Creative design shared database construction method

Abstract

The invention discloses a creative design shared database construction method, which belongs to the technical field of database construction, and comprises the following steps: carrying out content feature analysis on a received creative design work file, and extracting to obtain an implicit visual feature vector; generating a dominant semantic feature vector based on the received corresponding initial description information and user community behavior data, fusing the dominant semantic feature vector with an implicit visual feature vector, constructing enhanced metadata corresponding to the creative design work file, calculating an association strength value between works, and constructing a graph database taking the works as nodes and the association strength value as a weight; and in response to a query request of a user, carrying out association path mining in the graph database, and carrying out collaborative filtering to generate a personalized work recommendation list. The implicit visual features and the dominant semantic features are fused, and the graph database based on the association strength is constructed, so that accurate representation, efficient management and personalized recommendation of the creative design works can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of database construction, in particular to a creative design sharing database construction method. BACKGROUND

[0002] In the digital age, creative design works such as images, videos, three-dimensional models, digital art, etc. are increasingly rich, and their number is growing explosively. These works not only carry the artistic ideas of the creators, but also contain rich visual and semantic information. Building an efficient database to store, manage and retrieve these works is of great significance to stimulating creativity, promoting communication and reuse. The current technical challenge is how to extract and represent the features of heterogeneous creative design works uniformly and deeply, so as to facilitate effective management, retrieval and personalized recommendation.

[0003] The prior art in handling creative design work sharing database usually relies on manually annotated labels, text descriptions or simple file metadata to classify and retrieve works. Some systems may use content-based image retrieval technology to analyze the visual features of works, or use natural language processing technology to analyze text descriptions. In terms of recommendation, it is often based on user's historical behavior records for collaborative filtering, or based on explicit labels of works for similarity matching. These methods improve the discoverability of works to some extent.

[0004] However, the existing technical solutions have some technical defects. First, excessive reliance on manual annotation can lead to incomplete, inaccurate or outdated labels, and it is difficult to capture the deep implicit features of works. Second, based on single modal features such as pure visual or pure text, it is difficult to fully depict the complexity and ambiguity of creative design works. For example, considering only visual features may ignore the deep semantics of works, and considering only text semantics may overlook the innovative performance in vision. Third, the existing work association and recommendation mechanism often lacks the ability to model deep and multi-dimensional associations between works, resulting in insufficient relevance and personalization of recommendation results, making it difficult for users to find works that match their potential interests, and limiting the depth of work value mining. SUMMARY

[0005] To solve the above problems, the present application provides a creative design sharing database construction method, which fuses implicit visual features and explicit semantic features, and constructs a graph database based on association strength, which can realize accurate representation, efficient management and personalized recommendation of creative design works.

[0006] The above-mentioned object can be achieved by the following scheme: The application discloses a creative design sharing database construction method, which comprises the following steps: receiving a creative design work file and corresponding initial description information; performing content feature analysis on the creative design work file to extract an implicit visual feature vector; generating and dynamically updating an explicit semantic feature vector based on the corresponding initial description information and user community behavior data; fusing the implicit visual feature vector and the explicit semantic feature vector to construct enhanced metadata corresponding to the creative design work file; calculating the correlation strength value between works based on the enhanced metadata of all creative design work files, and constructing a graph database with works as nodes and the correlation strength value as weight; and in response to a query request of a user, performing correlation path mining in the graph database and collaborative filtering to generate a personalized work recommendation list.

[0007] Optionally, the step of extracting the implicit visual feature vector comprises the following steps: acquiring a file type parameter of the creative design work file; matching and selecting a corresponding feature extraction model from a preset model library according to the file type parameter; and processing the creative design work file by using the feature extraction model to generate the implicit visual feature vector.

[0008] Optionally, the step of generating and dynamically updating the explicit semantic feature vector comprises the following steps: performing semantic analysis on the initial description information to extract an initial label set; obtaining a label heat parameter and a label co-occurrence frequency parameter by counting the user community behavior data; and dynamically calculating a label weight based on the initial label set, the label heat parameter and the label co-occurrence frequency parameter to generate the explicit semantic feature vector.

[0009] Optionally, the step of constructing the graph database with works as nodes and the correlation strength value as weight comprises the following steps: calculating a similarity parameter between the enhanced metadata of different creative design works; comparing the similarity parameter with a preset dynamically adjusted correlation threshold to generate a correlation relationship judgment result; generating the correlation strength value according to the correlation relationship judgment result, and establishing a weighted edge with the correlation strength value as weight between corresponding work nodes, so as to construct the graph database.

[0010] Optionally, the step of generating the personalized work recommendation list comprises the following steps: responding to a query request of a user and performing analysis to generate a query feature vector; performing graph atlas traversal in the graph database with the query feature vector as a starting point to obtain an associated work set; obtaining historical behavior preference data of a current user, and reordering the associated work set by using collaborative filtering to generate the personalized work recommendation list.

[0011] Optionally, the method further comprises: identifying different iterative versions of the same creative design work file, generating version association identifiers; recording modification logs and difference information of each version to form version change data; and constructing a version tree structure in the graph database based on the version association identifiers and the version change data.

[0012] Optionally, the method further comprises: embedding structured rights management information in the enhanced metadata; parsing authorization types and usage permission parameters from the rights management information; and adding traceable digital watermarks to the creative design work file based on the usage permission parameters.

[0013] Optionally, the method further comprises: monitoring the tag co-occurrence frequency parameters, identifying a strong association tag combination that continuously exceeds a co-occurrence threshold; automatically synthesizing a new tag based on the strong association tag combination; integrating the new tag and its initial weight into the explicit semantic feature vector and reversely labeling it to the relevant creative design work file.

[0014] Optionally, the method further comprises: obtaining interactive behavior data of a user when browsing associated works, generating an interactive heat factor; updating the association strength value by combining the similarity parameter and the interactive heat factor; and using the updated association weight value as the maximum weight of the weighted edge to update the association relationship in the graph database.

[0015] Based on the same inventive concept, the application also provides a creative design sharing database construction system, which comprises: a data receiving module for receiving creative design work files and corresponding initial description information; an implicit feature extraction module for performing content feature analysis on the creative design work files to extract implicit visual feature vectors; an explicit feature extraction module for generating and dynamically updating explicit semantic feature vectors based on the corresponding initial description information and user community behavior data; a feature enhancement module for fusing the implicit visual feature vectors and the explicit semantic feature vectors to construct enhanced metadata corresponding to the creative design work files; a database construction module for calculating the association strength values between works based on the enhanced metadata of all creative design work files, and constructing a graph database with works as nodes and the association strength values as weights; and a work recommendation module for responding to a user's query request, performing association path mining in the graph database, and performing collaborative filtering to generate a personalized work recommendation list.

[0016] Compared with the prior art, the application has the following advantages: 1. The present application fuses the implicit visual features and explicit semantic features of works, constructs enhanced metadata, and realizes multi-dimensional and high-precision integrated representation of creative design works, which overcomes the single and semantic missing problems that may exist in traditional work representation methods, enables the system to understand and capture the internal attributes and external semantics of works, and lays a solid foundation for intelligent management and deep analysis of works. 2. Based on the enhanced metadata between works, the present application dynamically calculates the correlation strength and constructs a graph database with the correlation strength as the weight, improves the discovery ability and expression efficiency of complex correlations between works, and through the graph structure, the system can clearly show the diversification of works, such as derivation, similarity and theme correlation, so that users can not only see single works, but also understand the ecological network in which the works are located, and enrich the context information and explorability of works. 3. When responding to the query request of the user, the present application mines the correlation path in the graph database and performs collaborative filtering combined with the historical preference data of the user to generate a personalized recommendation list, ensures the high relevance and personalization of the recommendation result, solves the information cocooning and cold start problems that may exist in traditional recommendation systems, improves the efficiency and accuracy of users discovering potential interested works, and optimizes the user experience.

[0017] 4. The present application also has version management and digital copyright protection capabilities, constructs a version tree structure to clearly record the evolution process and change details of works, facilitates traceability and collaboration, and at the same time, embeds structured right management information in the enhanced metadata and adds traceable digital watermark to protect the copyright of creative design works, reduce the risk of infringement, and promote a healthy and orderly creative content sharing ecosystem.

[0018] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the present application. The objects and other advantages of the present application can be realized and obtained by the structure indicated in the specification, claims and drawings. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0020] Figure 1 is a flowchart of a creative design sharing database construction method according to an embodiment of the present application.

[0021] Figure 2is a version tree structure schematic diagram of an embodiment of the present application.

[0022] Figure 3 is a structure schematic diagram of an innovative design sharing database construction system of an embodiment of the present application. DETAILED DESCRIPTION

[0023] To make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0024] With reference to Figure 1 An embodiment of the present application proposes an innovative design sharing database construction method, which fuses implicit visual features and explicit semantic features, and constructs a graph database based on correlation strength, so as to realize accurate representation, efficient management and personalized recommendation of innovative design works.

[0025] The method of the embodiment specifically includes: receiving an innovative design work file and corresponding initial description information; performing content feature analysis on the innovative design work file to extract an implicit visual feature vector; generating and dynamically updating an explicit semantic feature vector based on the corresponding initial description information and user community behavior data; fusing the implicit visual feature vector and the explicit semantic feature vector to construct enhanced metadata corresponding to the innovative design work file; calculating correlation strength values between works based on the enhanced metadata of all innovative design work files, and constructing a graph database with works as nodes and the correlation strength values as weights; in response to a query request of a user, generating a personalized work recommendation list by performing correlation path mining in the graph database and performing collaborative filtering.

[0026] The present application extracts and integrates implicit visual features and explicit semantic features from creative design works in a multi-modal fusion manner. The implicit visual features are obtained by in-depth analysis of the content of the works, reflecting the visual style, structure and aesthetic properties of the works. The explicit semantic features combine the initial description information of the works and the interaction behavior data of users in the community, dynamically capturing the semantic theme of the works and the collective cognition of users. The fusion of these two features builds enhanced metadata, realizing the comprehensive representation of multi-dimensional information of the works. Based on these enhanced metadata, the system quantifies the correlation strength between works and constructs a graph database with this weight, so that the complex relationships between works can be structured and represented. When users query, the system uses the correlation structure of the graph database for deep mining, and combines the personalized preferences of users for collaborative filtering, thereby generating highly relevant recommendation results.

[0027] By deeply integrating the visual content of works and community semantics, the system can comprehensively and accurately understand and represent creative works. The construction of the graph database effectively solves the problem of discovering and expressing complex correlations between massive works, making the potential connections between works clear and visible. When users issue query requests, the system can not only rely on query content, but also fully consider the internal correlations between works and individual preferences of users to dynamically generate highly customized recommendation lists. This mechanism significantly improves the efficiency and experience of users in discovering, exploring and utilizing creative design works, promotes the dissemination and realization of the value of works, and can flexibly adapt to the characteristics of diverse types and variable semantic expressions of creative design works.

[0028] Optionally, the extracted implicit visual feature vector comprises: receiving creative design work files and corresponding initial description information; obtaining the file type parameter of the creative design work file; According to the file type parameter, matching and selecting the corresponding feature extraction model from the preset model library; processing the creative design work file using the feature extraction model to generate an implicit visual feature vector.

[0029] Specifically, creative design work files and their corresponding initial description information are received. These files may include various types, such as image files JPEG, PNG, PSD, vector graphics files SVG, AI, engineering files CAD, Blender, and video files MP4, MOV, GIF, etc. Then, the system needs to obtain the file type parameter of the received creative design work file. This parameter can be obtained by checking the file extension, reading the magic bytes of the file header, or using file recognition tool libraries such as the "file" command or Apache Tika, etc.

[0030] After obtaining the file type parameter, the system will match and select the corresponding feature extraction model from the preset model library according to the parameter. The model library pre-stores deep learning models or traditional machine learning models for different file types. For example, if the file type parameter indicates an image file JPEG or PNG, the system may select an image feature extraction model based on a convolutional neural network, such as ResNet, VGG, or Vision Transformer ViT, etc. If the file type parameter indicates a video file MP4, a video feature extraction model combining a 3D convolutional neural network or a spatiotemporal attention mechanism may be selected. If the file type parameter indicates a vector graph file SVG or a CAD engineering file, a graph neural network GNN model capable of processing structured graph data or geometric information may need to be selected, or a model based on traditional computer graphics algorithms to extract geometric features. For text-based creative design work files, such as font design or concept description documents, models in the natural language processing NLP field may be used, such as BERT, GPT, etc.

[0031] Once the appropriate feature extraction model is selected, the system will process the creative design work file using the model. This process involves inputting the original file data into the model, performing layer-by-layer operations and abstraction on the model, and finally outputting a fixed-length numerical vector, which is the implicit visual feature vector. For example, for an image file, the model will convert the pixel data into a high-dimensional vector, capturing the texture, color, shape, composition, and other visual elements of the image. For a video file, the model will extract the temporal relationship and motion information between frames. For a vector graph file, the model will encode line, shape, color, layout, and other graphical properties. The internal operation mechanism of these models is usually based on the forward propagation algorithm of neural networks, which gradually maps the original data to the semantic space through weighted summation, nonlinear activation, pooling, and other operations, forming an abstract representation of the visual content of the work. This implicit visual feature vector is a fixed-dimension, numerical vector that can reflect the deep visual characteristics of the creative design work, without relying on human-understandable labels or descriptions.

[0032] Optionally, the generating and dynamically updating the explicit semantic feature vector comprises: performing semantic analysis on the initial description information to extract an initial label set; statistically analyzing the user community behavior data to obtain a label popularity parameter and a label co-occurrence frequency parameter; based on the initial label set, the label popularity parameter, and the label co-occurrence frequency parameter, dynamically calculating a label weight to generate an explicit semantic feature vector.

[0033] Specifically, the process begins with semantic parsing of the initial descriptive information to extract an initial tag set. This initial descriptive information is typically the text accompanying a user's uploaded creative design work, such as the title, description, keywords, and category. Semantic parsing can be implemented using Natural Language Processing (NLP) techniques, including but not limited to word segmentation, part-of-speech tagging, named entity recognition, keyword extraction, and text classification. Word segmentation divides the text into individual words or phrases; part-of-speech tagging identifies the grammatical function of each word; named entity recognition identifies specific entities such as names of people, places, and organizations; keyword extraction uses methods such as TF-IDF, TextRank, and BERT-based keyword extraction models to automatically identify keywords from the descriptive text that accurately summarize the content of the work. Text classification categorizes the descriptive information into a predefined classification system, and its classification labels can also serve as initial tags. These extracted keywords, entities, and classification labels collectively constitute the initial tag set. For example, if the initial descriptive information is "a 3D rendering of a sci-fi cityscape for architectural visualization," semantic parsing might extract "architectural visualization," "sci-fi," "cityscape," and "3D rendering" as initial tags.

[0034] Next, the system will collect user community behavior data to obtain tag popularity and tag co-occurrence frequency parameters. User community behavior data includes user interactions such as browsing, liking, commenting, sharing, saving, downloading, and searching using tags related to the work. Tag popularity reflects the popularity or frequency of use of a single tag within the user group, and can be obtained by counting the number of times a specific tag is viewed, clicked, or searched, or the number of times it is mentioned by other users in comments or reposts. For example, the higher the frequency of a tag being searched by users, the higher its popularity parameter value. Tag co-occurrence frequency measures the frequency with which two or more tags appear simultaneously in user behavior. This can be calculated by analyzing the combination of tags in works viewed, liked, or saved by users, or by analyzing combinations of user search keywords. For example, if the tags "architectural visualization" and "3D rendering" frequently appear together in user interactions with the same work, or are frequently used as combined search keywords, then the co-occurrence frequency parameter of these two tags will be high.

[0035] Finally, based on the initial tag set, the tag popularity parameter, and the tag co-occurrence frequency parameter, tag weights are dynamically calculated to generate an explicit semantic feature vector. The explicit semantic feature vector is a high-dimensional vector whose dimensions correspond to all known tags, and each component in the vector represents the weight of a tag. The formula for calculating tag weights is as follows: , in, Indicates label The weights; Indicates label The popularity parameter is obtained by counting the number of times the tag appears in all works or user behaviors and then normalizing it to reflect its relative popularity. Indicates label and tags The co-occurrence frequency parameter, i.e., the tag and tags The number of times it appears in both the work description and user behavior is normalized. Indicates label The weight can be the weight from the previous round or the initial weight during the iterative update process. and It is a hyperparameter that adjusts the influence of tag popularity and tag co-occurrence frequency; its value typically ranges from 0 to 1. This formula determines the importance of each tag by combining its own popularity with its correlation with other tags. The explicit semantic feature vector is a vector composed of all tags and their calculated weights, for example... ,in This represents the total number of tags. This vector is dynamically updated. As new user community behavior data flows in, the tag popularity and co-occurrence frequency parameters are updated in real time or periodically, leading to a recalculation of tag weights. This ensures that the explicit semantic feature vector can promptly reflect the current community's understanding and preference for the semantics of the work.

[0036] Optionally, constructing a graph database with works as nodes and the association strength value as weights includes: Calculate the similarity parameter between the enhanced metadata of different creative design works; The similarity parameter is compared with a preset dynamically adjusted association threshold to generate an association determination result; Based on the association determination results, an association strength value is generated, and weighted edges with the association strength value are established between the corresponding work nodes to construct a graph database.

[0037] Specifically, this first involves fusing implicit visual feature vectors and explicit semantic feature vectors to construct enhanced metadata corresponding to creative design work files. Implicit visual feature vectors are numerical representations extracted from the content of the work, capturing its visual characteristics. Explicit semantic feature vectors reflect the semantic information and community perception of the work. The fusion process typically involves concatenating or weighting these two vectors to form a more comprehensive and richer composite vector. For example, if the implicit visual feature vector is... The explicit semantic feature vector is Enhanced metadata It can be represented as: , The square brackets here indicate a vector concatenation operation, which joins two vectors end-to-end to form a longer vector. Alternatively, a weighted summation method can be used, i.e.: , in and These are preset weighting coefficients used to adjust the relative importance of visual and semantic features in enhanced metadata. The sum of these two coefficients is typically 1, and their values ​​can be adjusted according to the specific application scenario; for example, when visual features are more important... The value will be larger. Enhanced metadata contains both visual and semantic information about the work, providing a foundation for subsequent work association analysis.

[0038] Next, the system will calculate the similarity parameter between the enhanced metadata of different creative design works. For any two works i and j, their enhanced metadata are respectively and Similarity parameters can be calculated using various distance or similarity metrics, such as cosine similarity, the reciprocal of Euclidean distance, or the Pearson correlation coefficient. Taking cosine similarity as an example, the calculation formula is as follows: , in, Represents the similarity parameter. Represents the dot product of two vectors. This represents the L2 norm of a vector, i.e., the magnitude of the vector. Cosine similarity measures the degree of proximity between two vectors in a direction, with values ​​ranging from -1 to 1; a higher value indicates a higher similarity. In this way, the degree of correlation between any two works in terms of content can be quantified.

[0039] Then, the similarity parameter is compared with a preset, dynamically adjusted association threshold to generate an association determination result. The association threshold is a critical value that determines whether an association is established between works. This threshold can be dynamically adjusted based on the similarity distribution of the overall dataset, user feedback, or empirical rules. For example, if the similarity parameter is higher than the current association threshold, an association is determined between the two works; otherwise, it is not. Dynamic adjustment means that this threshold is not fixed; it can be optimized in real-time or periodically based on the growth of the number of works in the database, changes in user behavior, or the evaluation of the system's recommendation effect.

[0040] Finally, based on the association determination results, an association strength value is generated, and weighted edges with the association strength value are established between the corresponding work nodes, thereby constructing a graph database. For work pairs determined to have an association relationship, their similarity parameters can be directly used as the association strength value, or further transformed and weighted, for example, by combining other attribute information of the work, such as publication time and author, to generate the final association strength value. Then, in the graph database, each creative design work is represented as a node. If there is an association relationship between two works, an edge is established between their corresponding nodes. The weight of this edge is the calculated association strength value. In this way, all works and their association strengths together constitute a complex weighted graph database.

[0041] Optionally, the generation of the personalized work recommendation list includes: Respond to user query requests and parse them to generate query feature vectors; The graph database is traversed starting from the query feature vector to obtain a set of related works. Obtain the current user's historical behavioral preference data, and use collaborative filtering to reorder the associated works set to generate a personalized works recommendation list.

[0042] Specifically, the system first responds to and parses user queries to generate a query feature vector. User queries can be text input, such as keywords, phrases, or natural language descriptions, or they can be uploaded image files as query examples, or selected work IDs. If the query is text, the system uses the same or similar Natural Language Processing (NLP) techniques as before for semantic feature extraction to perform semantic parsing, extracting keywords, entities, and calculating semantic weights to form a query semantic vector. If the query is an image file or work ID, the system extracts its implicit visual feature vector and combines it with its existing explicit semantic feature vector to construct a comprehensive query feature vector, similar to the method of constructing enhanced metadata for works. This query feature vector is a quantitative representation of the user's interests or needs in the vector space.

[0043] Next, a graph traversal is performed in the graph database starting with the query feature vector to obtain a set of related works. Graph traversal is a method for finding relevant nodes in graph-structured data. The traversal algorithm varies depending on the type of query feature vector and the specific recommendation strategy. If the query is a specific work ID, then that work node is used directly as the starting point for traversal. If the query is a keyword or image feature, the system first needs to find one or more work nodes most similar to the query feature vector as starting points. This can be done by calculating the similarity between the query feature vector and the enhanced metadata of all works in the graph database. The traversal method can be breadth-first search (BFS) or depth-first search (DFS), or more complex algorithms based on random walks or graph neural networks (GNNs). During the traversal, the system prioritizes exploring work nodes with stronger association with the query starting point based on the edge weights, i.e., the association strength values, thereby collecting works related to the user's query topic. For example, a traversal depth or an upper limit on the number of works collected can be set to ensure the efficiency of the traversal process and the relevance of the results. All work nodes obtained through the traversal constitute the initial set of related works.

[0044] Then, the user's historical behavioral preference data is obtained, and collaborative filtering is used to re-rank the set of related works to generate a personalized list of recommended works. The user's historical behavioral preference data includes records of works viewed, liked, favorited, commented on, downloaded, or purchased in the past, as well as the user's interest tags and demographic information. Collaborative filtering is a recommendation method that utilizes collective intelligence; it is divided into user-based and item-based methods. In this scenario, a hybrid or improved collaborative filtering strategy can be used. For example, item-based collaborative filtering can analyze the similarity between the user's historically preferred works and works in the set of related works, giving higher recommendation priority to related works that are highly similar to the user's already liked works. User-based collaborative filtering finds other users with similar interests to the current user and then recommends works liked by these similar users but not yet accessed by the current user. When re-ranking the set of related works, a ranking function can be designed to comprehensively consider the original association strength of the related works, the user's preference matching degree, and other factors such as the work's freshness and popularity. For example, the final recommendation score of a work can be calculated using the following formula: , in, It is a work Recommended score; It is a query feature vector With the work The similarity of enhanced metadata, or the strength of associations obtained through graph traversal; User On the work The degree of preference, which can be calculated using collaborative filtering algorithms, for example, could be the user's... Works that have been historically performed The average similarity; It is a work Popularity metrics, such as the number of likes and downloads; , , These are parameters that adjust the weights of various influencing factors. Ultimately, the system will sort the related works in descending order based on the calculated scores and select the top N works as a personalized recommendation list.

[0045] Optionally, the method further includes: Identify different iterations of the same creative design file and generate version association identifiers; Record the change logs and difference information for each version to form version change data; Based on the version association identifier and the version change data, a version tree structure is constructed in the graph database.

[0046] Specifically, the process begins by identifying different iterative versions of the same creative design file and generating version association identifiers. When a user uploads or modifies a creative design file, the system treats it as a new version. To identify the relationship between different versions, the following mechanism can be used: For the first uploaded work, the system assigns it a unique work ID. For subsequent modifications, the system treats the modified file as a new version of the work. When processing a new version, the system compares the newly uploaded file with previous versions of the work. Comparison methods can include file hash value verification, such as MD5 or SHA256 hash values. If the file hash value changes, it is confirmed as a new version. A more precise approach is to use specific file formats, such as CAD files or design drafts, such as PSD or AI files, to determine whether the modification is based on the original file by parsing the file's internal structure or metadata. If a new version of the same work is identified, the system assigns a unique version ID to the new version and associates it with the original work ID, generating a version association identifier, such as binding the original work ID with the new version ID.

[0047] Next, the system records the change logs and difference information for each version, forming version change data. The change logs include timestamps, modifier information, change type (e.g., text modification, element addition, color adjustment), and user-submitted comments. The difference information records the specific content changes between the current version and the previous version. For image files, differences can be quantified through pixel-level comparison or feature-based comparison; for vector graphics or text files, specialized file comparison tools or algorithms can be used to identify and record added, deleted, or modified elements. For example, for code files or text descriptions, the Diff tool can be used to generate difference patch files. For image files, image comparison algorithms can be used to calculate the mean squared error or structural similarity index between two versions and mark the key areas or elements that have changed. These change logs and difference information together constitute the version change data, which details the evolution of the work from one version to the next.

[0048] Finally, based on the version association identifier and the version change data, a version tree structure is constructed in the graph database. In the graph database, each version of a creative design work can be modeled as an independent node. The version association identifier is used to identify all different versions belonging to the same work and link them together. For example... Figure 2 As shown, for each work, all its version nodes are connected by directed edges, forming a version tree or version chain. Typically, each directed edge points from an older version node to a newer version node, and version change data can be attached to the edge as an attribute, for example: , in, The first version of the representative work, The second version of the representative work, This represents the third version of the work. These edges can have weights, which reflect the complexity or importance of the modifications between versions. For example, a version that modifies core design elements can be given a higher weight. If a work has branching versions—that is, two or more different subsequent versions derive from a certain version—multiple directed edges will be drawn from that version node, forming a tree structure. In this way, users can clearly trace the evolution history of the work and understand the inheritance relationship and specific modifications between different versions.

[0049] This method provides a version management mechanism for creative designs, enabling the systematic recording and organization of the iterative history of these works. By integrating version information into a graph database, it not only clearly displays the evolutionary path of a work but also supports source tracing, team collaboration, copyright management, and knowledge reuse. Users can easily view the differences between different versions and rewind to any historical version, enhancing the granularity and controllability of work management, thereby improving the efficiency of the design process and the value of the work.

[0050] Optionally, the method further includes: Structured rights management information is embedded in the enhanced metadata; The authorization type and usage permission parameters are parsed from the rights management information; Based on the usage permission parameters, a traceable digital watermark is added to the creative design work file.

[0051] Specifically, structured rights management information is first embedded in the enhanced metadata. Rights management information refers to data describing the copyright, licensing status, and usage restrictions of a work. This information should be stored in a structured format, such as JSON, XML, or RDF, to facilitate machine parsing and processing. The structured information should at least include the work's author information, copyright owner, creation date, license type, scope of use, geographical restrictions, validity period, and source tracking identifier. The license type can be CC BY-SA, commercial license, etc., and the scope of use can be personal use, commercial use, derivative works, etc. This information is provided by the author or copyright manager when the work is uploaded, and the system parses it and embeds it as new fields or attributes into the work's enhanced metadata.

[0052] Next, the system will parse the authorization type and usage permission parameters from the rights management information. The authorization type refers to the license model adopted by the work, such as open source, commercial, or restricted license. The usage permission parameters specify the specific operations the licensee can perform on the work, such as whether viewing, downloading, editing, modifying, distributing, and selling are allowed, whether attribution is required, and whether it can be used for commercial purposes. These parameters typically exist in the structured rights management information as Boolean values ​​(yes / no), enumerated values, or numerical ranges. For example, if the rights management information contains the fields "allow_commercial_use: true" and "allow_modification: false", the system will parse that commercial use is allowed and modification is not allowed.

[0053] Finally, based on the aforementioned usage permission parameters, a traceable digital watermark is added to the creative design work file. Digital watermarking is a technology that embeds hidden information into digital media files without affecting their perceptible quality, primarily used for copyright protection and source tracing. Specific operations include watermark information encoding, watermark embedding algorithm selection, watermark strength adjustment, and watermark file generation.

[0054] The watermark information encoding process involves encoding key data such as usage permission parameters, the work's unique identifier, owner information, and authorized user ID to generate a binary sequence as the watermark information. The watermark embedding algorithm is selected based on the file type of the work. For image files (JPEG and PNG), frequency domain watermarking algorithms, such as Discrete Cosine Transform (DCT), Discrete Wavelet Transform (DWT), or Fourier Transform (FFT), can be used, embedding the watermark information into the image's frequency coefficients. Spatial domain watermarking algorithms, such as Least Significant Bit (LSB) replacement, can also be used. For video files (MP4 and MOV), the watermark can be embedded in each frame or in the inter-frame motion vectors and encoding parameters, such as watermark embedding based on DCT, DWT, or H.264 / H.265 bitstreams. For audio files (MP3 and WAV), spread spectrum, phase modulation, or echo hiding watermarking techniques can be used. For vector graphics files (SVG, PDF), or engineering files (CAD), the watermark can be embedded in the file's metadata, layer names, node attributes, or encoded through minor geometric deformations. The watermark strength adjustment allows you to adjust the watermark strength according to actual needs. Excessive strength may affect the quality of the work, while insufficient strength may make it easily removable. Watermark file generation involves embedding the watermark information into the creative design file, creating a new file with an invisible watermark. This new file is visually or aurally almost indistinguishable from the original file, but it contains authorization and provenance information that can be extracted by specific detectors.

[0055] When a work is used or disseminated illegally, watermark information can be extracted from suspicious files using a watermark extractor and compared with the authorization type and usage permission parameters recorded in the database, thereby tracing the source of the work, the authorization status, and possible infringing users or channels.

[0056] This method provides a digital rights management and protection mechanism for creative design works. By tightly integrating structured rights management information with enhanced metadata of the work and embedding it using digital watermarking technology, the authorization information of the work can be disseminated along with the work file itself, improving the attachment and robustness of copyright information. This method not only helps to clarify the boundaries of authorization and prevent unauthorized use, but also provides strong evidence for tracing infringement, thereby effectively protecting creators' works and promoting the healthy development of the creative design ecosystem.

[0057] Optionally, the method further includes: Monitor the tag co-occurrence frequency parameter and identify strongly associated tag combinations that consistently exceed the co-occurrence threshold; Based on the strongly associated tag combinations, new tags are automatically synthesized; The new labels and their initial weights are integrated into the explicit semantic feature vector and then back-annotated to the relevant creative design work files.

[0058] Specifically, the process begins by monitoring the tag co-occurrence frequency parameter to identify strongly correlated tag combinations that consistently exceed a co-occurrence threshold. The tag co-occurrence frequency parameter represents the frequency at which two or more tags appear simultaneously in user behavior data. The system continuously statistically analyzes these co-occurrence frequencies. A co-occurrence threshold is set; this threshold is a preset value representing the minimum co-occurrence frequency a tag combination must reach to be considered "strongly correlated." For example, if the tags "architectural visualization" and "3D rendering" consistently appear together in works viewed, saved, or searched by users within a recent period, exceeding the set co-occurrence threshold, the system will identify them as a strongly correlated tag combination. This monitoring can be performed in real-time or periodically to ensure the latest user behavior trends are captured. To filter out sporadic or unstable co-occurrences, the "consistency" condition is usually considered; for example, the co-occurrence frequency of a tag combination exceeding the threshold for the most recent N time windows.

[0059] Next, based on the strongly related tag combinations, the system automatically synthesizes new tags. Once a persistently strongly related tag combination is identified, the system treats it as a new concept with specific compound semantics. For example, the tags "architectural visualization" and "3D rendering" can be combined into a new, more general and specific tag, such as "architectural rendering".

[0060] There are several ways to synthesize new tags. One approach is direct concatenation, where multiple strongly related tags are combined into a new phrase, such as "architectural visualization_3D rendering". Another approach is based on dictionaries or ontology, where a pre-built domain dictionary or knowledge ontology is consulted to find superordinate terms or technical terms that summarize the tag combination. A third approach is based on machine learning models, where a text generation model capable of recognizing compound concepts, such as a Seq2Seq model, is trained to take a strongly related tag combination as input and output the corresponding natural language description as the new tag. Automatically synthesized new tags aim to more accurately describe the content of the work, reduce redundancy in user input, and better capture implicit semantic concepts.

[0061] Then, the new tag and its initial weight are integrated into the explicit semantic feature vector and back-labeled to the relevant creative design work files. For each automatically synthesized new tag, the system needs to assign it an initial weight. This initial weight can be calculated based on the popularity and co-occurrence frequency parameters of the original tags that constitute the new tag. For example, the initial weight of the new tag can be the average weight or weighted sum of the original tags that constitute it. Subsequently, the new tag will be added to the global tag set as a new dimension. Correspondingly, the explicit semantic feature vectors of all creative design works will be updated, and the dimension corresponding to this new tag will be filled with its initial weight value (if the work content is related to the new tag) or 0 (if not). At the same time, the system will back-label this newly synthesized tag to all creative design work files that prompted the generation of the new tag and are highly related to its content. This means that the initial tag set of these works will be expanded to include this new tag. For example, works that were previously only labeled as "architectural visualization" and "3D rendering" will now also be additionally labeled as "architectural rendering".

[0062] This method continuously optimizes and enriches the semantic representation of creative design works, making it more accurate and dynamic. By automatically discovering and synthesizing new tags, the system can capture emerging and complex semantic concepts in the user community, thereby continuously improving the depth and breadth of understanding the content of the works. This dynamic tagging system not only reduces the burden of manual annotation but also enables semantic feature vectors to more accurately reflect the essential characteristics of the works and the user's focus. By integrating new tags into explicit semantic feature vectors and performing reverse annotation, the system can improve the retrieval accuracy and recommendation quality of works, helping users find works that meet their potential needs faster and more accurately, thereby optimizing the user experience and promoting the accumulation and evolution of community knowledge.

[0063] Optionally, the method further includes: Collect user interaction data when browsing related works and generate interaction popularity factors; The association strength value is updated by combining the similarity parameter and the interaction heat factor; The updated association weight values ​​are used as the final weights of the weighted edges to update the associations in the graph database.

[0064] Specifically, the system first acquires user interaction data while browsing related works and generates an interaction popularity factor. When a user browses a creative design work, the system continuously monitors and records various interactions related to that work. These interactions include, but are not limited to: browsing time (the time a user spends on the work's page), clicks (such as clicking to zoom in and view details), favorites, likes, comments, shares, downloads, or clicks on related recommended works. This data reflects the user's level of interest and appreciation for the currently viewed work and its related works. The interaction popularity factor is an indicator that quantifies user interaction activity. For example, different weights can be assigned to different interaction behaviors; for instance, favorites and likes may have a higher weight than simply browsing. The interaction popularity factor $IF_{ij}$ can be calculated using the following formula: , in, The interaction popularity factor between work i and work j; It is the weight of the k-th interaction behavior. For example, the weight of the like behavior may be higher than the browsing time. This refers to the number or intensity of interactions a user performs on work j after viewing work i. By accumulating all user interactions between work i and work j and weighting the different interactions by importance, we can obtain the interaction heat factor, which reflects the strength of the user interest association between the two works.

[0065] Next, the system will update the association strength value by combining the similarity parameter and the interaction popularity factor. The similarity parameter is calculated based on the content features of the work (enhanced metadata resulting from the fusion of implicit visual feature vectors and explicit semantic feature vectors). Association strength value The initial value could be a simple similarity parameter, but now it needs to incorporate the popularity generated by actual user interactions. The formula for updating the association strength value can be designed as follows: , in, This is the updated association strength value; It is the similarity parameter between work i and work j calculated based on enhanced metadata; It is a normalized interaction heat factor, which is usually normalized to between 0 and 1 so that it is in the same dimension as the similarity parameter; This is a weighting parameter, ranging from 0 to 1, used to adjust the relative importance of content similarity and user interaction popularity in the final association strength value. When When the similarity is close to 1, content similarity becomes dominant; when... When the frequency approaches zero, user interaction activity becomes dominant. The model can be adjusted based on the actual results to ensure that it considers both the relevance of the content itself and the collective wisdom of the users.

[0066] Finally, the system uses the updated association weight values ​​as the final weights of the weighted edges to update the associations in the graph database. In the graph database, the weights of the weighted edges established between works will be updated to the weights calculated above. This means that even if two works have a high degree of similarity in content, if users rarely observe switching or interacting between them, their association strength may be moderately reduced due to a lower interaction popularity factor. Conversely, if two works have only moderate content similarity but users frequently interact between them, their association strength will increase accordingly. This dynamic update process can be performed periodically to ensure that the associations in the graph database can reflect changes in user behavior and new popular trends in a timely manner.

[0067] This method improves the accuracy and dynamism of the relationships between creative design works. It relies not only on the objective similarity of the content but also incorporates social feedback generated by actual user interactions. This allows the association weights between works to more accurately reflect the user group's recognition of the potential connections and their usability. In this way, the work association structure in the graph database can more sensitively capture changes in user interests and trends, providing more personalized, real-time, and accurate association information for subsequent recommendations and searches, thus improving user discovery efficiency and experience satisfaction when exploring creative design works.

[0068] Based on the same inventive concept, such as Figure 3 The present invention also provides a creative design sharing database construction system, the system comprising: The data receiving module is used to receive creative design work files and corresponding initial description information; The latent feature extraction module is used to perform content feature analysis on the creative design work file and extract latent visual feature vectors. The explicit feature extraction module is used to generate and dynamically update explicit semantic feature vectors based on the corresponding initial description information and user community behavior data. The feature enhancement module is used to fuse the implicit visual feature vectors and the explicit semantic feature vectors to construct enhanced metadata corresponding to the creative design work file; The database construction module is used to calculate the association strength value between works based on the enhanced metadata of all creative design work files, and to construct a graph database with works as nodes and the association strength value as weights. The works recommendation module is used to respond to user query requests by mining related paths in the graph database and performing collaborative filtering to generate a personalized works recommendation list.

[0069] To verify the feasibility of this invention in practice, it was applied to an online creative design sharing platform. This platform aggregates a large number of creative design works of various types, including images, videos, 3D models, vector graphics, and conceptual text. Previously, the platform faced challenges in work retrieval, related recommendations, and copyright management. Works of different formats were difficult to manage uniformly, keyword-based recommendations were often inaccurate and homogenized, and work theft and copyright tracing were difficult.

[0070] In this embodiment, the creative design sharing platform utilizes the method of this invention to construct its core database. First, when a designer uploads their work, the data receiving module receives creative design work files, including JPEG and PNG image files, MP4 video files, and PSD and AI project files, along with corresponding initial description information. The latent feature extraction module selects a ResNet image model or a 3D CNN video model based on the file type to perform content analysis on the work and extract latent visual feature vectors. For example, for an uploaded "cyberpunk cityscape" design, the system will extract its visual features such as color, texture, and composition. The explicit feature extraction module performs semantic parsing on the initial description information, for example, extracting initial tags such as "cyberpunk," "cityscape," and "futurism." Simultaneously, it collects user community behavior data, such as the number of times the "cyberpunk" tag is viewed and the frequency of co-occurrence of the "cityscape" and "futurism" tags, and dynamically calculates tag weights accordingly to generate explicit semantic feature vectors.

[0071] The feature enhancement module concatenates and fuses implicit visual feature vectors and explicit semantic feature vectors to generate enhanced metadata. For example, it fuses the visual style feature vector of "Cyberpunk City Landscape" with its semantic tag vector to form its comprehensive representation. Subsequently, the database construction module calculates the cosine similarity between works based on their enhanced metadata and compares it with a dynamically adjusted association threshold. For instance, if the enhanced metadata similarity between "Cyberpunk City Landscape" and another work, "Science Fiction Architectural Design," is higher than the threshold of 0.8, the system establishes a weighted edge between them, with the weight representing the association strength value. In this way, a graph database is constructed, with all works as nodes and the association strength value as the weight.

[0072] Furthermore, the platform also implements version management and copyright protection functions. For example, when a designer modifies and uploads a new version of the "Cyberpunk City Landscape" design, the system recognizes that it is a different iteration of the same work, generates a version association identifier, and records the pixel differences and modification logs between the old and new versions. This version information is stored in the graph database in a version tree structure for easy historical tracing. Simultaneously, structured rights management information, such as author and license type "personal non-commercial use only," is embedded in the enhanced metadata. When a user downloads the work, the system adds a traceable digital watermark to the downloaded file based on the parsed usage permission parameters, such as "no commercial use" and "no modification," effectively preventing illegal dissemination.

[0073] The work recommendation module responds to user queries, such as searching for "futuristic architectural design," and parses the query into a feature vector. Starting with this vector, the system traverses the graph database to obtain an initial set of works associated with "futuristic architectural design." Then, considering the user's historical behavioral preferences (e.g., the user has previously liked numerous works tagged with "cyberpunk" or "3D rendering"), a collaborative filtering algorithm is used to reorder the associated work set. Finally, a highly personalized work recommendation list is generated, which not only includes visually similar works but also considers semantic relevance and the user's personalized interests.

[0074] Regarding the discovery of work-related associations, the graph database constructed by this invention enables the platform to identify cross-modal associations that are difficult to detect using traditional methods. For example, when a user searches for an image of "Chinese-style garden design," the platform can recommend not only other images with similar styles but also videos or 3D models that share the semantic tags "Chinese style" and "garden." Compared to traditional recommendation systems based on keyword matching or single feature similarity, the average recommendation relevance of the method in this invention is improved.

[0075] In terms of personalized recommendations, the system can dynamically adjust recommendation weights based on user behavior data. For example, after a user likes and saves a series of animated short videos tagged with "abstract geometry" and "motion graphics," the system will prioritize recommending videos or interactive design works with similar characteristics in their subsequent queries.

[0076] Regarding automatic tag synthesis, the system continuously monitors tag co-occurrence frequency. In its second month of operation, it was found that the co-occurrence frequency of the tags "UI design" and "mobile application" consistently exceeded the co-occurrence threshold. The system was able to automatically synthesize a new tag "mobile UI" and integrate it and its initial weight into the explicit semantic feature vector. Subsequently, "mobile UI" was back-labeled onto all related works, enabling users searching for "mobile UI" to find related works more directly and comprehensively, rather than relying solely on searching for the combination of "UI design" and "mobile application," thus improving search efficiency and accuracy.

[0077] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method is applicable to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0078] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A method for constructing a creative design shared database, characterized in that, The method includes: Receive creative design submissions and their corresponding initial descriptions; Content feature analysis is performed on the creative design work files to extract implicit visual feature vectors; Based on the corresponding initial description information and user community behavior data, an explicit semantic feature vector is generated and dynamically updated. By fusing the implicit visual feature vectors with the explicit semantic feature vectors, enhanced metadata corresponding to the creative design work file is constructed; Based on the enhanced metadata of all creative design work files, the association strength value between works is calculated, and a graph database is constructed with works as nodes and the association strength value as weights. In response to user queries, the system performs path mining in the graph database and collaborative filtering to generate a personalized list of recommended works.

2. The method for constructing a creative design shared database according to claim 1, characterized in that, The extracted latent visual feature vectors include: Obtain the file type parameters of the creative design work file; Based on the file type parameter, the corresponding feature extraction model is matched and selected from the preset model library; The creative design file is processed using the aforementioned feature extraction model to generate implicit visual feature vectors.

3. The method for constructing a creative design shared database according to claim 2, characterized in that, The generation and dynamic updating of explicit semantic feature vectors include: The initial description information is semantically parsed to extract an initial tag set; By analyzing the user community behavior data, we can obtain the tag popularity parameter and the tag co-occurrence frequency parameter. Based on the initial tag set, the tag popularity parameter, and the tag co-occurrence frequency parameter, the tag weights are dynamically calculated to generate explicit semantic feature vectors.

4. The method for constructing a creative design shared database according to claim 3, characterized in that, The construction of the graph database, with works as nodes and the association strength value as weight, includes: Calculate the similarity parameter between the enhanced metadata of different creative design works; The similarity parameter is compared with a preset dynamically adjusted association threshold to generate an association determination result; Based on the association determination results, an association strength value is generated, and weighted edges with the association strength value are established between the corresponding work nodes to construct a graph database.

5. The method for constructing a creative design shared database according to claim 4, characterized in that, The generated personalized work recommendation list includes: Respond to user query requests and parse them to generate query feature vectors; The graph database is traversed starting from the query feature vector to obtain a set of related works. Obtain the current user's historical behavioral preference data, and use collaborative filtering to reorder the associated works set to generate a personalized works recommendation list.

6. The method for constructing a creative design shared database according to claim 1, characterized in that, The method further includes: Identify different iterations of the same creative design file and generate version association identifiers; Record the change logs and difference information for each version to form version change data; Based on the version association identifier and the version change data, a version tree structure is constructed in the graph database.

7. The method for constructing a creative design shared database according to claim 1, characterized in that, The method further includes: Structured rights management information is embedded in the enhanced metadata; The authorization type and usage permission parameters are parsed from the rights management information; Based on the usage permission parameters, a traceable digital watermark is added to the creative design work file.

8. The method for constructing a creative design shared database according to claim 3, characterized in that, The method further includes: Monitor the tag co-occurrence frequency parameter and identify strongly associated tag combinations that consistently exceed the co-occurrence threshold; Based on the strongly associated tag combinations, new tags are automatically synthesized; The new labels and their initial weights are integrated into the explicit semantic feature vector and then back-annotated to the relevant creative design work files.

9. A method for constructing a creative design shared database according to claim 4, characterized in that, The method further includes: Collect user interaction data when browsing related works and generate interaction popularity factors; The association strength value is updated by combining the similarity parameter and the interaction heat factor; The updated association weight values ​​are used as the final weights of the weighted edges to update the associations in the graph database.

10. A creative design shared database construction system, characterized in that, The system includes: The data receiving module is used to receive creative design work files and corresponding initial description information; The latent feature extraction module is used to perform content feature analysis on the creative design work file and extract latent visual feature vectors. The explicit feature extraction module is used to generate and dynamically update explicit semantic feature vectors based on the corresponding initial description information and user community behavior data. The feature enhancement module is used to fuse the implicit visual feature vectors and the explicit semantic feature vectors to construct enhanced metadata corresponding to the creative design work file; The database construction module is used to calculate the association strength value between works based on the enhanced metadata of all creative design work files, and to construct a graph database with works as nodes and the association strength value as weights. The works recommendation module is used to respond to user query requests by mining related paths in the graph database and performing collaborative filtering to generate a personalized works recommendation list.

Images