, thereby reducing the probability of generation model misalignment. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this embodiment of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is an exemplary flowchart of a method for digital reconstruction of cultural and tourism resources based on a generative intelligent model provided in this application;

[0050] Figure 2 This is a flowchart illustrating the process of constructing a cultural knowledge graph provided in this application;

[0051] Figure 3 This is a flowchart illustrating the process of determining the consistency verification result based on the information provided in this application;

[0052] Figure 4 This is a module structure diagram of a digital reconstruction system for cultural and tourism resources based on a generative intelligent model, provided in this application. Detailed Implementation

[0053] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0054] This application provides a system and method for digital reconstruction of cultural and tourism resources based on a generative intelligent model. Its core is to acquire user requirements for the reconstruction of a target historical scene and historical document data from different eras; to perform feature analysis on the reconstruction requirements to obtain constraints on the era characteristics during the visual restoration of the historical scene; to construct a cultural knowledge graph using all historical document data; to perform constrained reconstruction of the target historical scene according to the generative intelligent model and the constraints on the era characteristics, resulting in a digital painting of the target historical scene; to perform era consistency verification on the digital painting using the cultural knowledge graph, obtaining a consistency verification result for the era characteristics within the digital painting; when the consistency verification result passes, the digital painting of the target historical scene is output; when the consistency verification result fails, the era conflict characteristics of the digital painting are identified through the era style characteristics of the target historical scene; and the conflict is corrected in the digital painting using the era conflict characteristics and the generative intelligent model, resulting in a corrected digital painting.

[0055] Example 1: To better understand the above technical solution, the following will provide a detailed description of the technical solution in conjunction with the accompanying drawings and specific implementation methods. (Refer to...) Figure 1 As shown in the figure, this is an exemplary flowchart of a method for digital reconstruction of cultural and tourism resources based on a generative intelligent model according to this embodiment of the application. The method for digital reconstruction of cultural and tourism resources includes the following steps:

[0056] In step S1, the user's reconstruction requirements for the target historical scene and historical document data from different eras are obtained.

[0057] It should be noted that the reconstruction requirement mentioned in this application refers to the textual information describing the target historical scene by the user; the historical document data refers to the historical document information of the relevant era.

[0058] In practice, a web-based interactive front-end interface provides form input boxes and natural language input areas to receive a piece of free text submitted by the user (for example, the user inputs: "I want a picture of a wine shop by West Lake in Lin'an City during the Southern Song Dynasty, with literati, elegant scholars, winding railings and waterside pavilions, in the style of Song Dynasty court paintings"). The user's submitted free text is converted into a JSON string and used as the user's reconstruction requirement for the target historical scene. Secondly, a backend data scheduling service is invoked, which is connected to a centralized historical and cultural big data platform. This platform has pre-integrated resources such as the "Basic Ancient Books Database of China", digital collection databases of museums across the country, and archaeological report databases of academic institutions, and has indexed them at multiple levels according to dynasty, region, and theme. The scheduling service parses the time and location keywords in the reconstruction requirement (such as "Southern Song Dynasty" and "Lin'an"), generates database query statements, and retrieves all relevant historical documents from the platform: including text data (such as the record of West Lake wine shops in "Menglianglu") and image data (such as high-definition scans of representative works of Song Dynasty court paintings). All retrieved documents are used as historical documents for the corresponding era.

[0059] In step S2, feature analysis is performed on the reconstruction requirements to obtain the constraints of the era characteristics in the process of visual restoration of historical scenes, and a cultural knowledge graph is constructed through all historical document data.

[0060] In this embodiment, the constraints of era characteristics in the process of visual restoration of historical scenes can be obtained by performing feature analysis on the reconstruction requirements using the following steps:

[0061] Keyword extraction of historical scene types is performed on the reconstruction requirements to obtain keywords of historical scene types in the process of visual restoration of historical scenes;

[0062] Semantic analysis of the era information is performed on the reconstruction requirements to obtain the semantic features of the era information in the process of visual restoration of historical scenes;

[0063] Based on the keywords of historical scene types and the semantic features of era information, the constraints of era characteristics in the process of visually restoring historical scenes are determined.

[0064] It should be noted that the keywords mentioned in this application refer to nouns that are the core constituent elements of historical scene types; the semantic features refer to the label features of the era style; and the constraints refer to the era-specific restrictions on the creation of content by the generative intelligent model during the visual restoration of historical scenes.

[0065] In specific implementation, firstly, the reconstruction requirements are input into a pre-trained named entity recognition model (e.g., a BERT model jointly trained on general Chinese corpora and historical texts. This model is based on the Transformer architecture and uses word embedding layers, multi-head self-attention mechanisms, and feedforward neural networks to achieve deep semantic encoding of the input text. Based on pre-training on general Chinese corpora, further domain-adaptive training is performed using historical corpora such as the *Twenty-Four Histories* and *Local Chronicles* to improve the recognition ability of ancient Chinese entities (such as building names, official titles, and artifacts). For a user's natural language description (e.g., "Lin'an Restaurant in the Southern Song Dynasty"), after word segmentation and vectorization, the model outputs a sequence of entities with BIO tags (e.g., "Lin'an -> Location", "Restaurant -> Building"), providing structured information for subsequent extraction of historical constraints. During training, masked language modeling and next-sentence prediction are used. In the fine-tuning stage, an annotated historical text set is used, with AdamW as the optimizer and a learning rate of 5e-5 to achieve accurate extraction of historical scene keywords. This model is then used to identify all entities belonging to "location," "building," and "person" in the reconstruction requirements. For entities categorized as "objects" and "activities," the identified entities, such as "Lin'an City," "West Lake," "restaurants," "literati," and "winding railings and waterside pavilions," are used as keywords for historical scene types. Secondly, the reconstruction requirements are input into a semantic rule parsing and mapping module. This module has a built-in historical dynasty epoch table, an ancient and modern place name comparison knowledge base, and an art style classification dictionary. First, the module uses rules to match the time phrases in the reconstruction requirements (e.g., "Southern Song Dynasty"), and queries the epoch table to map them to the precise date range label "1127-1279" and the dynasty code "Southern_Song." Then, it matches the location phrases (e.g., "Lin'an City"), queries the place name comparison database to map them to the standard historical geographical coordinate label "Song_Dynasty_Lin'an." Finally, it matches the style descriptions (e.g., "Song Dynasty court painting"), queries the style classification dictionary to map them to the visual style code "Song_Academy_Style." This set of structured labels (date range, dynasty code, geographical coordinates, style code) is packaged as semantic features of the era information.Finally, the obtained keywords and semantic features are input into a constraint synthesis engine. This engine synthesizes the data according to a fixed logical template of "era-region-core elements-visual style-negative constraints": the dynasty code and geographical coordinates from the semantic features are directly filled into the corresponding fields; the keyword list is categorized and organized and then used as the core element list; based on the style code, corresponding rendering parameters (such as brushstrokes and color saturation) are selected from a pre-set style template library; simultaneously, the engine automatically connects to a historical and cultural knowledge base based on the dynasty code and regional coordinates to retrieve elements that are impossible to appear in that era and region (such as "reinforced concrete" and "glass curtain wall") as the negative constraint list; finally, the engine outputs a standardized JSON object containing all the above fields, which serves as the constraint condition for the era features in the visual restoration of historical scenes.

[0066] Preferably, in this embodiment, a cultural knowledge graph is constructed using all historical document data, with reference to... Figure 2 As shown in the figure, this is a schematic diagram of the process of constructing a cultural knowledge graph in some embodiments of this application. The construction of the cultural knowledge graph in this embodiment can be achieved by the following steps:

[0067] In step S21, named entity recognition is performed on the text-based historical document data to obtain different cultural knowledge entities;

[0068] In step S22, relation extraction is performed on the historical document data of text type to obtain the functional matching relationship and era attribution relationship between cultural knowledge entities;

[0069] In step S23, visual element annotation is performed on historical document data of image type to obtain the visual attribute relationship between cultural knowledge entities;

[0070] In step S24, a cultural knowledge graph is constructed based on different cultural knowledge entities, the functional matching relationships and era affiliation relationships between cultural knowledge entities, and the visual attribute relationships.

[0071] It should be noted that, in this application, the cultural knowledge entities refer to specific objects with independent historical and cultural significance, such as specific buildings, artifacts, figures, systems, etc.; the functional matching relationship refers to the subordinate relationship between cultural knowledge entities in which they are used together in a specific context; the epochal affiliation relationship refers to the affiliation of cultural knowledge entities in the historical period in which they exist; the visual attribute relationship refers to the association between cultural knowledge entities in terms of visual characteristics; and the cultural knowledge graph refers to the graph used for reasoning about historical and cultural knowledge.

[0072] In specific implementation, first, a domain named entity recognition model fine-tuned on a large number of historical literature corpora is adopted (for example, a model based on the ERNIE architecture, which introduces a knowledge-enhanced masking strategy on the basis of BERT and can identify compound cultural entities such as "colorful building facade" as a whole unit instead of disassembling them into sub-words. The model structure includes an entity-aware attention layer and a knowledge fusion module. The input is the segmented text of historical literature, and the output is a list of entities with categories (for example, "lottery tube -> utensil", "colorful building facade -> building component"), which are directly used as the nodes of the cultural knowledge graph. During training, a large-scale historical literature corpus is used for pre-training and fine-tuning on the labeled entity dataset, and the cross-entropy loss function is adopted to improve the recognition accuracy of historical proper nouns and support the construction of the entity layer of the graph). After segmenting the text-based historical literature data using the domain named entity recognition model, each sentence is recognized one by one, and the phrases belonging to categories such as "architecture", "utensil", "clothing", "occupation", etc. (for example, "colorful building facade", "lottery tube", "hair toss", "hawker") are output together with their category labels. Each such phrase-label pair is used as a cultural knowledge entity; secondly, the original text sentences containing the recognized cultural knowledge entities are input into a joint relation extraction model (for example, a relation extraction model based on sequence annotation and attention mechanism. This model is based on sequence annotation and attention mechanism, encodes the text sequence through bidirectional LSTM, and uses multi-head self-attention to capture the long-distance dependencies between entities. The output layer uses a pointer network to annotate the relation types. The input is the sentence and the positions of the recognized entities, and the output is a triple (for example, "colorful building facade - belongs to - wine shop"), which is used to establish the functional collocation and era attribution relation edges in the cultural knowledge graph. The training data comes from the labeled historical literature relation set, and FocalLoss is used to alleviate the class imbalance. The number of attention heads is set to 8, and the hidden layer dimension is 512 to ensure the effective extraction of ancient Chinese grammar and implicit relations).The joint relation extraction model extracts the relations between cultural knowledge entities in a sentence according to preset relation types (such as "usage", "location", "belonging to"), and takes the extracted relation such as "Caillou Huanmen - belonging to -> Wine Shop" as a functional collocation relation. At the same time, it attaches a relation of "Entity Name - belonging to the era -> Southern Song Dynasty" to all entities extracted from the text-based historical literature data for the metadata of the historical literature data (such as the publication era "Southern Song Dynasty"), as the era attribution relation; then, for the image-based historical literature data (such as the high-definition scanned file of "耕织图" in the Song Dynasty), a visual large model combination process is adopted for processing: first, use an image segmentation model (such as Segment Anything Model) to perform general segmentation on the image to obtain multiple region masks, then use a visual entity classification model trained on an ancient painting dataset to classify and identify each mask region to obtain its category (such as "plow", "spinning wheel"), then, use a dense description generation model (such as BLIP-2 model) to generate a detailed natural language description for the identified region, and finally, extract keywords about shape, material, color, and pattern from a predefined "visual attribute - value" dictionary to establish associations for cultural knowledge such as "Plow - material feature -> wooden with iron blade", "Spinning wheel - structural feature -> hand-operated single spindle", and take such associations as visual attribute relations; finally, all cultural knowledge entities, functional collocation relations, era attribution relations, and visual attribute relations generated in the above steps are uniformly converted into the RDF triple format of "head entity - relation - tail entity", and use a knowledge graph construction tool (such as Apache Jena) to batch import all triples into a graph database (such as Neo4j), set the unique ID of the entity, establish an index, and finally form an interconnected and multi-hop query-supported knowledge network in the database, and use this database instance as the cultural knowledge graph.

[0073] In step S3, the target historical scene is constrained and reconstructed according to the constraint conditions of the generative intelligent model and the era characteristics to obtain a digital painting of the target historical scene, and the era consistency verification of the digital painting is performed through the cultural knowledge graph to obtain the consistency verification result of the era characteristics in the digital painting.

[0074] In this embodiment, the constraint reconstruction of the target historical scene according to the constraint conditions of the generative intelligent model and the era characteristics to obtain a digital painting of the target historical scene can be achieved by the following steps:

[0075] Initialize the generative intelligent model;

[0076] Convert the constraint conditions of the era characteristics into positive and negative prompt words recognizable by the generative intelligent model;

[0077] Based on a generative intelligent model, the target historical scene is reconstructed by combining the positive and negative prompts to obtain a digital painting of the target historical scene.

[0078] It should be noted that the generative intelligent model mentioned in this application refers to an artificial intelligence model that generates images; the positive prompt words refer to text strings used to describe and limit the specific elements, scenes, attributes and styles that should appear in the generated image; the negative prompt words refer to text strings used to prohibit the appearance of certain specific elements or styles in the generated image; and the digital painting displays an image of the target historical scene.

[0079] In practice, firstly, on the computing server, a pre-trained weight file of an open-source, basic large-scale text-to-image diffusion model (e.g., Stable Diffusion XL 1.0) is loaded. Simultaneously, a low-rank adapter weight trained using a dataset of images and texts related to ancient Chinese architecture, clothing, and artifacts is loaded. The pre-trained weight file and the low-rank adapter weights are merged in memory, completing the initialization of a finely tuned generative intelligent model specifically for generating historical and cultural scenes. This generative intelligent model is based on the open-source SDXL architecture, including a CLIP text encoder, a UNet denoising network, and a VAE codec. Domain adaptation is achieved by injecting a low-rank adapter, which inserts a rank-4 trainable parameter matrix into the UN. The cross-attention layer of et is linearly merged with the original weights to form a fine-tuned model specifically for historical scene generation. Training uses a dataset of paired images and texts of ancient Chinese architecture, clothing, and artifacts. The objective function is a combination of reconstruction and perceptual loss, with a batch size of 4, a learning rate of 1e-4, and the optimizer Adam. After 20,000 iterations, the generative intelligent model can generate digital paintings with consistent styles based on historical constraints. Secondly, by querying a pre-defined "history-style" mapping table, the "history" field value in the constraints (e.g., "Southern Song Dynasty") is converted into the style descriptor "SouthernSong dynasty style, Chinese classical painting". Similarly, by querying a "place name-environment" mapping table, the "location" field value in the constraints (e.g., "by the West Lake in Lin'an, with lakeside") is converted into the environment descriptor "by the West Lake in Lin'an, with lakeside". The script first selects "view". Then, it merges these two descriptive words with the direct translations of the "core elements" list in the constraints (e.g., "restaurant", "literati") to form a comprehensive description. This description is then input into an open-source cue word optimization tool (e.g., Gustavosta's Phraser) to automatically expand it into a detailed long English text, which is then used as a positive cue word. Simultaneously, the script directly translates the "negative constraints" list in the constraints into English phrases and uses them as negative cue words. Finally, the initialized generative intelligent model is run, using the positive and negative cue words as core inputs. To precisely control the scene composition, a spatial layout control network (e.g., ControlNet) is used. The simplified line drawing generated based on the "core elements" and spatial relationships is input into the model as additional spatial conditions. Generation parameters (e.g., sampling steps, size) are set, and the iterative denoising inference process of the model is initiated to generate a batch of candidate images. An image filter (e.g., sorting based on CLIP scores of cue word-image matching degree) selects the image with the highest matching degree from the candidate images, which is then used as the digital painting of the target historical scene.

[0080] Preferably, in this embodiment, the digital painting is subjected to a time consistency check using the cultural knowledge graph to obtain a consistency check result for the time characteristics within the digital painting, which is then referenced. Figure 3 As shown in the figure, this is a flowchart illustrating the process of determining the consistency verification result in some embodiments of this application. In this embodiment, the consistency verification result can be determined by the following steps:

[0081] In step S31, the digital painting is deconstructed to obtain the constituent elements of different era scenes in the digital painting;

[0082] In step S32, the cultural knowledge graph is checked for consistency with all the elements that make up the historical scenes to obtain consistency check values ​​for the elements that make up the historical scenes.

[0083] In step S33, the consistency verification result of the epochal characteristics within the digital painting is determined based on all the consistency verification values.

[0084] It should be noted that, in this application, the elements constituting the era scene represent the visual features describing object instances within the era scene; the consistency verification value represents the degree of matching between the attribute description of the elements constituting the era scene and the corresponding entity record in the cultural knowledge graph; and the consistency verification result represents the verification result of the degree of fit between the visual content of the era features in the digital painting and the historical era background.

[0085] In specific implementation, firstly, the digital painting is input into a pre-trained YOLOv8 segmentation model (this segmentation model is based on the CSPDarkNet backbone network and PANet feature pyramid structure to achieve end-to-end panoramic segmentation capability. The input is a 1024×1024 pixel digital painting image, and the output is a mask, category label (e.g., "pavilion" or "figure") and its visual feature vector for each instance. These instances are defined as "elements of the era scene" for subsequent consistency verification with the cultural knowledge graph. During the training phase, a labeled historical scene segmentation dataset is used, and data augmentation methods such as random cropping, color dithering, and overlay of ancient painting textures are adopted to improve the robustness of the model in segmenting historical elements). The digital painting is segmented panoramically using a pre-trained YOLOv8 segmentation model to identify all salient object instances (e.g., "a pavilion," "a wooden boat," "several figures"), and a category label, bounding box coordinates, and visual feature vector are generated for each physical instance. The combination of the category label, bounding box coordinates, and visual feature vector of each physical instance is used as a period scene component in the digital painting, thus obtaining different period scene components in the digital painting. Secondly, for each period scene component, a search and comparison are performed in the constructed cultural knowledge graph: using the category label of the period scene component (e.g., "pavilion") as the query keyword, a Cypher query is executed in a graph database (e.g., Neo4j) to find all "pavilion" type entities with the "period" attribute set to the target era (e.g., "Southern Song Dynasty"), and these entities are obtained. The system first sets text descriptions of the entities and associated reference image feature vectors. Then, it calculates the cosine similarity between the visual feature vectors of the elements constituting the historical scene and the reference image feature vectors associated with all candidate entities in the cultural knowledge graph. The highest value is taken as the visual matching degree. At the same time, a text similarity model (such as the Sentence-BERT model) is loaded. This model adopts a dual encoder structure. Based on BERT, the input text pairs are encoded into sentence vectors respectively. The semantic matching degree is evaluated by calculating the cosine similarity. The input is the text description of the elements constituting the historical scene (such as "two-story wooden pavilion") and the standard description set of entities in the cultural knowledge graph. The output is a similarity score between 0 and 1, which is used as a component of the consistency check value. The model is fine-tuned on the historical text semantic matching dataset and uses a contrastive learning loss function. The negative samples come from entity descriptions from different eras to enhance the model's sensitivity to semantic differences related to different eras.Next, a text similarity model is used to calculate the semantic similarity between the text descriptions of the constituent elements of the era scene and the text description set of candidate entities. The highest value is taken as the semantic matching degree. The visual matching degree and semantic similarity are weighted and averaged (e.g., each with a weight of 0.5) to obtain a value between 0 and 1. This value is used as the consistency verification value of the constituent elements of the era scene, thus obtaining the consistency verification value of each constituent element of the era scene. Finally, an overall pass threshold is set (e.g., the average consistency verification value of all constituent elements of the era scene must be greater than 0.75, and the consistency verification value of a single constituent element of the era scene must not be lower than 0.4). The average consistency verification value of all constituent elements of the era scene is calculated and the minimum value is found. If the average value is greater than 0.75 and the minimum value is greater than 0.4, the consistency verification result containing the "pass" state is output; if the threshold condition is not met, the consistency verification result containing the "fail" state is output.

[0086] In step S4, when the consistency check result passes, a digital painting of the target historical scene is output; when the consistency check result fails, the period conflict characteristics of the digital painting are identified through the period style characteristics of the target historical scene.

[0087] In practice, when the consistency verification result passes, the digital painting of the target historical scene is directly output and used as the final output of the generative intelligent model.

[0088] In this embodiment, identifying the period conflict characteristics of the digital painting based on the stylistic features of the target historical scene can be achieved through the following steps:

[0089] Obtain the elements constituting the era scene that fail verification within the digital painting;

[0090] Acquiring cultural knowledge graphs;

[0091] Determine the stylistic characteristics of the target historical scene;

[0092] Based on the cultural knowledge graph and the stylistic features of the era, conflict identification is performed on the constituent elements of the era scene that fail the verification, and the era conflict features of the digital painting are obtained.

[0093] It should be noted that the stylistic features of the era described in this application refer to the characteristics of the visual norms and styles of the target historical scene; the conflict features of the era refer to the conflict features of digital painting that violate the norms of the era.

[0094] In specific implementation, firstly, when the consistency check fails, the consistency check value of the period scene components within the digital painting is obtained, and the period scene components with consistency check values ​​lower than the overall pass threshold are considered as the period scene components within the digital painting that have failed the check; secondly, a cultural knowledge graph is obtained; then, the constraints of the period characteristics in the historical scene visual restoration process are read, and the "period" and "location" fields (e.g., "Ming Dynasty" and "Nanjing") are extracted and combined (e.g., "Ming Dynasty"). Taking "Nanjing" as input, a structured summary generation service is invoked. Based on this keyword combination, the service automatically retrieves and generates a condensed text description of the visual characteristics of architecture, clothing, and artifacts in that era and region from pre-organized historical materials (e.g., "Ming Dynasty Nanjing official buildings mostly used blue glazed tiles, while residential buildings had white walls and black tiles; officials wore black gauze hats, and commoners mostly wore straight-cut clothing"). This text description is used as the stylistic feature of the target historical scene. Finally, for each element of the historical scene that fails validation, the following process is executed: based on its bounding box coordinates, a partial image of the element is cropped from the digital painting. This partial image, the category label of the element, and the aforementioned stylistic feature text are then input into a visual language model. (For example, the BLIP-2 model consists of a visual encoder (ViT-G / 14), a query transformer (Q-Former), and a large language model (FlanT5-XXL). It enables cross-modal alignment and reasoning between images and language. The input includes a partially cropped image of a digital painting, the category label of the region, and the text describing the style of the era. The model outputs the conflict analysis results of the natural language description through a preset prompt template (e.g., "What is the discrepancy between the [category] in this image and the given style description?"). During training, it is fine-tuned based on visual question answering and historical conflict annotation data, so that the model can accurately identify the specific contradictions between visual elements and the historical background, providing a clear basis for conflict correction.)Use visual language models to construct specific cue words (e.g., "Given the style description '[era style characteristics text]', what is wrong or anachronistic about this [category label]"). The model directly analyzes and generates text descriptions of the local image content, pointing out specific discrepancies between the image content and the given stylistic features of the era (e.g., "The hat of the figure in the picture has the Qing Dynasty's unique style of top hat and peacock feather, which does not match the black gauze hat that should be worn in the Ming Dynasty"). Simultaneously, using the category labels of the elements constituting the era (e.g., "hat") and the era keywords in the stylistic features (e.g., "Ming Dynasty") as joint query conditions, a query is performed in the cultural knowledge graph to find all "hat" entities that match the era background and obtain their standard names (e.g., "black gauze hat") and entity IDs. Based on this, a structured record is generated containing "boundary box coordinates of the elements constituting the era scene," "category labels of the elements constituting the era scene," "specific conflict descriptions identified by the visual language model," and "standard names and entity IDs provided by the cultural knowledge graph." After traversing all records not constituting the era scene, all structured records are compiled into a list, which is used as the era conflict feature of the digital painting.

[0095] In step S5, the digital painting is corrected for conflicts using the era conflict characteristics and generative intelligent model to obtain the corrected digital painting.

[0096] In this embodiment, the digital painting is corrected for conflicts using the aforementioned era conflict characteristics and generative intelligent model. The corrected digital painting can be obtained through the following steps:

[0097] Based on the aforementioned characteristics of historical conflict, a conflict redraw instruction is generated for the elements constituting the historical scene in the digital painting that fail verification.

[0098] The digital painting is modified by using a generative intelligent model and conflict redrawing instructions for elements of the era scene that fail verification, resulting in a corrected digital painting.

[0099] It should be noted that the conflict redraw instruction described in this application refers to an instruction to redraw the conflict information of the constituent elements of the era scene.

[0100] In practice, firstly, for each conflict record in the era's conflict characteristics, its "bounding box coordinates" are extracted, expanded by 5 pixels, and converted into a binary rectangular mask. Then, its "specific conflict description recognized by the visual language model" and "standard name provided by the cultural knowledge graph" are extracted. These are then combined and formatted into a clear repair prompt word using an instruction compilation script (for example, combining the conflict description "top hat and peacock feather style" with the standard name "black gauze hat" to generate "a Ming dynasty official wearing a black wushamao"). The system retrieves a standard reference image (e.g., a line drawing of a Ming Dynasty official hat) from a cultural knowledge graph using its "entity ID". It then packages the rectangular mask, repair prompts, and the standard reference image, using this as a conflict redrawing instruction for the elements of the era scene that failed verification within the corresponding digital painting. This process generates conflict redrawing instructions for the elements of the era scene that failed verification within the digital painting. Next, the system loads the generative intelligent model, using the original digital painting as the base image input, and processes each conflict redrawing instruction sequentially: inputting the base image, the matrix mask from the current conflict redrawing instruction, and the repair prompts into the generative intelligent model, while simultaneously using an image adapter. (For example, the IP-Adapter) takes the standard reference image in the current conflict repaint instruction as the style condition input to the generative intelligent model to guide the style consistency of local repainting, sets the repair parameters (e.g., only repaint the mask area, denoising intensity of 0.75), executes the model's inference, generates the repaired local image patch, and then uses the Poisson image fusion algorithm to seamlessly stitch the repaired image patch back into the corresponding mask area in the base image, ensuring natural edge transitions and consistent lighting. The result of this correction is used as the new base image to continue processing the next instruction. When all conflict repaint instructions have been processed, an image in which all identified conflicts have been corrected is obtained, and this image is used as the corrected digital painting.

[0101] Therefore, this application demonstrates that the digital painting can be corrected for conflicts using the aforementioned era conflict characteristics and generative intelligent model, resulting in a corrected digital painting. Firstly, by acquiring user reconstruction needs and historical document data from different eras, a clear target era and comprehensive knowledge base are established for the entire restoration process, defining the range of non-target era characteristics that need to be suppressed from the source. Secondly, feature analysis is performed on the reconstruction needs to extract the constraints of era characteristics, and a structured cultural knowledge graph is constructed using historical documents. This transforms the complex and unstructured multi-era knowledge into a computable and queryable feature rule base, providing precise guiding standards for the generation process and fundamentally solving the problems of mixed historical knowledge and its inability to be systematically utilized. Furthermore, during the generation stage, epochal constraints are directly applied to the generative intelligent model, guiding it to activate visual features of the target era. Simultaneously, a cultural knowledge graph is used to perform fine-grained epochal consistency verification on the output digital paintings. This achieves dual protection through proactive constraints before generation and precise detection after generation, effectively intercepting cross-epochal errors in elements such as architectural forms and clothing styles, and significantly reducing the risk of imperceptible epochal misalignment. Finally, when verification fails, the identified epochal conflict features are used to make targeted corrections using the generative model, enabling the system to adaptively reduce similar errors. This systematically improves the epochal accuracy of the generated model, suppresses cross-epochal feature confusion, and ensures the authenticity and reliability of the results of digital reconstruction of cultural and tourism resources.

[0102] In summary, the technical solution adopted in this application can suppress the erroneous generation of cross-era features in the cultural and tourism resource generation model, thereby reducing the probability of generation model misalignment.

[0103] Example 2: This application provides a digital reconstruction system for cultural and tourism resources based on a generative intelligent model, referencing... Figure 4 As shown in the figure, this is a module structure diagram of a digital reconstruction system for cultural and tourism resources based on a generative intelligent model according to this embodiment of the present application. The digital reconstruction system for cultural and tourism resources includes:

[0104] The data acquisition module 100 is used to acquire the user's reconstruction requirements for the target historical scene and historical document data from different eras;

[0105] The feature extraction module 200 is used to perform feature analysis on the reconstruction requirements, obtain the constraints of the era characteristics in the process of visual restoration of historical scenes, and construct a cultural knowledge graph through all historical document data;

[0106] The consistency verification module 300 is used to reconstruct the target historical scene according to the constraints of the generative intelligent model and the characteristics of the era, to obtain a digital painting of the target historical scene, and to perform an era consistency verification on the digital painting through the cultural knowledge graph to obtain the consistency verification result of the era characteristics within the digital painting.

[0107] The conflict identification module 400 is used to output a digital painting of the target historical scene when the consistency verification result passes, and to identify the era conflict characteristics of the digital painting through the era style characteristics of the target historical scene when the consistency verification result fails.

[0108] The conflict correction module 500 is used to correct the conflict in the digital painting by means of the conflict characteristics of the era and the generative intelligent model, so as to obtain the corrected digital painting.

[0109] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0110] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compactdisc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0111] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

Claims

1. A method for digital reconstruction of cultural and tourism resources based on a generative intelligent model, characterized in that, The method for digitally reconstructing cultural and tourism resources includes the following steps: Obtain users' needs for reconstructing target historical scenarios and historical document data from different eras; The reconstruction requirements are analyzed to obtain the constraints of the era characteristics in the process of visual restoration of historical scenes, and a cultural knowledge graph is constructed through all historical document data. The target historical scene is reconstructed based on the constraints of the generative intelligent model and the characteristics of the era, resulting in a digital painting of the target historical scene. The digital painting is then verified for consistency with the era through the cultural knowledge graph, and the consistency verification result of the characteristics of the era within the digital painting is obtained. Specifically, the process of reconstructing the target historical scene based on constraints imposed by a generative intelligent model and the characteristics of the era, resulting in a digital painting of the target historical scene, includes: Initialize the generative intelligent model; The constraints of the era's characteristics are transformed into positive and negative prompts that can be recognized by the generative intelligent model. The generative intelligent model refers to the artificial intelligence model that generates the image. The positive prompts are text strings used to describe and limit the specific elements, scenes, attributes, and styles that should appear in the generated image. The negative prompts are text strings used to prohibit certain specific elements or styles from appearing in the generated image. The digital painting displays images of the target historical scene. Based on a generative intelligent model, the target historical scene is reconstructed by combining the positive and negative prompt words to obtain a digital painting of the target historical scene; When the consistency check result passes, the digital painting of the target historical scene is output; when the consistency check result fails, the period conflict characteristics of the digital painting are identified by the period style characteristics of the target historical scene. The digital painting is corrected by applying the characteristics of the conflict of the era and the generative intelligent model to obtain the corrected digital painting.

2. The method for digital reconstruction of cultural and tourism resources based on a generative intelligent model as described in claim 1, characterized in that, The specific constraints on the era characteristics during the visual restoration of historical scenes, obtained by performing feature analysis on the reconstruction requirements, include: Keyword extraction of historical scene types is performed on the reconstruction requirements to obtain keywords of historical scene types in the process of visual restoration of historical scenes; Semantic analysis of the era information is performed on the reconstruction requirements to obtain the semantic features of the era information in the process of visual restoration of historical scenes; Based on the keywords of historical scene types and the semantic features of era information, the constraints of era characteristics in the process of visually restoring historical scenes are determined.

3. The method for digital reconstruction of cultural and tourism resources based on a generative intelligent model as described in claim 1, characterized in that, Constructing a cultural knowledge graph using all historical document data specifically includes: Named entity recognition is performed on text-based historical document data to obtain different cultural knowledge entities; Relationships are extracted from textual historical documents to obtain functional pairings and epochal affiliations among cultural knowledge entities. Visual element annotation is performed on image-based historical document data to obtain the visual attribute relationships between cultural knowledge entities; A cultural knowledge graph is constructed based on different cultural knowledge entities, the functional relationships and temporal affiliations between cultural knowledge entities, and the visual attribute relationships.

4. The method for digital reconstruction of cultural and tourism resources based on a generative intelligent model as described in claim 1, characterized in that, The consistency verification of the digital painting's epochal characteristics obtained by using the cultural knowledge graph specifically includes: The digital painting is deconstructed to obtain the constituent elements of different era scenes in the digital painting; The cultural knowledge graph is then compared with all the elements that make up the historical scenes to obtain the consistency verification values ​​of the elements that make up the historical scenes. The consistency check result of the epochal characteristics within the digital painting is determined based on all consistency check values.

5. The method for digital reconstruction of cultural and tourism resources based on a generative intelligent model as described in claim 1, characterized in that, The specific characteristics of temporal conflict in the digital painting, identified by the stylistic features of the target historical scene, include: Obtain the elements constituting the era scene that fail verification within the digital painting; Acquiring cultural knowledge graphs; Determine the stylistic characteristics of the target historical scene; Based on the cultural knowledge graph and the stylistic features of the era, conflict identification is performed on the constituent elements of the era scene that fail the verification, and the era conflict features of the digital painting are obtained.

6. The method for digital reconstruction of cultural and tourism resources based on a generative intelligent model as described in claim 1, characterized in that, By applying the aforementioned characteristics of historical conflict and a generative intelligent model to correct the conflict in the digital painting, the corrected digital painting is obtained, specifically including: Based on the aforementioned characteristics of historical conflict, a conflict redraw instruction is generated for the elements constituting the historical scene in the digital painting that fail verification. The digital painting is modified by using a generative intelligent model and conflict redrawing instructions for elements of the era scene that fail verification, resulting in a corrected digital painting.

7. The method for digital reconstruction of cultural and tourism resources based on a generative intelligent model as described in claim 1, characterized in that, The consistency verification result represents the verification result of how well the visual content of the digital painting matches the historical background in terms of the characteristics of the era.

8. The method for digital reconstruction of cultural and tourism resources based on a generative intelligent model as described in claim 1, characterized in that, The term "era conflict feature" refers to the conflicting characteristics of digital painting that violate era norms.

9. A system for digital reconstruction of cultural and tourism resources based on a generative intelligent model, used to execute a method for digital reconstruction of cultural and tourism resources based on a generative intelligent model as described in any one of claims 1 to 8, characterized in that, The digital reconstruction system for cultural and tourism resources includes: The data acquisition module is used to acquire users' reconstruction needs for target historical scenes and historical document data from different eras; The feature extraction module is used to perform feature analysis on the reconstruction requirements, obtain the constraints of the era characteristics in the process of visual restoration of historical scenes, and construct a cultural knowledge graph through all historical document data; The consistency verification module is used to reconstruct the target historical scene based on the constraints of the generative intelligent model and the characteristics of the era, to obtain a digital painting of the target historical scene. The digital painting is then verified for consistency with the era through the cultural knowledge graph to obtain the consistency verification result of the characteristics of the era within the digital painting. The conflict identification module is used to output a digital painting of the target historical scene when the consistency verification result passes, and to identify the period conflict characteristics of the digital painting by the period style characteristics of the target historical scene when the consistency verification result fails. The conflict correction module is used to correct the conflicts in the digital painting by using the conflict characteristics of the era and the generative intelligent model, so as to obtain the corrected digital painting.

Images