Big data-based animation design material library management system

Abstract

The application relates to the technical field of animation design material management, in particular to an animation design material library management system based on big data, which comprises a material collection module, a preprocessing standardization module, an intelligent clustering archiving module, a network construction module, an intention recommendation module and a dynamic optimization module. The system collects multi-source heterogeneous animation original data, carries out standardization processing, and automatically classifies and archives the data according to visual and semantic correlation degrees by using an improved clustering algorithm. On this basis, a multi-level semantic correlation network containing derivation, co-occurrence and style subordination relations is constructed. The network supports an intelligent recommendation process to respond to user creation intention description and output an ordered list of associated materials. The system also dynamically optimizes clustering parameters and network connection strength by analyzing user usage behavior. The application realizes deep semantic organization and intelligent recommendation of materials, and improves the management and utilization efficiency of animation materials.

Description

Technical Field

[0001] This invention relates to the field of animation design material management technology, and in particular to an animation design material library management system based on big data. Background Technology

[0002] In the current animation industry, the efficiency of managing and accessing design materials directly impacts the creative process. Existing animation material libraries mostly employ manual or semi-automatic classification and archiving methods based on keyword tags. This approach relies on the accuracy and completeness of manual annotation. However, given the massive amounts of diverse and heterogeneous animation materials, such as 3D models and motion sequences, manual annotation is costly and inconsistent. Some systems attempt to introduce automated clustering techniques based on visual features, but these typically calculate similarity based only on low-level features like color and texture, failing to grasp the high-level semantic concepts and creative connections within the materials. For example, a model of a "cyberpunk-style cityscape" and a vector image of "retro mechanical gears" may differ significantly in purely visual features, but they possess a strong semantic connection under a specific creative theme. Conventional clustering methods cannot effectively identify such relationships.

[0003] At the material retrieval and recommendation level, existing technologies mostly rely on precise matching of user-input keywords and material tags, or recommendations based on simple collaborative filtering. The connections established in this way are flat and static, usually limited to basic attributes such as category and author, lacking the exploration and organization of deep and diverse creative connections between materials. Multidimensional networks that are highly inspiring for creation, such as derivative relationships, stylistic inheritance, and co-occurrence relationships in typical scenarios, have not been effectively constructed. This makes it difficult for the system to understand vague, intent-based user descriptions, and it cannot proactively recommend material combinations that are logically closely related but have different surface characteristics, limiting the level of creative intelligence assisted by the material library. This invention aims to address the problem of insufficient semantic understanding in automated archiving and construct a deep material association network capable of supporting creative reasoning. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a big data-based animation design resource library management system.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a big data-based animation design resource library management system, comprising: The material acquisition module collects multi-source heterogeneous raw data for animation creation, forming an initial material set that includes images, vector graphics, 3D models, and motion capture sequences. The preprocessing standardization module performs standardization preprocessing operations on the initial material set to generate a material stream to be analyzed with a unified format encoding. The standardization preprocessing operations include parsing file header information, extracting metadata tags, and performing color space conversion. The intelligent clustering and archiving module calls an improved clustering algorithm to automatically classify and archive the material stream to be analyzed. The improved clustering algorithm is optimized based on the visual similarity and semantic relevance between the materials. The network construction module establishes a multi-level semantic association network for automatically classified and archived materials. The multi-level semantic association network includes derivative relationships between materials, co-occurrence relationships of usage scenarios, and style and genre affiliation relationships. The intent recommendation module constructs an intelligent material recommendation process based on the multi-level semantic association network. The intelligent material recommendation process responds to the user's creative intent description and outputs an ordered list of related materials. The dynamic optimization module dynamically updates the parameter weights of the improved clustering algorithm and the connection strength of the multi-level semantic association network based on user behavior data regarding recommended materials.

[0006] As a further aspect of the present invention, the initial material set is subjected to a standardized preprocessing operation to generate a material stream to be analyzed with a unified format encoding, including: Identify the inherent encoding methods and structural features of each media format in the initial material set; For image-based materials, resolution normalization and color profile conversion are performed to convert all images to the preset color space and bit depth. For vector graphics materials, their path description language is parsed, converted into a unified and standardized path data structure, and the characteristic instructions related to specific rendering engines are stripped away. For 3D model materials, their mesh topology, material definition and skeleton binding information are separated and converted into neutral geometric description format, material description format and skeleton hierarchy description format, respectively. For motion capture sequences, their timelines are aligned to a unified time base, and all skeletal animation data is redirected to a standard humanoid skeleton template; To generate metadata tags containing basic attributes, technical parameters, and conversion logs for each type of material to be converted, all materials and their corresponding metadata tags are encapsulated according to a unified data encapsulation standard to form a continuous stream of materials to be analyzed.

[0007] As a further aspect of the present invention, the improved clustering algorithm is optimized based on the visual similarity and semantic relevance between materials, and its working principle includes: Extract the low-level visual feature vector and high-level semantic feature vector of the material from the material stream to be analyzed; Initial visual weights are assigned to the bottom-level visual feature vectors, and initial semantic weights are assigned to the high-level semantic feature vectors. The initial semantic weights are adaptively set according to the material type. Calculate the comprehensive distance between any two materials, where the comprehensive distance is the sum of the Euclidean distance between the visual feature vectors of the two materials multiplied by the visual weight, plus the sum of the cosine distance between the semantic feature vectors of the two materials multiplied by the semantic weight; Initialize a specified number of cluster centers, and based on the comprehensive distance, use an iterative optimization strategy to assign each material to the category of the nearest cluster center; After each iteration, the cluster center of each category is recalculated, and the ratio of the visual weight to the semantic weight in the next iteration is dynamically adjusted based on the dispersion of the material feature vectors within the category. When the difference between the categories generated by two consecutive iterations is less than the set threshold, or when the maximum number of iterations is reached, the iteration stops, and the final material clustering results and the most representative features of each cluster are output.

[0008] As a further aspect of the present invention, extracting the low-level visual feature vector and high-level semantic feature vector of the material from the material stream to be analyzed includes: For images and vector graphics, activation values ​​of multiple layers are extracted through a pre-trained deep convolutional neural network. The activation values ​​of specific layers are pooled to form the bottom-level visual feature vector. At the same time, text descriptions are obtained through an image description generation model, and the text descriptions are input into a semantic encoding model to obtain the high-level semantic feature vector. For a 3D model, its shape distribution descriptor, reflectance statistical features, and normal direction histogram are calculated and fused to form the underlying visual feature vector. At the same time, its attached material name and usage label are parsed and encoded through a word vector model to form the high-level semantic feature vector. For motion capture sequences, the key point sequence of their motion trajectory, motion energy spectrum, and posture angle change curve are calculated and fused to form the underlying visual feature vector. At the same time, based on their action category label and emotion label, the high-level semantic feature vector is obtained through semantic encoding.

[0009] As a further aspect of the present invention, the step of dynamically adjusting the ratio of the visual weight to the semantic weight in the next iteration based on the dispersion of the feature vectors of materials within the category includes: Calculate the average distance between the visual feature vectors of all materials in each category and the visual cluster center of the current category in the current iteration, and denote it as the visual dispersion. Calculate the average distance between the semantic feature vectors of all materials in each category and the semantic cluster center of the current category in the current iteration, and denote it as semantic dispersion. For a category, if its visual dispersion is greater than its semantic dispersion, then in the next iteration, when calculating the comprehensive distance for the current category, the proportion of visual weight will be increased. For a category, if its semantic dispersion is greater than its visual dispersion, then in the next iteration, when calculating the comprehensive distance for the current category, the proportion of semantic weight will be increased. The specific magnitude of the weight adjustment is related to the ratio of the two discretenesses by a preset function, and the adjustment must ensure that the sum of the visual weight and the semantic weight remains constant in each iteration.

[0010] As a further aspect of the present invention, the step of establishing a multi-level semantic association network for the automatically classified and archived materials includes: Construct an initial relationship graph using each material as a network node and the relationships between materials as edges; Establish derivative relationship edges between materials. If a material is directly modified, combined, or redrawn based on another material, then establish a strong connection derivative relationship edge between the two materials and record the derivative direction. Establish co-occurrence relationship edges between materials. If two materials are frequently used simultaneously in the same scene or shot in historical projects, then establish a co-occurrence relationship edge between the two materials. The weight of the edge is positively correlated with the co-occurrence frequency. Establish style and genre affiliation edges between materials. Based on the visual characteristics and metadata tags of the materials, classify them into specific style and genre, and establish affiliation edges between materials belonging to the same style and genre. The derived relation edges, co-occurrence relation edges, and subordinate relation edges are superimposed to form a directed weighted graph with multiple edge types and weights, namely the multi-level semantic association network.

[0011] As a further aspect of the present invention, the intelligent material recommendation process constructed based on the multi-level semantic association network includes: Receive text describing the user's creative intent; Natural language processing is performed on the text describing the creative intent to extract key themes, sentiment words, and desired style keywords; Using the aforementioned key themes, sentiment terms, and style keywords as seeds, multiple steps are taken in the multi-level semantic association network; During the exploration, prioritize exploring along the edges of derivative and subordinate relationships to discover materials with consistent style or shared origins; At the same time, based on the weight of the co-occurrence relationship edge, material nodes that have historically frequently co-occurred with seed features are assigned a high access probability; Collect the material nodes visited during the walk, their visit counts, and semantic similarity to the seed features, as an initial candidate set; For the materials in the initial candidate set, a comprehensive recommendation score is calculated by taking into account their access frequency, semantic similarity, freshness and popularity. Sort all candidate materials in descending order according to the overall recommendation score, generate an ordered list of related materials, and output it.

[0012] As a further aspect of the present invention, natural language processing is performed on the text describing the creative intent, including: The text is segmented and stop words are removed to obtain the original keyword sequence; The original keywords were expanded using a domain dictionary, and synonyms and related terms related to animation design were added to form an expanded keyword set. Sentiment analysis models are used to determine the overall sentiment polarity and intensity of the text, core words representing sentiment tendencies are extracted, and potential topic distributions are extracted from the text through topic models, selecting the most probable topics as key topics. The extended keyword set, core emotional vocabulary, and key themes are collectively used as a set of semantic features representing creative intent.

[0013] As a further aspect of the present invention, the step of dynamically updating the parameter weights of the improved clustering algorithm and the connection strength of the multi-level semantic association network based on user behavior data regarding recommended materials includes: Record user interactions with the ordered list of recommended materials, including viewing details, downloading, saving, and referencing in actual projects; When a user downloads or uses a recommended material, the semantic association strength between the recommended material and the seed intent feature on which it was recommended is enhanced, and the weight of the corresponding edge is increased in the multi-level semantic association network. When a user views or downloads multiple materials from the same cluster consecutively, the internal consistency of the cluster is strengthened, and higher weights are assigned to the feature dimensions represented by the cluster in the next round of clustering algorithm execution. If a user frequently combines materials from different clusters, then in the multi-level semantic association network, co-occurrence relationship edges are established or strengthened between these two clusters or between materials representing the clusters. The initial weight allocation strategy and distance calculation formula in the improved clustering algorithm are fine-tuned periodically using accumulated user behavior data, and the edge weights in the semantic association network are recalibrated.

[0014] As a further aspect of the present invention, the system further includes: The version and evolution management module is used to build a version evolution tree of the material, recording the different versions of each material from the original version through each modification and derivative creation. In the multi-level semantic association network, the different versions of the material are treated as independent but closely related nodes, and they are connected by strong derivative relationship edges. When recommending materials, if a version of a material is recommended and adopted by a user, the system will also trace its version evolution tree and include other versions that are visually or semantically highly consistent with the adopted version in the scope of subsequent recommendations. The system continuously monitors changes in the usage popularity and correlation strength of materials, downgrades or archives material nodes that have not been used for a long time and have weak correlation, and quickly identifies and enhances the network of emerging hot materials and their correlation relationships.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: An improved clustering algorithm, optimized based on visual similarity and semantic relevance, is employed to process the stream of materials to be analyzed, overcoming the limitations of traditional methods that rely solely on single-dimensional features. The algorithm simultaneously analyzes the low-level visual features and high-level semantic information of the materials, performing joint optimization and measurement in the feature space. This allows the system to aggregate materials with different visual representations but belonging to the same theme, style, or fulfilling the same functional requirements into the same category. For example, it can group together the same character drawn from different perspectives or texture materials with significant visual differences but all conveying a sense of "fragmentation." Simultaneously, it can effectively distinguish visually similar but semantically dissimilar materials, reducing misclassification. This processing significantly improves the accuracy and semantic rationality of automated archiving of massive amounts of multi-source heterogeneous materials. The resulting classification structure is closer to the cognitive logic of creative personnel, providing a high-quality, semantically consistent archiving foundation for subsequent deep retrieval and recommendation.

[0016] This system constructs a multi-layered semantic association network encompassing derivative relationships between materials, co-occurrence relationships in usage scenarios, and stylistic affiliations, transforming the traditional flat tag-based association structure of material libraries. This network dynamically captures and formally records the complex, non-linear creative logical connections between materials. For example, a basic model forms a derivative relationship chain with its multiple variant models; background and prop materials that frequently appear together in the same scene design establish strong co-occurrence relationships; and different materials belonging to specific art styles form stylistic affiliations. This networked organization prevents materials from existing in isolation, embedding them within a rich creative context. The recommendation process based on this network enables multi-hop reasoning and association mining, accurately responding to user descriptions based on creative intent. When a user inputs their intent, the system not only finds directly matching materials but also discovers sets of materials related in style coordination, scene adaptation, or creative evolution along the network path, achieving a leap from "retrieval" to "creative association," directly stimulating creative inspiration and improving material reuse efficiency. Attached Figure Description

[0017] Figure 1 This is a sequence diagram of the big data-based animation design resource library management system described in this invention; Figure 2 A flowchart for generating the stream of materials to be analyzed is generated through standardized preprocessing of the initial material set; Figure 3 A flowchart for constructing a multi-level semantic association network for the materials. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] See Figure 1The big data-based animation design resource library management system described in this invention has the following overall implementation scheme: The system, through a resource acquisition module, collects multi-source heterogeneous animation creation raw data, including images, vector graphics, 3D models, and motion capture sequences, from multiple sources such as publicly available internet resource libraries, professional design platforms, enterprise internal asset libraries, and motion capture equipment, forming an initial resource set. A preprocessing and standardization module performs standardization preprocessing operations on this initial resource set, specifically including parsing the header information of various files, extracting and standardizing their metadata tags, and performing unified conversion across color spaces, thereby generating a unified format encoded resource stream to be analyzed. An intelligent clustering and archiving module calls an improved clustering algorithm to process the resource stream to be analyzed. This algorithm optimizes the data by combining visual similarity and semantic relevance between materials, achieving automatic classification and archiving. Based on this, a network construction module establishes a multi-level semantic association network for the classified and archived materials. This network characterizes the derivative relationships, co-occurrence relationships of usage scenarios, and style / genre affiliation relationships between material nodes. The intent recommendation module, based on the constructed multi-level semantic association network, designs an intelligent material recommendation process. This process can parse the user's input description of creative intent, perform relevant calculations in the network, and finally output an ordered list of related materials. The dynamic optimization module continuously collects user behavior data such as viewing, downloading, and referencing recommended materials. Based on this feedback, it dynamically adjusts the parameter weights in the improved clustering algorithm and updates the connection strength of various relationships in the multi-level semantic association network, enabling the system to self-evolve.

[0021] In one embodiment of the present invention, see [reference] Figure 2 The process involves standardizing and preprocessing the initial media set to generate a stream of media to be analyzed, encoded in a unified format. This process identifies the inherent encoding methods and structural characteristics of each media format in the initial set. For image-based media, resolution normalization and color profile conversion are performed, transforming all images to a preset sRGB color space and 8-bit channel depth. For vector graphics, the path description language in their SVG or AI files is parsed, converted into a unified standardized path data structure, and specific rendering engine-related feature instructions are removed. For 3D model-based media, their mesh topology, material definitions, and skeleton binding information are separated. Mesh data is converted to OBJ or glTF format, material information is converted to a standardized Physically Based Rendering (PBR) description, and skeleton information is converted to a common hierarchical joint description format. For motion capture sequences, their timelines are aligned to a unified 30 frames per second time base, and all skeletal animation data is redirected to a standard humanoid skeleton template. Metadata tags containing basic attributes, technical parameters, and conversion logs are generated for each type of media after conversion. All media and their corresponding metadata tags are encapsulated according to a unified data encapsulation specification, forming a continuous stream of media to be analyzed.

[0022] The intelligent clustering and archiving module uses an improved clustering algorithm to automatically classify and archive the source material stream. This improved algorithm extracts the low-level visual feature vectors and high-level semantic feature vectors from the stream. Initial visual weights are assigned to the low-level visual feature vectors, and initial semantic weights are assigned to the high-level semantic feature vectors. The initial semantic weights are adaptively set based on the source material type. The comprehensive distance between any two source materials is calculated, which is the sum of the Euclidean distance between their visual feature vectors multiplied by the visual weights, plus the cosine distance between their semantic feature vectors multiplied by the semantic weights. A specified number of cluster centers are initialized. Based on this comprehensive distance, an iterative optimization strategy is used to assign each source material to the category containing the nearest cluster center. After each iteration, the cluster centers for each category are recalculated, and the ratio of visual weights to semantic weights in the next iteration is dynamically adjusted based on the dispersion of the source material feature vectors within each category. When the difference between the category divisions generated by two consecutive iterations is less than a set threshold, or when the maximum number of iterations is reached, the iteration stops, and the final source material clustering results and the most representative features of each cluster are output.

[0023] In practical implementation, the preprocessing standardization module of the big data-based animation design material library management system performs standardized preprocessing operations on the initial material set. This initial material set includes images, vector graphics, 3D models, and motion capture sequences collected from multiple sources. Identifying the inherent encoding methods and structural characteristics of each media format in the initial material set is the starting point for the preprocessing standardization module. For image materials, resolution normalization and color profile conversion are performed, converting all images to the preset sRGB color space and 8-bit channel depth. For vector graphics materials, the path description language in their SVG or AI files is parsed and converted into a unified standardized path data structure. By stripping away feature instructions specific to a particular rendering engine, and for 3D model-type materials, their mesh topology, material definitions, and skeletal binding information are separated. Mesh data is converted to OBJ format, material information is converted to a standardized description based on physically based rendering, and skeletal information is converted to a universal hierarchical joint description format. For motion capture sequences, their timelines are aligned to a unified 30 frames per second time base, and all skeletal animation data is redirected to a standard humanoid skeleton template. Metadata tags containing basic attributes, technical parameters, and conversion logs are generated for each type of material that has been converted. All materials and their corresponding metadata tags are encapsulated according to a unified data encapsulation specification to form a continuous stream of materials to be analyzed.

[0024] In some embodiments, the intelligent clustering and archiving module invokes an improved clustering algorithm to automatically classify and archive the material stream to be analyzed. The improved clustering algorithm extracts the low-level visual feature vectors and high-level semantic feature vectors of the material stream to be analyzed, assigns initial visual weights to the low-level visual feature vectors, and assigns initial semantic weights to the high-level semantic feature vectors. The initial semantic weights are adaptively set according to the material type. The comprehensive distance between any two materials is calculated, and the formula for the comprehensive distance is defined as:

[0025] in: This represents the combined distance between material i and material j. and Let i and j represent the visual feature vectors of material i and material j, respectively. and Let i and j represent the semantic feature vectors of material i and material j, respectively. Indicates visual weight, Represents semantic weight, Represents Euclidean distance. This represents the vector dot product. A specified number of cluster centers are initialized. Based on the comprehensive distance, an iterative optimization strategy is used to assign each material to the category containing the nearest cluster center. After each iteration, the cluster centers for each category are recalculated. Based on the dispersion of the material feature vectors within a category, the ratio of visual weights to semantic weights in the next iteration is dynamically adjusted. The iteration stops when the difference between category divisions generated by two consecutive iterations is less than a set threshold, or when the maximum number of iterations is reached. The final material clustering result and the most representative features of each cluster are output. In essence, the improved clustering algorithm optimizes based on the visual similarity and semantic relevance between materials. In practice, the dynamic adjustment of visual and semantic weights is based on a comparison of dispersion within a category. For a category, if the visual dispersion is greater than the semantic dispersion, the proportion of visual weights is increased in the next iteration; if the semantic dispersion is greater than the visual dispersion, the proportion of semantic weights is increased. The specific magnitude of the weight adjustment has a preset functional relationship with the ratio of the two dispersions, and the adjustment must ensure that the sum of visual and semantic weights remains constant in each iteration.

[0026] Optionally, in the standardized preprocessing operations, for image materials, resolution normalization is performed by scaling while maintaining the aspect ratio, scaling the longest side of the image to 1024 pixels, and filling any insufficient areas with transparent pixels. Color space conversion involves converting images with non-standard color profiles to a preset sRGB color space to ensure color consistency. In some embodiments, the iterative process of the improved clustering algorithm starts with randomly selecting cluster centers and gradually optimizes the clustering results through multiple iterations. When calculating the comprehensive distance, both the visual feature vector and the semantic feature vector are normalized to eliminate the influence of dimensions. It can be understood that the recalculation of cluster centers adopts the mean method, that is, taking the average value of the feature vectors of all materials within the category.

[0027] Optionally, the preprocessing standardization module generates metadata tags for the materials, including material name, format type, creation time, file size, color mode, resolution, copyright information, as well as parameters and logs used during the conversion process. These metadata tags are encapsulated together with the material data in a unified container format to form the material stream to be analyzed.

[0028] In one embodiment of the present invention, low-level visual feature vectors and high-level semantic feature vectors of the material to be analyzed are extracted from the material stream. For images and vector graphics, activation values ​​of multiple layers are extracted using a pre-trained deep convolutional neural network. The activation values ​​of specific layers are pooled to form low-level visual feature vectors. Simultaneously, text descriptions are obtained through an image description generation model, and these text descriptions are input into a semantic encoding model to obtain high-level semantic feature vectors. For 3D models, shape distribution descriptors, reflectance statistical features, and normal direction histograms are calculated and fused to form low-level visual feature vectors. Simultaneously, the attached material names and usage labels are parsed and encoded using a word vector model to form high-level semantic feature vectors. For motion capture sequences, keypoint sequences, motion energy spectra, and attitude angle change curves of the motion trajectory are calculated and fused to form low-level visual feature vectors. Simultaneously, high-level semantic feature vectors are obtained through semantic encoding based on the action category labels and sentiment labels.

[0029] The improved clustering algorithm dynamically adjusts the ratio of visual weights to semantic weights in the next iteration based on the dispersion of feature vectors within each category. The average distance between the visual feature vectors of all materials within each category and the visual cluster center of the current category is calculated; this distance is denoted as visual dispersion. The average distance between the semantic feature vectors of all materials within each category and the semantic cluster center of the current category is also calculated; this distance is denoted as semantic dispersion. For a category, if its visual dispersion is greater than its semantic dispersion, the proportion of visual weights is increased when calculating the overall distance for the current category in the next iteration. For a category, if its semantic dispersion is greater than its visual dispersion, the proportion of semantic weights is increased when calculating the overall distance for the current category in the next iteration. The specific magnitude of the weight adjustment follows a preset functional relationship with the ratio of the two dispersions, and the adjustment must ensure that the sum of visual weights and semantic weights remains constant in each iteration.

[0030] In practice, the process of extracting the low-level visual feature vectors and high-level semantic feature vectors from the source material to be analyzed involves, for images and vector graphics, extracting the activation values ​​of multiple layers through a pre-trained deep convolutional neural network. The pre-training of the deep convolutional neural network is completed on a large public image dataset. The activation values ​​of a specific layer are then globally averaged to form the low-level visual feature vectors. At the same time, the text descriptions are obtained through an image description generation model. The image description generation model is based on an encoder-decoder architecture, which converts the visual information of images and vector graphics into natural language descriptions. The obtained text descriptions are then input into a semantic encoding model to obtain high-level semantic feature vectors. The semantic encoding model adopts a bidirectional encoding representation based on Transformer.

[0031] For 3D models, the shape distribution descriptor, reflectance statistical features, and normal direction histogram are calculated. The shape distribution descriptor is obtained by statistically analyzing the distance distribution between random point pairs on the 3D model surface. The reflectance statistical features are extracted from the material texture of the 3D model, showing the mean and variance of color and gloss. The normal direction histogram statistically quantifies the directional distribution of all normal vectors on the 3D model surface. After fusion, a low-level visual feature vector is formed. Simultaneously, the material name and usage tags attached to the 3D model are parsed and encoded using a word vector model to form a high-level semantic feature vector. The word vector model maps text tags to dense vectors of fixed dimensions. For motion capture sequences, the keypoint sequence, motion energy spectrum, and posture angle change curve of the motion trajectory are calculated. The keypoint sequence of the motion trajectory records the temporal position of the main joints of the human body in 3D space. The motion energy spectrum is obtained by performing a Fourier transform on the joint motion velocity sequence. The posture angle change curve describes the change of joint angles over time. After fusion, a low-level visual feature vector is formed. Simultaneously, based on the motion category tags and sentiment tags of the motion capture sequence, a high-level semantic feature vector is obtained through semantic encoding.

[0032] The improved clustering algorithm dynamically adjusts the ratio of visual weights to semantic weights in the next iteration based on the dispersion of feature vectors within each category. The average Euclidean distance between the visual feature vectors of all materials within each category and the visual cluster center of the current category in the current iteration is calculated; this average distance is denoted as the visual dispersion. The average cosine distance between the semantic feature vectors of all materials within each category and the semantic cluster center of the current category in the current iteration is also calculated; this average distance is denoted as the semantic dispersion. For a category, if the visual dispersion is greater than the semantic dispersion, the proportion of visual weights is increased when calculating the overall distance for the current category in the next iteration. For a category, if the semantic dispersion is greater than the visual dispersion, the proportion of semantic weights is increased when calculating the overall distance for the current category in the next iteration. The specific magnitude of the weight adjustment has a preset functional relationship with the ratio of the two dispersions. The sum of the visual weights and semantic weights remains constant in each adjustment. The adjustment function is defined as:

[0033] in: This represents the absolute value of the weight adjustment. This represents the preset learning rate parameter. Indicates the category in the current iteration Visual dispersion, Indicates the category in the current iteration The direction of weight adjustment for semantic dispersion is determined based on the relationship between visual dispersion and semantic dispersion.

[0034] In some embodiments, visual dispersion is calculated using the average Euclidean distance from all visual feature vectors of all materials within a category to the visual cluster center of the category, and semantic dispersion is calculated using the average cosine distance from all semantic feature vectors of all materials within a category to the semantic cluster center of the category. It can be understood that the visual cluster center is the mean vector of all visual feature vectors of all materials within a category, and the semantic cluster center is the mean vector of all semantic feature vectors of all materials within a category. In each iteration, the visual weights and semantic weights of each category are calculated and adjusted independently. Optionally, a preset learning rate parameter is used. A positive number less than 1 is set to control the magnitude of weight adjustment, preventing drastic weight changes in a single iteration. In some embodiments, both the visual feature vector and the semantic feature vector undergo L2 norm normalization before input distance calculation to ensure the stability of the distance metric. Optionally, for newly added materials, the extraction process of their low-level visual feature vector and high-level semantic feature vector is the same as that of existing materials. After extraction, the feature vectors of the materials are sent to an improved clustering algorithm for archiving.

[0035] In one embodiment of the present invention, a multi-level semantic association network is established for the automatically classified and archived materials. See also... Figure 3 An initial association graph is constructed using each material as a network node and the relationships between materials as edges. Derivative relationship edges are established between materials. If a material is directly modified, combined, or redrawn from another material, a strong derivative relationship edge is established between the two materials, and the direction of derivation is recorded. Co-occurrence relationship edges are established between materials based on usage scenarios. If two materials are frequently used simultaneously in the same scene or shot in historical projects, a co-occurrence relationship edge is established between the two materials, with the edge weight positively correlated with the co-occurrence frequency. Style / genre affiliation relationship edges are established between materials. Based on the visual characteristics and metadata tags of the materials, they are categorized into specific style / genre, and affiliation relationship edge is established between materials belonging to the same style / genre. The derivative relationship edges, co-occurrence relationship edges, and affiliation relationship edges are superimposed to form a directed weighted graph with multiple edge types and weights, i.e., the multi-level semantic association network.

[0036] In practical implementation, a multi-level semantic association network is established for the automatically categorized and archived materials. Each material is treated as a network node, and the relationships between materials are used as edges to construct an initial association graph. This initial graph contains all material nodes that have undergone intelligent clustering and archiving. Derivative relationship edges are established between materials. If a material is directly modified, combined, or redrawn from another material, a strong derivative relationship edge is established between the two materials, and the derivative direction is recorded, pointing from the original material to the derived material. Co-occurrence relationship edges are established between materials based on usage scenarios. If two materials are frequently used simultaneously in the same scene or shot in historical projects, a co-occurrence relationship edge is established between them, with the weight of the co-occurrence relationship edge positively correlated with the co-occurrence frequency. Style / genre affiliation relationship edges are established between materials. Based on the visual characteristics and metadata tags of the materials, they are categorized into specific style / genre, and affiliation relationship edges are established between materials belonging to the same style / genre. The derivative relationship edges, co-occurrence relationship edges, and affiliation relationship edges are superimposed to form a directed weighted graph with multiple edge types and weights. This directed weighted graph is the multi-level semantic association network.

[0037] In practical implementation, the multi-level semantic association network is stored and computed in the form of an adjacency matrix. The edge weights in the network are initialized according to the relationship type and strength, with initial values ​​set based on historical data statistics or preset rules. It can be understood that derived relationship edges are directional, and their weights are usually set to fixed high values ​​to represent strong associations. The weights of co-occurrence relationship edges are obtained by statistically analyzing the number of times two materials co-occur in the historical project dataset and then normalizing the calculation. Dependent relationship edges are undirected, and their weights are calculated based on the similarity between the material and the central features of the style / genre. Edges of different relationship types have different merging strategies when superimposed. Optionally, the rules for initializing and calculating edge weights are shown in Table 1.

[0038] Table 1: Initialization Rules for Edge Weights in Multi-Level Semantic Association Networks Derivative relation edge promising The weight is fixed at 0.9. The derivation direction points from the parent material to the child material. Co-occurrence relation edge Undirected Weight = min(1.0, co-occurrence count / 10). That is, the number of times material A and material B appear together in all historical projects divided by 10, with an upper limit of 1.0. Subordinate relationship edge Undirected Weight = Cosine Similarity (feature vector of source material, feature vector of style / genre center). The feature vector of style / genre center is obtained by averaging the feature vectors of all source materials belonging to that style / genre. In some embodiments, the establishment of derivative relationship edges relies on version management logs of materials or explicitly labeled derivative relationships by users. Co-occurrence relationship edges are established by analyzing project files and statistically analyzing the occurrence of different material identifiers within the same scene container. Dependent relationship edges are established based on the results of intelligent clustering archiving, treating materials within the same cluster as belonging to the same style or genre, with the cluster center feature vector serving as the style or genre center feature vector. It can be understood that the multi-level semantic association network is a dynamically evolving graph structure. During system operation, the network construction module continuously monitors new material additions, project usage records, and user feedback, and adds, deletes, and modifies nodes and edges in the network based on the new data. The network construction module periodically executes community detection algorithms to identify closely connected material clusters in the network; these clusters may correspond to unlabeled sub-styles or theme series.

[0039] Optionally, the edge weights are superimposed following a linear weighting principle. For the various possible relationships between node A and node B, the final connection strength is determined. Calculated using the following formula:

[0040] in: This represents the overall connection strength between node A and node B. This indicates the weight of the derived relation edge. This indicates the weight of the co-occurrence relation edge. This indicates the weight of the dependency edge. , , These represent the preset fusion coefficients for corresponding edge types, used to adjust the contribution ratio of different relationship types to the overall connection strength. Each node in the multi-level semantic association network maintains a neighbor list, which records the identifiers of all other nodes connected to that node, the connection relationship type, and the corresponding edge weights. The network construction module provides an application programming interface (API) for the intent recommendation module to query node relationships and paths in the network.

[0041] In one embodiment of the present invention, an intelligent material recommendation process is constructed based on a multi-level semantic association network. This process receives a text describing the creative intent input by the user. Natural language processing is performed on the text to extract key themes, sentiment terms, and desired style keywords. Using the extracted key themes, sentiment terms, and style keywords as seeds, a multi-step walk is performed within the multi-level semantic association network. During the walk, priority is given to exploring along derivative and subordinate relationship edges to discover materials with consistent styles or shared origins. Simultaneously, based on the weight of co-occurrence relationship edges, material nodes that historically frequently co-occur with seed features are assigned a high access probability. The material nodes visited during the walk, their visit counts, and semantic similarity to seed features are collected as an initial candidate set. For the materials in this initial candidate set, a comprehensive recommendation score is calculated by comprehensively considering their visit counts, semantic similarity, freshness, and popularity. All candidate materials are sorted in descending order according to the comprehensive recommendation score, generating an ordered list of related materials and outputting it.

[0042] Natural language processing (NLP) was performed on the text describing the creative intent, including word segmentation and stop word removal, to obtain the original keyword sequence. The original keywords were expanded using a domain dictionary, incorporating synonyms and hypernyms related to animation design to form an expanded keyword set. A sentiment analysis model was used to determine the overall sentiment polarity and intensity of the text, extracting core words representing sentiment tendencies. A topic model was used to extract potential topic distributions from the text, selecting the most probable topics as key topics. The expanded keyword set, sentiment-oriented core words, and key topics were combined to form a semantic feature set representing the creative intent.

[0043] In its implementation, an intelligent material recommendation process is constructed based on a multi-level semantic association network. This process receives user-inputted text describing the creative intent, which is a description of the desired material features in natural language. Natural language processing is performed on the text to extract key themes, sentiment terms, and desired style keywords. These extracted keywords serve as seeds, and the process involves multiple walks within the multi-level semantic association network. During these walks, priority is given to exploring derived and subordinate relationships to discover materials with consistent styles or shared origins. Simultaneously, based on the weights of co-occurrence edges, material nodes that historically frequently co-occur with seed features are assigned a high access probability. The process collects the material nodes visited during the walks, their visit counts, and semantic similarity to the seed features, forming an initial candidate set. For materials in the initial candidate set, a comprehensive recommendation score is calculated, considering visit counts, semantic similarity, freshness, and popularity. All candidate materials are then sorted in descending order according to the comprehensive recommendation score, generating and outputting an ordered list of related materials.

[0044] In practice, natural language processing is applied to the text describing the creative intent, including word segmentation and stop word removal to obtain the original keyword sequence. The original keywords are then expanded using a domain dictionary by adding synonyms and hypernyms related to animation design, forming an expanded keyword set. A sentiment analysis model is used to determine the overall sentiment polarity and intensity of the text, extracting core words representing sentiment tendencies. A topic model is then used to extract latent topic distributions from the text, selecting the most probable topics as key topics. The expanded keyword set, sentiment-oriented core words, and key topics are collectively used as a semantic feature set representing the creative intent. The domain dictionary includes professional terms such as animation characters, scenes, props, special effects, and art style, along with their related vocabulary. The sentiment analysis model is trained on an annotated text sentiment dataset and can identify positive, neutral, and negative sentiment polarities. The topic model employs a latent Dirichlet distribution model and is trained on a corpus related to animation design.

[0045] In some embodiments, the multi-step walk in the multi-level semantic association network employs a restarted random walk algorithm. Each keyword or topic in the seed feature set corresponds to a virtual seed node in the network. The virtual seed node is connected to the actual material nodes in the network through semantic similarity. The walk starts from the virtual seed node. At each step, the walker randomly walks to a neighboring node of the current node with a certain probability, or jumps back to any seed node with a certain probability. The transition probability settings for different types of edges in the walk algorithm are different, with higher transition probability factors set for derived and subordinate relationship edges. Optionally, the access probability calculation during the walk process is shown in Table 2.

[0046] Table 2: Node Walk Transition Probability Weight Allocation Table Derivative relation edge 0.5 Wanderers are more likely to jump along this side, which is used to explore derivative chains. Subordinate relationship edge 0.3 Wanderers have a moderate probability of jumping along this side, making it suitable for exploring materials of the same style. Co-occurrence relation edge 0.2 The probability weight of a wanderer jumping along this side is based on 0.2, multiplied by the actual weight of the side, and used to explore historical combinations. It is understandable that the number of steps in the multi-step walk is set to a fixed value, and the walk is repeated multiple times to count the total number of times all actual material nodes are accessed. Access count, semantic similarity, material freshness, and popularity are weighted and fused when calculating the overall recommendation score. Material freshness is determined by the material's upload time, and popularity is determined by the number of times the material has been cited in historical projects. In some embodiments, the formula for calculating the overall recommendation score is:

[0047] in: This represents the overall recommendation score for the candidate materials. This represents the total number of times the resource node was visited during the random walk. This represents the average semantic similarity between the semantic feature vectors of candidate materials and the set of semantic features of user intent. The freshness of the material is measured by the reciprocal of the difference (in days) between the material's upload timestamp and the current timestamp. For time normalization factor, The popularity of the material is indicated by the number of times it has been cited throughout history. , , , These are weighting coefficients for visit count, semantic similarity, freshness, and popularity, used to adjust the importance of different factors in the final score. Optional, This can be set to a constant, such as 1000 days, to control the magnitude of the freshness factor. The calculated overall recommendation score is used to rank all materials entering the initial candidate set.

[0048] In one embodiment of the present invention, the parameter weights of the improved clustering algorithm and the connection strength of the multi-level semantic association network are dynamically updated based on user behavior data regarding recommended materials. User interaction with the ordered list of recommended materials is recorded, including viewing details, downloading, saving, and referencing in actual projects. When a user downloads or references a recommended material, the semantic association strength between the recommended material and the seed intent feature upon which it was recommended is enhanced, and the weight of the corresponding edge in the multi-level semantic association network is increased. When a user continuously views or downloads multiple materials from the same cluster, the internal consistency of the cluster is strengthened, and higher weights are assigned to the feature dimensions represented by the cluster in the next round of clustering algorithm execution. If a user frequently combines materials from different clusters, co-occurrence relationship edges are established or strengthened between these two clusters or between the materials represented by the clusters in the multi-level semantic association network. The initial weight allocation strategy and distance calculation formula in the improved clustering algorithm are fine-tuned periodically using accumulated user behavior data, and the edge weights in the semantic association network are recalibrated.

[0049] The system also includes a version and evolution management module, used to build a version evolution tree for the materials, recording the different versions of each material generated from each modification and derivative creation starting from the original version. In the multi-level semantic association network, different versions of materials are treated as independent but closely related nodes, connected by strong derivative relationship edges. When recommending materials, if a version of a material is recommended and adopted by a user, the system will simultaneously trace its version evolution tree, including other versions that are highly consistent with the adopted version visually or semantically in the subsequent recommendation scope. The system continuously monitors changes in the usage popularity and association strength of materials, downgrading or archiving material nodes that have not been used for a long time and have weak association, and quickly identifying and enhancing emerging hot materials and their associations.

[0050] In practical implementation, the parameter weights of the improved clustering algorithm and the connection strength of the multi-level semantic association network are dynamically updated based on user behavior data regarding recommended content. User interactions with the ordered list of recommended content are recorded, including viewing details, downloading, saving, and referencing in actual projects. When a user downloads or references a recommended content, the semantic association strength between the recommended content and the seed intent features upon which it was recommended is enhanced, and the weight of the corresponding edge in the multi-level semantic association network is increased. When a user continuously views or downloads multiple content from the same cluster, the internal consistency of the cluster is strengthened, and higher weights are assigned to the feature dimensions represented by the cluster in the next round of clustering algorithm execution. If a user frequently combines content from different clusters, co-occurrence relationship edges are established or strengthened between these two clusters or between the content represented by the clusters in the multi-level semantic association network. The initial weight allocation strategy and distance calculation formula in the improved clustering algorithm are fine-tuned periodically using accumulated user behavior data, and the edge weights in the semantic association network are recalibrated.

[0051] In its implementation, the version and evolution management module establishes a version evolution tree for the materials. This tree records the different versions of each material, starting from the original version and through various modifications and derivative creations. Within a multi-layered semantic association network, different versions of the material are treated as independent but closely related nodes, connected by strong derivative relationship edges. When recommending materials, if a version of a material is recommended and adopted by a user, the system simultaneously traces its version evolution tree, including other versions that are visually or semantically highly consistent with the adopted version in subsequent recommendations. The system continuously monitors changes in the usage popularity and association strength of materials, downgrading or archiving material nodes that have not been used for a long time and have weak associations, and quickly identifying and enhancing emerging trending materials and their associations. User behavior data is recorded in a specific log database; each record includes a user identifier, material identifier, interaction type, timestamp, and a descriptive text of the creative intent that triggered the recommendation.

[0052] In some embodiments, the weights of the edges in the multi-level semantic association network are adjusted based on the type and intensity of user interaction behavior, with user feedback such as downloading or referencing materials being stronger than viewing details or adding them to favorites. The dynamic update formula for the edge weights is:

[0053] in: This represents the new weight of the edge connecting node i and node j after adjustment. This represents the old weight of the connecting edge before adjustment. This represents the global learning rate parameter. Indicates information about user behavior types The feedback function. The values ​​are preset based on the behavior type. For example, downloading is assigned a larger positive value, referencing is assigned an even larger positive value, while viewing details is assigned a smaller positive value or zero. Optionally, the weight adjustment of the connection edges has upper and lower bounds to prevent the weight from increasing or decreasing indefinitely due to continuous feedback. For the adjustment of the clustering algorithm parameter weights, when a user continuously provides positive feedback to materials within the same cluster, the system will increase the weight of the component corresponding to the cluster center feature vector in the next round of clustering calculation. The adjustment of the cluster feature dimension weights is based on the variance of that dimension in the set of user positive feedback materials; the smaller the variance of the dimension, the greater the increase in its weight.

[0054] It is understandable that the periodic recalibration cycle of the clustering algorithm and network can be set to a fixed time interval, such as weekly or monthly. The calibration process uses all new user behavior data accumulated since the start of the calibration cycle. During calibration, the weight parameters in the distance calculation formula of the improved clustering algorithm and the initial positions of the cluster centers are optimized based on user behavior data. The weights of all edges in the multi-level semantic association network are globally recalculated based on the edge type, time decay, and user interaction data. Optionally, the system maintains a "popularity" score for each media node. The popularity score is calculated based on the frequency with which the node has been accessed, downloaded, and referenced by users recently, and decays over time. The system continuously monitors changes in the usage popularity and association strength of media, downgrading or archiving media nodes that have not been used for a long time and have weak associations, and quickly identifying and enhancing emerging hot media and their associations. The identification of hot media is based on the rate at which their popularity score exceeds a preset threshold within a short period of time.

[0055] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A big data-based animation design resource library management system, characterized in that, The system includes: The material acquisition module collects multi-source heterogeneous raw data for animation creation, forming an initial material set that includes images, vector graphics, 3D models, and motion capture sequences. The preprocessing standardization module performs standardization preprocessing operations on the initial material set to generate a material stream to be analyzed with a unified format encoding. The standardization preprocessing operations include parsing file header information, extracting metadata tags, and performing color space conversion. The intelligent clustering and archiving module calls an improved clustering algorithm to automatically classify and archive the material stream to be analyzed. The improved clustering algorithm is optimized based on the visual similarity and semantic relevance between the materials. The network construction module establishes a multi-level semantic association network for automatically classified and archived materials. The multi-level semantic association network includes derivative relationships between materials, co-occurrence relationships of usage scenarios, and style and genre affiliation relationships. The intent recommendation module constructs an intelligent material recommendation process based on the multi-level semantic association network. The intelligent material recommendation process responds to the user's creative intent description and outputs an ordered list of related materials. The dynamic optimization module dynamically updates the parameter weights of the improved clustering algorithm and the connection strength of the multi-level semantic association network based on user behavior data regarding recommended materials.

2. The big data-based animation design resource library management system according to claim 1, characterized in that, The initial set of materials is subjected to standardized preprocessing to generate a stream of materials to be analyzed with a unified format encoding, including: Identify the inherent encoding methods and structural features of each media format in the initial material set; For image-based materials, resolution normalization and color profile conversion are performed to convert all images to the preset color space and bit depth. For vector graphics materials, their path description language is parsed, converted into a unified and standardized path data structure, and the characteristic instructions related to specific rendering engines are stripped away. For 3D model materials, their mesh topology, material definition and skeleton binding information are separated and converted into neutral geometric description format, material description format and skeleton hierarchy description format, respectively. For motion capture sequences, their timelines are aligned to a unified time base, and all skeletal animation data is redirected to a standard humanoid skeleton template; To generate metadata tags containing basic attributes, technical parameters, and conversion logs for each type of material to be converted, all materials and their corresponding metadata tags are encapsulated according to a unified data encapsulation standard to form a continuous stream of materials to be analyzed.

3. The big data-based animation design resource library management system according to claim 1, characterized in that, The improved clustering algorithm is optimized based on the visual similarity and semantic relevance between materials, and its working principle includes: Extract the low-level visual feature vector and high-level semantic feature vector of the material from the material stream to be analyzed; Initial visual weights are assigned to the bottom-level visual feature vectors, and initial semantic weights are assigned to the high-level semantic feature vectors. The initial semantic weights are adaptively set according to the material type. Calculate the comprehensive distance between any two materials, where the comprehensive distance is the sum of the Euclidean distance between the visual feature vectors of the two materials multiplied by the visual weight, plus the sum of the cosine distance between the semantic feature vectors of the two materials multiplied by the semantic weight; Initialize a specified number of cluster centers, and based on the comprehensive distance, use an iterative optimization strategy to assign each material to the category of the nearest cluster center; After each iteration, the cluster center of each category is recalculated, and the ratio of the visual weight to the semantic weight in the next iteration is dynamically adjusted based on the dispersion of the material feature vectors within the category. When the difference between the categories generated by two consecutive iterations is less than the set threshold, or when the maximum number of iterations is reached, the iteration stops, and the final material clustering results and the most representative features of each cluster are output.

4. The big data-based animation design resource library management system according to claim 3, characterized in that, Extracting the low-level visual feature vector and high-level semantic feature vector of the material from the material stream to be analyzed includes: For images and vector graphics, activation values ​​of multiple layers are extracted through a pre-trained deep convolutional neural network. The activation values ​​of specific layers are pooled to form the bottom-level visual feature vector. At the same time, text descriptions are obtained through an image description generation model, and the text descriptions are input into a semantic encoding model to obtain the high-level semantic feature vector. For a 3D model, its shape distribution descriptor, reflectance statistical features, and normal direction histogram are calculated and fused to form the underlying visual feature vector. At the same time, its attached material name and usage label are parsed and encoded through a word vector model to form the high-level semantic feature vector. For motion capture sequences, the key point sequence of their motion trajectory, motion energy spectrum, and posture angle change curve are calculated and fused to form the underlying visual feature vector. At the same time, based on their action category label and emotion label, the high-level semantic feature vector is obtained through semantic encoding.

5. The big data-based animation design resource library management system according to claim 3, characterized in that, The dynamic adjustment of the ratio of visual weights to semantic weights in the next iteration based on the dispersion of feature vectors within a category includes: Calculate the average distance between the visual feature vectors of all materials in each category and the visual cluster center of the current category in the current iteration, and denote it as the visual dispersion. Calculate the average distance between the semantic feature vectors of all materials in each category and the semantic cluster center of the current category in the current iteration, and denote it as semantic dispersion. For a category, if its visual dispersion is greater than its semantic dispersion, then in the next iteration, when calculating the comprehensive distance for the current category, the proportion of visual weight will be increased. For a category, if its semantic dispersion is greater than its visual dispersion, then in the next iteration, when calculating the comprehensive distance for the current category, the proportion of semantic weight will be increased. The specific magnitude of the weight adjustment is related to the ratio of the two discretenesses by a preset function, and the adjustment must ensure that the sum of the visual weight and the semantic weight remains constant in each iteration.

6. The big data-based animation design resource library management system according to claim 1, characterized in that, The process of establishing a multi-level semantic association network for automatically classified and archived materials includes: Construct an initial relationship graph using each material as a network node and the relationships between materials as edges; Establish derivative relationship edges between materials. If a material is directly modified, combined, or redrawn based on another material, then establish a strong connection derivative relationship edge between the two materials and record the derivative direction. Establish co-occurrence relationship edges between materials. If two materials are frequently used simultaneously in the same scene or shot in historical projects, then establish a co-occurrence relationship edge between the two materials. The weight of the edge is positively correlated with the co-occurrence frequency. Establish style and genre affiliation edges between materials. Based on the visual characteristics and metadata tags of the materials, classify them into specific style and genre, and establish affiliation edges between materials belonging to the same style and genre. The derived relation edges, co-occurrence relation edges, and subordinate relation edges are superimposed to form a directed weighted graph with multiple edge types and weights, namely the multi-level semantic association network.

7. The big data-based animation design resource library management system according to claim 6, characterized in that, The intelligent material recommendation process based on the multi-level semantic association network includes: Receive text describing the user's creative intent; Natural language processing is performed on the text describing the creative intent to extract key themes, sentiment words, and desired style keywords; Using the aforementioned key themes, sentiment terms, and style keywords as seeds, multiple steps are taken in the multi-level semantic association network; During the exploration, prioritize exploring along the edges of derivative and subordinate relationships to discover materials with consistent style or shared origins; At the same time, based on the weight of the co-occurrence relationship edge, material nodes that have historically frequently co-occurred with seed features are assigned a high access probability; Collect the material nodes visited during the walk, their visit counts, and semantic similarity to the seed features, as an initial candidate set; For the materials in the initial candidate set, a comprehensive recommendation score is calculated by taking into account their access frequency, semantic similarity, freshness and popularity. Sort all candidate materials in descending order according to the overall recommendation score, generate an ordered list of related materials, and output it.

8. The big data-based animation design resource library management system according to claim 7, characterized in that, Natural language processing is performed on the text describing the creative intent, including: The text is segmented and stop words are removed to obtain the original keyword sequence; The original keywords were expanded using a domain dictionary, and synonyms and related terms related to animation design were added to form an expanded keyword set. Sentiment analysis models are used to determine the overall sentiment polarity and intensity of the text, core words representing sentiment tendencies are extracted, and potential topic distributions are extracted from the text through topic models, selecting the most probable topics as key topics. The extended keyword set, core emotional vocabulary, and key themes are collectively used as a set of semantic features representing creative intent.

9. The big data-based animation design resource library management system according to claim 1, characterized in that, The step of dynamically updating the parameter weights of the improved clustering algorithm and the connection strength of the multi-level semantic association network based on user behavior data regarding recommended content includes: Record user interactions with the ordered list of recommended materials, including viewing details, downloading, saving, and referencing in actual projects; When a user downloads or uses a recommended material, the semantic association strength between the recommended material and the seed intent feature on which it was recommended is enhanced, and the weight of the corresponding edge is increased in the multi-level semantic association network. When a user views or downloads multiple materials from the same cluster consecutively, the internal consistency of the cluster is strengthened, and higher weights are assigned to the feature dimensions represented by the cluster in the next round of clustering algorithm execution. If a user frequently combines materials from different clusters, then in the multi-level semantic association network, co-occurrence relationship edges are established or strengthened between these two clusters or between materials representing the clusters. The initial weight allocation strategy and distance calculation formula in the improved clustering algorithm are fine-tuned periodically using accumulated user behavior data, and the edge weights in the semantic association network are recalibrated.

10. The big data-based animation design resource library management system according to claim 1, characterized in that, The system also includes: The version and evolution management module is used to build a version evolution tree of the material, recording the different versions of each material from the original version through each modification and derivative creation. In the multi-level semantic association network, the different versions of the material are treated as independent but closely related nodes, and they are connected by strong derivative relationship edges. When recommending materials, if a version of a material is recommended and adopted by a user, the system will also trace its version evolution tree and include other versions that are visually or semantically highly consistent with the adopted version in the scope of subsequent recommendations. The system continuously monitors changes in the usage popularity and correlation strength of materials, downgrades or archives material nodes that have not been used for a long time and have weak correlation, and quickly identifies and enhances the network of emerging hot materials and their correlation relationships.

Images