A generative 3D modeling method for cultural resources and related equipment

Abstract

The application discloses a generative 3D modeling method for cultural resources and related equipment, and relates to the technical field of cultural resources, and can accurately present the structural features and cultural attributes of the generated final three-dimensional model.

Description

Technical Field

[0001] This invention relates to the field of digital culture and three-dimensional computer graphics technology, and in particular to a generative 3D modeling method and related equipment for cultural resources. Background Technology

[0002] With the rapid development of digital culture and virtual reality technology, the digital protection and display of cultural resources has become an important means of cultural inheritance and dissemination. 3D modeling technology, as a crucial foundation for the digitization of cultural resources, has been widely applied in fields such as cultural relic protection, historical scene restoration, digital museums, and immersive displays. However, traditional 3D modeling methods mainly rely on manual modeling or rule-based modeling processes, which not only heavily depend on professionals but also involve long modeling cycles and high costs, making it difficult to meet the practical needs of rapid digitization of large-scale and diverse cultural resources.

[0003] To address the aforementioned issues, various automated or semi-automated 3D modeling methods have been developed in existing technologies. These methods are mostly based on multi-view reconstruction, point cloud reconstruction, or image measurement techniques. They collect multiple sets of images or point cloud data of cultural resources, perform calculations and fusion to reconstruct the 3D geometric form of the cultural resources. In recent years, with the development of generative artificial intelligence technology, generative 3D modeling methods have emerged. These methods, by learning the data distribution characteristics of large amounts of 3D data, can automatically generate 3D models with reasonable geometric structures and surface textures without human intervention, thus reducing the dependence on manual operation and the modeling cost to a certain extent.

[0004] However, existing 3D modeling methods still have many shortcomings. Methods based on multi-view reconstruction, point cloud reconstruction, or image measurement have high requirements for acquisition conditions and data quality. When cultural resources have complex structures, degraded textures, missing data, or limited shooting conditions, the modeling effect is easily and significantly affected. Moreover, these methods usually focus on the restoration of geometric forms and are insufficient in modeling the historical semantics, stylistic features, and cultural attributes contained in cultural resources, making it difficult to effectively express the cultural connotations. Existing generative 3D models are mostly general-purpose models, lacking targeted modeling of the professional characteristics of cultural resources. They still have shortcomings in terms of cultural style consistency, detail realism, and generation controllability. The generated 3D models often fail to accurately reflect the historical value and cultural connotations carried by cultural resources. Summary of the Invention

[0005] This invention aims to address the shortcomings of existing technologies, specifically providing a generative 3D modeling method and related equipment for cultural resources, as detailed below: 1) In a first aspect, the present invention provides a generative 3D modeling method for cultural resources, the specific technical solution of which is as follows: Acquire multimodal data of cultural resources; The multimodal data is preprocessed to obtain structured multimodal data; Structural features, texture features, and semantic information of cultural resources are extracted from structured multimodal data. The structural features, texture features, and semantic information are jointly encoded to generate cultural attribute feature vectors that constrain the 3D generation process. Input the cultural attribute feature vectors into the generative 3D model to generate the initial 3D model; The geometry and surface texture of the initial 3D model are iteratively optimized to obtain the optimized 3D model; The optimized 3D model is determined as the final 3D model.

[0006] The beneficial effects of the generative 3D modeling method for cultural resources provided by this invention are as follows: By acquiring and preprocessing multimodal data of cultural resources to generate structured multimodal data, a unified data foundation is provided for subsequent feature extraction, effectively reducing reliance on high-precision acquisition equipment and professional manual modeling. Structural features, texture features, and semantic information of cultural resources are extracted from the structured multimodal data and jointly encoded to generate cultural attribute feature vectors that constrain the 3D generation process. This ensures that the generative 3D model is constrained by the structural, textural, and semantic attributes of the cultural resources during 3D generation, guaranteeing that the generated initial 3D model conforms to the original characteristics of the cultural resources in terms of geometric shape and texture. Iterative optimization of the geometric structure and surface texture of the initial 3D model continuously improves its geometric accuracy and texture realism, effectively ensuring the integrity and detail of the final 3D model. This method reduces manual intervention while improving the automation and efficiency of 3D modeling, enabling the generated final 3D model to accurately represent the structural features and cultural attributes of the cultural resources.

[0007] 2) Secondly, the present invention also provides a generative 3D modeling system for cultural resources, the specific technical solution of which is as follows: It includes a data acquisition module, a data preprocessing module, an extraction and generation module, a model generation module, an optimization module, and a determination module; The data acquisition module is used to acquire multimodal data of cultural resources; The data preprocessing module is used to preprocess multimodal data to obtain structured multimodal data; The extraction and generation module is used to: extract the structural morphological features, texture features, and semantic information of cultural resources from structured multimodal data, jointly encode the structural morphological features, texture features, and semantic information, and generate cultural attribute feature vectors to constrain the 3D generation process; The model generation module is used to: input cultural attribute feature vectors into the generative 3D model to generate an initial 3D model; The optimization module is used to iteratively optimize the geometry and surface texture of the initial 3D model to obtain the optimized 3D model. The determination module is used to determine the optimized 3D model as the final 3D model.

[0008] 3) In a third aspect, the present invention also provides an electronic device, the electronic device including a processor coupled to a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor, so as to enable the electronic device to implement any of the above-mentioned generative 3D modeling methods for cultural resources.

[0009] 4) In a fourth aspect, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-mentioned generative 3D modeling methods for cultural resources.

[0010] It should be noted that the beneficial effects of the technical solutions of the second to fourth aspects of the present invention and their corresponding possible implementations can be found in the above description of the technical effects of the first aspect and its corresponding possible implementations, and will not be repeated here. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments of the present invention will be briefly introduced below: Figure 1 This is a flowchart illustrating a generative 3D modeling method for cultural resources according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a generative 3D modeling system for cultural resources according to an embodiment of the present invention. Detailed Implementation

[0012] The principles and features of the present invention are described below. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0013] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0014] like Figure 1 As shown in the figure, a generative 3D modeling method for cultural resources according to an embodiment of the present invention includes the following steps: S1. Acquire multimodal data of cultural resources; multimodal data specifically includes image data, video data, scanned data, and text description data. The acquisition process is as follows: 1) Image data is two-dimensional visual information obtained by photographing cultural resources using a high-resolution digital camera or professional photography equipment. During the acquisition process, it is necessary to reasonably set the camera's aperture, shutter speed, ISO, and white balance according to the material, surface reflectivity, and ambient lighting conditions of the cultural resources to ensure that the acquired images clearly present the overall shape, local details, color composition, and surface texture of the cultural resources. For large cultural resources such as historical buildings, it is necessary to shoot from multiple angles and different heights to ensure coverage of all visible surfaces of the resource; for small cultural resources such as cultural relics and artifacts, it is necessary to shoot in a professional photography studio using uniform lighting and a standard color chart to ensure accurate color reproduction. After acquisition, the image data is stored in its original format, retaining complete color and detail information for subsequent preprocessing.

[0015] 2) Video data is a time-series image sequence obtained by continuously recording cultural resources using cameras or devices with video recording capabilities. Video data can capture the state changes of cultural resources over time, and has unique advantages in recording the dynamic display process of cultural resources, their appearance under changing ambient lighting, and their morphological characteristics when observed continuously from different perspectives. During acquisition, it is necessary to shoot around or directionally at the cultural resources according to a preset shooting path and angle to ensure that the video sequence covers visual information from all directions of the resources. The acquisition frame rate is generally set to 24 frames per second or higher to ensure smooth motion and sufficient detail capture. The acquired video data is stored in a common video format, containing a continuous sequence of image frames and their corresponding temporal information.

[0016] 3) Scanned data is a collection of spatial geometric information obtained through non-contact active measurement of cultural resources using 3D scanning equipment. The 3D scanning equipment emits laser beams or structured light and receives signals reflected from the surface of the cultural resource, calculating the 3D spatial coordinates of each point on the resource surface relative to the scanning equipment, thus generating point cloud data composed of a large number of spatial points. Each point in the point cloud data contains its X-axis, Y-axis, and Z-axis coordinates in 3D space. Some scanning equipment can also record the corresponding color value and surface reflection intensity value. During the acquisition process, it is necessary to select an appropriate type of 3D scanning equipment and set the corresponding scanning resolution and scanning range according to the size, shape complexity, and accuracy requirements of the cultural resource. For cultural resources with complex structures, multiple scans from multiple different locations are usually required. Then, the point cloud data obtained from each scan are registered and fused to obtain scan data that completely covers all surfaces of the cultural resource. The acquired scanned data is stored as a point cloud file for use in subsequent preprocessing steps.

[0017] 4) Textual description data refers to textual descriptive information related to cultural resources collected through methods such as literature review, archival retrieval, field survey records, and expert knowledge compilation. This type of data comes from a wide range of sources, including artifact archives, historical documents, archaeological excavation reports, museum collection descriptions, academic research papers, and records of the orderly transmission of cultural resources. During the collection process, it is necessary to extract textual descriptions related to the artifact type, chronological attributes, geographical origin, historical background, cultural connotations, and craftsmanship techniques of the cultural resources from the aforementioned various types of documents. For paper documents and ancient books, optical character recognition technology can be used to convert the text in scanned images into editable text data; for electronic documents and databases, the text content can be directly extracted and organized. The collected textual description data is stored in a structured or semi-structured text format, preserving the original semantic information and contextual relationships.

[0018] S2. Preprocess the multimodal data to obtain structured multimodal data, including: Standardized image data is obtained by unifying resolution and converting color space for image and video data in multimodal data. Standardized point cloud data is obtained by point cloud filtering and outlier removal for scanned data in multimodal data. Standardized text data is obtained by word segmentation and entity recognition for text description data, extracting cultural resource attribute descriptive words. The standardized image data, standardized point cloud data, and standardized text data are then mapped to the same spatial coordinate system and stored in association according to the spatial relationship of cultural resources, generating structured multimodal data. The specific implementation process is as follows: S20. Resolution unification processing is performed on image and video data separately. For image data, bicubic interpolation is used to adjust the resolution of all images to uniform width and height values. For example, images from different sources and of different sizes are all adjusted to a fixed resolution of 1280×1024. For video data, keyframe images are first extracted from the video sequence at fixed time intervals. Then, the same bicubic interpolation method is used to adjust the width and height values ​​of each extracted frame to uniform values, ensuring that each frame in the video has the same spatial resolution. After resolution unification, each frame in the image and video data is converted from its original color space to a unified color space, for example, from the red-green-blue color space to the luminance-chrominance color space. During the conversion, a color space transformation matrix is ​​used to linearly transform the three channel values ​​of each pixel to obtain standardized image data. Standardized image data refers to image data with consistent spatial dimensions and a unified color representation after resolution unification and color space conversion processing.

[0019] S21. Point cloud filtering is performed on the point cloud data in the scanned data. Statistical filtering is used: for each spatial point in the point cloud data, the nearest neighboring points are searched, the average distance between the spatial point and all neighboring points is calculated, and the global mean and global standard deviation of the average distances of all spatial points are calculated. When the average distance of a spatial point exceeds the sum of the global mean and three times the global standard deviation, the spatial point is identified as an outlier and removed from the point cloud data. After point cloud filtering, a radius filtering method is used for supplementary processing: for each spatial point in the point cloud data, the number of other spatial points contained within a spherical neighborhood centered on the spatial point and with a preset radius is counted. When the number is less than a preset threshold, the spatial point is identified as an isolated point and removed from the point cloud data. After the above processing, standardized point cloud data is obtained. Standardized point cloud data refers to a set of three-dimensional spatial coordinate points that have been filtered and removed from noisy and isolated points. Each point contains its X-axis, Y-axis, and Z-axis coordinates in three-dimensional space.

[0020] S22. The text description data is segmented using a method combining dictionary and statistical methods. The continuous Chinese character sequence in the text description data is divided into independent word sequences according to semantic units. After segmentation, entity recognition is performed on the word sequences. Entity recognition uses a sequence labeling model to label each word in the sequence with a corresponding entity category label. Entity category labels include artifact type, era, region, material type, craftsmanship, and cultural symbol name. Based on the entity category labels generated by entity recognition, words related to cultural resource attributes are extracted from the word sequence as cultural resource attribute descriptors. Cultural resource attribute descriptors refer to words that reflect specific attribute information of cultural resources in terms of artifact type classification, era, regional style, material composition, craftsmanship, and cultural symbols. All extracted cultural resource attribute descriptors are organized into a word list according to their original order of appearance; this word list constitutes the standardized text data.

[0021] S23. Establish a common spatial coordinate system. A common spatial coordinate system refers to a three-dimensional Cartesian spatial coordinate system with a common origin, a common X-axis direction, a common Y-axis direction, and a common Z-axis direction. For standardized point cloud data, since the scanned data itself contains spatial coordinate information, the coordinates of each spatial point it contains are directly transformed from the original scanned coordinate system to the common spatial coordinate system. A rigid body transformation matrix is ​​used for coordinate transformation during this transformation. For standardized image data, based on the camera position and shooting angle parameters recorded when acquiring each image or video frame, the inverse perspective projection transformation method is used to map each pixel in the image to a three-dimensional spatial point in the same spatial coordinate system, obtaining the spatial position and orientation information of each image in the same spatial coordinate system. For each cultural resource attribute descriptor in standardized text data, based on the spatial coordinate range of the specific cultural resource components described by the cultural resource attribute descriptor in the same spatial coordinate system, a corresponding spatial coordinate interval label is assigned to the cultural resource attribute descriptor. After completing the above mapping, the standardized image data, standardized point cloud data, and standardized text data are linked and stored according to the spatial location relationships of cultural resources: for each spatial point in the standardized point cloud data, the coordinate value of the spatial point and the image texture pixel value of the surface area associated with the spatial point are stored; for each image in the standardized image data, the position and orientation parameters of the image in the same spatial coordinate system and the index mapping relationship between the image and the corresponding spatial point in the standardized point cloud data are stored; for each cultural resource attribute descriptor in the standardized text data, the cultural resource attribute descriptor and its corresponding spatial coordinate interval label are stored. Through the above-mentioned linked storage, structured multimodal data is generated. Structured multimodal data refers to a multi-source data set formed by organizing and linking image data, point cloud data, and text data according to their spatial location relationships in the same spatial coordinate system.

[0022] S3. Extract the structural morphological features, texture features, and semantic information of cultural resources from structured multimodal data, and jointly encode the structural morphological features, texture features, and semantic information to generate cultural attribute feature vectors that constrain the 3D generation process. This includes extracting structural morphological features, texture features, and semantic information of cultural resources from structured multimodal data, including: S30. Perform three-dimensional spatial coordinate analysis on image data and scan data in structured multimodal data to extract the geometric topological relationships and surface curvature distribution of cultural resources as structural morphological features. Specifically: Image data has been mapped to the same spatial coordinate system, with each image having a known camera position and shooting orientation parameters. Utilizing multi-view stereo vision principles, feature point matching is performed on multiple images covering the same cultural resource area, calculating the 3D spatial coordinates of each feature point in the global spatial coordinate system to generate a sparse 3D point cloud. Subsequently, dense matching and triangulation are performed on the sparse 3D point cloud to generate a dense 3D mesh covering the surface of the cultural resource. For scanned data in structured multimodal data, the scanned data itself exists in the form of 3D spatial coordinate points, and the initial triangular mesh is directly constructed using these 3D spatial coordinate points. The dense 3D mesh generated from the image data and the initial triangular mesh constructed from the scanned data are fused and registered to form a unified 3D mesh surface. On the 3D mesh surface, the dihedral angles between adjacent faces are calculated. By analyzing the distribution of dihedral angles, the edges and planar regions of the mesh surface are identified, and the vertex connection relationships and face adjacency relationships are extracted as geometric topological relationships. Geometric topological relationships refer to the connection structure and adjacency hierarchy between vertices, edges, and faces on the 3D mesh surface, reflecting the basic framework and morphological composition of the cultural resource. On the surface of a 3D mesh, for each vertex, a local quadratic surface is fitted using the spatial coordinates of the vertex and its adjacent vertices. The Gaussian curvature and mean curvature of the vertex on the local surface are calculated. The curvature values ​​of all vertices are organized according to their spatial distribution on the mesh surface to obtain the surface curvature distribution. The surface curvature distribution refers to the spatial distribution map of the quantified values ​​of the degree of curvature at various points on the 3D surface of the cultural resource, reflecting the surface's unevenness and morphological undulations. The extracted geometric topological relationships and the surface curvature distribution are used together as structural morphological features.

[0023] S31. Perform color distribution statistics and surface texture recognition on image and video data in structured multimodal data to extract color composition, decorative patterns, and surface material information of cultural resources as texture features. Specifically: Color distribution statistics were performed on image and video data from structured multimodal data. Each frame of the image and video data was standardized and represented in a unified color space. For each image, the numerical distribution of all pixels in each channel of the color space was statistically analyzed, and a color histogram was plotted. The color values ​​corresponding to the peak values ​​in the color histogram were extracted as the set of dominant hues, and the area ratio of each dominant hue was calculated to form the color composition. Color composition refers to the main color types contained on the surface of cultural resources and their respective proportions. Surface texture recognition was performed on each frame of the image and video data. The texture features of the image, including contrast, correlation, and energy value, were calculated using a gray-level co-occurrence matrix method. These texture features were used to distinguish smooth regions, striped regions, and speckled regions of the image. In the identified striped and speckled regions, an edge detection algorithm was further used to extract the region boundaries. The connected regions enclosed by the boundaries were matched with a pre-set pattern template library. When a match was successful, the corresponding pattern type and spatial position were recorded as the pattern. Decorative patterns refer to recurring decorative graphic units on the surface of cultural resources, including geometric patterns, animal patterns, plant patterns, and human figures. Based on surface roughness parameters, gloss parameters, and local color variation gradients calculated during color distribution statistics and surface texture identification, the material type of each area on the surface of the cultural resource is comprehensively analyzed to determine its composition, including pottery, porcelain, stone, wood, and metal. The material type and its spatial distribution are then used as surface material information. The extracted color composition, decorative patterns, and surface material information are collectively used as texture features.

[0024] S32. Jointly analyze the identifiable cultural symbols in the text description data and image data of the structured multimodal data, and extract the artifact type classification, chronological attributes, regional style labels, and cultural attribute information related to historical background of the cultural resources as semantic information. Specifically: The text description data is standardized text data, which includes a list of cultural resource attribute descriptors. Words related to vessel form are directly extracted from the list, such as ding, bottle, jar, zun, gui, etc., and a preset vessel form classification knowledge graph is searched based on these words to determine the vessel form category to which the cultural resource belongs. The vessel form classification refers to categories divided according to the three-dimensional shape and functional use of cultural resources, such as food vessels, wine vessels, water vessels, musical instruments, and ritual vessels. Words related to time and era, including dynasty names and AD chronology ranges, are extracted from the list, and these words are mapped to specific time intervals as the era attribute. The era attribute refers to the quantified time range of the historical period in which the cultural resource is located. Words related to geographical location and ethnicity are extracted from the list, and these words are matched with the labels in the preset regional style label library to obtain regional style labels. The regional style label refers to a marker that characterizes the geographical region where the cultural resource originated or to which it belongs and its ethnic characteristics. For image data, a target detection network is used to detect and locate identifiable cultural symbols in the image, and the identifiable cultural symbols include inscriptions, clan emblem marks, religious symbols, and special shape identifiers. The category information and spatial position information of the detected identifiable cultural symbols are jointly analyzed with the cultural resource attribute descriptors in the text description data, in a specific manner: taking the category of the identifiable cultural symbol as a query key, searching the text description data for context statements containing the category, and extracting description content related to the historical background from the context statements. The description content related to the historical background includes the historical evolution of manufacturing techniques, unearthed information, circulation experiences, and association records with historical events or historical figures. The vessel form classification, era attribute, regional style labels, and the above-mentioned cultural attribute information associated with the historical background are integrated as semantic information. The cultural attribute information refers to the inherent attributes of cultural resources in the social and historical dimension in addition to basic forms and visual features, including cultural functions, symbolic meanings, and historical values.

[0025] S33. Mapping the structural morphological features, texture features, and semantic information to the same feature space for vector splicing to generate a cultural attribute feature vector, specifically: A common feature space refers to a real-valued vector space with a fixed number of dimensions, which is pre-set to a fixed value. The geometric topological relationships and surface curvature distribution in structural morphology features are transformed into fixed-length numerical vectors: for geometric topological relationships, the number of vertices, edges, and faces of the mesh, as well as the connectivity statistics of each vertex, are extracted to form a topological description vector; for surface curvature distribution, the curvature values ​​of all vertices are uniformly sampled to obtain a fixed-length curvature vector. The topological description vector and the curvature vector are concatenated end-to-end to form the structural morphology feature vector. The color composition in texture features is transformed into a color histogram vector, the pattern is transformed into a pattern type encoding vector, and the surface material information is transformed into a material distribution vector. These three vectors are concatenated end-to-end to form the texture feature vector. The object type classification, age attribute, and regional style label in semantic information are transformed into one-hot encoded vectors or embedding vectors, and the cultural attribute information associated with historical background is transformed into a keyword set vector. These vectors are concatenated end-to-end to form the semantic feature vector. The structural morphology feature vector, texture feature vector, and semantic feature vector each have their own number of dimensions and reside in different feature spaces. The structural morphological feature vector, texture feature vector, and semantic feature vector are each standardized to unify their numerical ranges to a preset interval. Then, the three vectors are concatenated sequentially, end to end, to form a new vector. The total dimension of the new vector equals the sum of the dimensions of the three vectors; this operation is called vector concatenation. The vector obtained after vector concatenation is the cultural attribute feature vector. The cultural attribute feature vector refers to a unified vector representation formed by encoding and fusing the structural morphological, texture, and semantic information of cultural resources, used to constrain the subsequent 3D generation process.

[0026] Specifically, structural morphological features, texture features, and semantic information are jointly encoded to generate cultural attribute feature vectors that constrain the 3D generation process, including: S34. Input the structural morphological features, texture features, and semantic information into three parallel feature encoding sub-networks. Each feature encoding sub-network performs normalization and dimension mapping on the input features to obtain structural morphological encoding vectors, texture encoding vectors, and semantic encoding vectors. Specifically: S340. Establish three parallel feature encoding sub-networks. These three sub-networks have the same network structure; each is a fully connected neural network consisting of an input layer, multiple hidden layers, and an output layer. However, the network parameters of the three sub-networks are independent and not shared. Structural features are input into the first feature encoding sub-network, texture features into the second feature encoding sub-network, and semantic information into the third feature encoding sub-network. These three input operations are performed simultaneously without waiting for each other.

[0027] S341. Each feature encoding sub-network performs normalization processing on the input features. Normalization processing refers to standardizing the value of each dimension of the input features according to the mean and standard deviation of that dimension across all samples, so that the mean of each dimension after transformation is zero and the standard deviation is one. Specifically, for each dimension of the input feature vector, the mean and standard deviation of that dimension are pre-calculated across all training samples. During the inference phase, the original value of the current input feature vector in that dimension is subtracted from the mean and then divided by the standard deviation to obtain the normalized value of that dimension. The above operation is performed sequentially on all dimensions of the input feature vector to complete the normalization processing of the entire feature vector. Normalization processing eliminates the influence of differences in units and numerical ranges between different feature dimensions.

[0028] S342. Each feature encoding sub-network performs dimension mapping on the normalized features. Dimension mapping refers to converting the original dimensions of the input features into a preset unified target dimension. The normalized feature vector is input to the input layer of the feature encoding sub-network. The input layer receives all dimensions of the feature vector and passes it through multiple hidden layers of nonlinear transformation. Each hidden layer linearly weights and sums the vectors output from the previous layer before performing a nonlinear mapping through an activation function, passing this process layer by layer. The output layer of the feature encoding sub-network maps the vector output from the last hidden layer to the preset unified target dimension, outputting a fixed-length vector. The fixed length is equal to the preset unified target dimension value. The three feature encoding sub-networks output three fixed-length vectors respectively. The vector output by the first feature encoding sub-network is used as the structural morphology encoding vector, the vector output by the second feature encoding sub-network is used as the texture encoding vector, and the vector output by the third feature encoding sub-network is used as the semantic encoding vector. The structural morphology encoding vector is a fixed-length vector representation obtained by normalizing and mapping structural morphological features through a feature encoding subnetwork. The texture encoding vector is a fixed-length vector representation obtained by normalizing and mapping texture features through a feature encoding subnetwork. The semantic encoding vector is a fixed-length vector representation obtained by normalizing and mapping semantic information through a feature encoding subnetwork. All three types of encoding vectors—structural morphology, texture, and semantic—have the same number of dimensions, which is equal to a pre-defined, unified target dimension.

[0029] S35. The structural morphology encoding vector, texture encoding vector, and semantic encoding vector are weighted and fused according to preset weight coefficients to obtain the cultural attribute feature vector, specifically: Three sets of weighting coefficients are pre-defined, each a real number between zero and one, and the sum of the three sets equals one. The weighted structural morphology encoding vector is obtained by multiplying each dimension value of the structural morphology encoding vector by its corresponding weighting coefficient. Similarly, the weighted texture encoding vector is obtained by multiplying each dimension value of the texture encoding vector by its corresponding weighting coefficient. The semantic encoding vector is then multiplied by its corresponding weighting coefficient. Element-wise addition is performed on the weighted structural morphology, texture, and semantic encoding vectors, adding the values ​​at the same dimensional positions to obtain a new vector with the same dimensions as the three vectors. This process is called weighted fusion. The resulting vector is used as the cultural attribute feature vector. This cultural attribute feature vector is a unified vector representation formed by fusing structural morphology, texture, and semantic information according to pre-defined importance weights, and is used in the subsequent constraint 3D generation process.

[0030] S4. Input the cultural attribute feature vector into the generative 3D model to generate the initial 3D model. The specific implementation process is as follows: S40. Prepare the training dataset by collecting a large number of 3D model samples of cultural resources. These samples include cultural object models from different cultural resource categories, eras, and regional styles. Each 3D model sample is accompanied by corresponding structural morphological feature annotations, texture feature annotations, and semantic information annotations. Combine each 3D model sample and its corresponding structural morphological feature annotations, texture feature annotations, and semantic information annotations into training sample pairs. Divide all training sample pairs into training and validation sets.

[0031] S41. Construct the basic network structure for a generative 3D model. The generative 3D model consists of four parts: an encoder network, a latent space layer, a generator network, and a discriminator network. The encoder network receives 3D model samples as input and encodes the geometric mesh data and texture image data of the 3D model into latent space encoded vectors. The latent space layer, located between the encoder network and the generator network, is responsible for storing and managing the latent space encoded vectors. The generator network receives the latent space encoded vectors and cultural attribute feature vectors as joint input and outputs the reconstructed 3D model's geometry and surface texture. The discriminator network receives the 3D model output from the generator network or real 3D model samples from the training set as input and outputs a discrimination result, which represents the probability value of whether the input 3D model is a real sample or a generated sample.

[0032] S42. In the first stage of training, a variational autoencoder is used. 3D model samples from the training set are input into the encoder network, which outputs a latent space encoding vector. The generator network receives the latent space encoding vector and outputs the reconstructed 3D model. The geometrical difference loss and texture difference loss between the reconstructed 3D model and the input 3D model samples are calculated, along with the KL divergence loss between the latent space encoding vector and the standard normal distribution. The weighted sum of the geometrical difference loss, texture difference loss, and KL divergence loss is used as the total loss for the first stage, and the parameters of the encoder and generator networks are updated using the backpropagation algorithm. This process is repeated until the total loss for the first stage converges, completing the first stage of training. After the first stage of training, the encoder network can encode any input 3D model into a latent space encoding vector, and the generator network can reconstruct the 3D model from the latent space encoding vector.

[0033] S43. In the second stage of training, a generative adversarial network (GAN) training method is adopted. The parameters of the encoder network trained in the first stage are kept unchanged. Latent space encoding vectors are randomly sampled from the latent space layer, and cultural attribute feature vectors are randomly sampled from the training set. The randomly sampled latent space encoding vectors and the randomly sampled cultural attribute feature vectors are concatenated and input into the generator network, which outputs a 3D model. The 3D model output by the generator network is used as a negative sample, and real 3D model samples from the training set are used as positive samples, both input into the discriminator network. The discriminator network outputs a discrimination result for each input sample. The adversarial loss of the generator network and the discriminator network are calculated based on the output of the discriminator network. The parameters of the generator network and the discriminator network are alternately updated using the backpropagation algorithm. When updating the generator network parameters, the parameters of the discriminator network are kept unchanged; when updating the discriminator network parameters, the parameters of the generator network are kept unchanged. This process is repeated until the adversarial losses of the generator network and the discriminator network reach Nash equilibrium, completing the second stage of training. After the second stage of training, the generator network can receive joint inputs of latent space encoding vectors and cultural attribute feature vectors to generate 3D models with reasonable geometric structures and realistic textures.

[0034] S44. During the inference phase, a latent space encoding vector is sampled from the latent space layer. This latent space encoding vector is concatenated with the cultural attribute feature vector along the vector dimension to form a joint input vector. This joint input vector is then input into the generator network of the generative 3D model. After one forward propagation computation, the generator network outputs the geometric mesh data and surface texture image data of the 3D model. The geometric mesh data includes the coordinates of all vertices, vertex connectivity, and face definitions of the 3D model. The surface texture image data includes color texture maps covering the surface of the 3D model. The geometric mesh data and surface texture image data output by the generator network are combined to form the 3D model, which is the initial 3D model.

[0035] S5. Iteratively optimize the geometry and surface texture of the initial 3D model to obtain the optimized 3D model, including: S50. Calculate the geometric difference between the initial 3D model's geometric structure and the target geometric structure constrained by the cultural attribute feature vector, and adjust the mesh vertex coordinates and surface normal directions of the initial 3D model according to the geometric difference to update the initial 3D model's geometric structure. The specific implementation process is as follows: S500. Decouple the geometric constraint component from the cultural attribute feature vector. The cultural attribute feature vector is a unified vector obtained by weighted fusion of the structural morphology encoding vector, texture encoding vector, and semantic encoding vector. The information related to geometric structure is mainly encoded in the structural morphology encoding vector. Using a preset dimension index range, extract the vector segment corresponding to the structural morphology constraint from the cultural attribute feature vector as the target geometric structure constraint vector.

[0036] S501. Decode the target geometric constraint vector into target geometric parameters. Pre-train a geometric decoder, which consists of a multi-layer fully connected network, capable of mapping the geometric constraint vector into target vertex coordinate offsets and target surface normal direction offsets in the 3D mesh. Input the target geometric constraint vector into the geometric decoder. The geometric decoder outputs target vertex coordinate offset vectors with the same number of vertices as the initial 3D model mesh, and corresponding target surface normal direction offset vectors. Each component of the target vertex coordinate offset vector corresponds to the 3D coordinate offset of a mesh vertex in the initial 3D model, and each component of the target surface normal direction offset vector corresponds to the angular offset of the surface normal direction at a mesh vertex in the initial 3D model.

[0037] S502. Calculate the geometric difference. Geometric difference refers to the degree of deviation between the current geometry of the initial 3D model and the target geometry. Add the current 3D coordinates of each mesh vertex of the initial 3D model to the offset of the target vertex coordinates corresponding to that mesh vertex to obtain the target coordinate value. Calculate the Euclidean distance between the current coordinate value and the target coordinate value. Sum the squares of the Euclidean distances of all mesh vertices to obtain the vertex coordinate difference value. Simultaneously, calculate the cosine of the angle between the current surface normal direction at each mesh vertex of the initial 3D model and the offset of the target surface normal direction corresponding to that mesh vertex. Average and invert the cosine values ​​of all vertices to obtain the normal direction difference value. Sum the vertex coordinate difference value and the normal direction difference value according to a preset weighted ratio to obtain the geometric difference value.

[0038] S503. Adjust the mesh vertex coordinates and surface normal directions of the initial 3D model based on geometric differences. Calculate the gradient of the geometric difference value with respect to the coordinates of each mesh vertex. The gradient indicates the direction in which the current coordinates can be moved to reduce the geometric difference. Update the 3D coordinates of each mesh vertex along the gradient descent direction with a preset learning rate to obtain the updated mesh vertex coordinates. Calculate the gradient of the geometric difference value with respect to the surface normal direction at each mesh vertex, and update the angle value of the surface normal direction along the gradient descent direction. Perform the above coordinate and normal direction update operations on all mesh vertices to complete the update of the initial 3D model's geometry.

[0039] S51. Calculate the texture difference between the surface texture of the updated 3D model and the target texture feature constrained by the cultural attribute feature vector, and adjust the surface texture coordinates and texture pixel values ​​of the updated 3D model according to the texture difference to update the surface texture of the 3D model. The specific implementation process is as follows: S510. Decouple the texture constraint component from the cultural attribute feature vector, and extract the vector segment corresponding to the texture encoding vector from the cultural attribute feature vector through the preset dimension index range, as the target texture feature constraint vector.

[0040] S511. Decode the target texture feature constraint vector into target texture parameters. Pre-train a texture decoder, which consists of a multi-layer fully connected network, capable of mapping the texture constraint vector into target texture coordinate offsets and target texture pixel value adjustments on the 3D model surface. Input the target texture feature constraint vector into the texture decoder. The texture decoder outputs the same number of target texture coordinate offsets as the updated 3D model surface texture coordinates, and the same number of target texture pixel value adjustments as the updated 3D model texture image pixels.

[0041] S512. Texture difference refers to the degree of deviation between the current surface texture and the target texture feature of the 3D model after updating the geometry. The target texture coordinates are obtained by adding the offset of each surface texture coordinate of the updated 3D model to the corresponding target texture coordinate. The Euclidean distance between the current and target texture coordinates is calculated, and the sum of the squares of the Euclidean distances for all texture coordinates is obtained to get the texture coordinate difference value. Simultaneously, the target pixel value is obtained by adding the current pixel value of each pixel in the texture image of the updated 3D model to the corresponding target texture pixel value adjustment. The absolute value of the difference between the current and target pixel values ​​is calculated, and the sum of the squares of the absolute values ​​of the differences for all pixels is obtained to get the texture pixel difference value. The texture coordinate difference value and the texture pixel difference value are then weighted and summed according to a preset ratio to obtain the texture difference numerical value.

[0042] S513. Adjust the surface texture coordinates and texture pixel values ​​of the updated 3D model based on the texture differences. Calculate the gradient of the texture difference value with respect to each surface texture coordinate, and update the texture coordinate values ​​along the gradient descent direction. Calculate the gradient of the texture difference value with respect to each pixel value in the texture image, and update the pixel values ​​along the gradient descent direction. Perform the above update operation on all texture coordinates and all texture pixels to complete the update of the 3D model surface texture.

[0043] S52. Repeat the geometric structure adjustment and surface texture adjustment until the sum of geometric differences and texture differences is less than a preset threshold, then stop the iteration and use the current 3D model as the optimized 3D model. Specifically: After each geometric structure and surface texture adjustment, the current geometric difference value and the current texture difference value are calculated and added together to obtain the total difference. This total difference is compared to a preset threshold, a positive real number pre-set based on the accuracy requirements of cultural resource modeling. When the total difference is less than the preset threshold, the iteration process stops, and the current 3D model, after multiple geometric structure and surface texture adjustments, is output as the optimized 3D model. When the total difference is greater than or equal to the preset threshold, the next round of geometric structure and surface texture adjustments continues until the stopping condition is met.

[0044] In this context, vertex coordinates refer to the X, Y, and Z axis coordinates of each vertex in the 3D mesh model, used to define the geometry of the 3D model. Surface normal direction refers to the direction vector perpendicular to the tangent plane at a point on the 3D model surface, used to describe the surface's orientation and smoothness. Texture difference refers to the quantized deviation between the current surface texture of the 3D model and the target texture features constrained by the cultural attribute feature vector, including a comprehensive measure of surface texture coordinate deviation and texture pixel value deviation. Surface texture coordinates refer to the position coordinates of each point on the 3D model surface mapped onto a 2D texture image, used to fit the texture image onto the 3D model surface. Texture pixel values ​​refer to the values ​​of each pixel in the texture image across each channel of the color space, used to define the color and pattern of the 3D model surface.

[0045] S6. Determine the optimized 3D model as the final 3D model.

[0046] Optionally, in the above technical solution, before determining the optimized 3D model as the final 3D model, the following steps are also included: S060. Conduct integrity assessment, structural rationality assessment, and cultural consistency assessment on the optimized 3D model to obtain the assessment results, including: S0600. The optimized 3D model is input into a pre-trained evaluation network. The evaluation network includes at least three branches: integrity evaluation, structural rationality evaluation, and cultural consistency evaluation. Then: ① The integrity evaluation branch calculates the mesh closure, hole ratio, and patch continuity of the optimized 3D model, and generates an integrity evaluation score based on these parameters; ② The structural rationality evaluation branch calculates the structural deviation between the geometric structure of the optimized 3D model and the target geometric structure constrained by the cultural attribute feature vector, and generates a structural rationality evaluation score based on the structural deviation; ③ The cultural consistency evaluation branch calculates the semantic deviation between the texture style features and decorative morphological features of the optimized 3D model and the target cultural attributes constrained by the cultural attribute feature vector, and generates a cultural consistency evaluation score based on the semantic deviation. The specific implementation process is as follows: The S06000 evaluation network employs a multi-task learning deep neural network architecture, sharing a bottom-level feature extractor and dividing the upper layers into three parallel branch networks. The shared bottom-level feature extractor consists of multiple graph convolutional layers used to process the mesh structure data of the 3D model. The input to the graph convolutional layers is the mesh vertex coordinates of the 3D model, the connecting edges between vertices, and the texture feature vectors on the vertices. Each graph convolutional layer extracts spatial structure features at different levels of abstraction by weighted aggregation of the features of each vertex and its neighboring vertices. After multiple layers of graph convolution, a vector representing the global features of the entire 3D model is obtained. This global feature vector is copied three times and input into the integrity evaluation branch, the structural rationality evaluation branch, and the cultural consistency evaluation branch, respectively. Each branch consists of two fully connected layers. The first fully connected layer maps the global feature vector to intermediate features, and the second fully connected layer maps the intermediate features to a scalar value, which serves as the evaluation score for that branch. The parameters of the fully connected layers in the three branches are independent and not shared.

[0047] S06001. Collect a large number of 3D model samples, each corresponding to a 3D model and its cultural attribute feature vector. For each sample, three true scores need to be assigned: integrity true score, structural rationality true score, and cultural consistency true score. The integrity true score is obtained by professionals based on the mesh closure, hole situation, and surface continuity of the 3D model. The structural rationality true score is obtained by quantitatively comparing the geometric structure of the 3D model with the target geometric structure constrained by the corresponding cultural attribute feature vector. The comparison method includes calculating vertex coordinate deviation and normal direction deviation, and normalizing the deviation values ​​as the score. The cultural consistency true score is obtained by cultural resource experts based on whether the texture style and decorative form of the 3D model conform to the cultural characteristics corresponding to the same cultural attribute feature vector. All samples are divided into training set and validation set.

[0048] S06002. Input the 3D model from the training set into the evaluation network. A shared low-level feature extractor extracts global features. The three branches output predicted integrity, structural rationality, and cultural consistency scores, respectively. Calculate the mean squared error between the predicted scores and the corresponding true scores to obtain the loss values ​​for each branch. Sum the loss values ​​of the three branches according to preset weighting coefficients to obtain the total loss. Calculate the gradient of the total loss with respect to all trainable parameters of the evaluation network using the backpropagation algorithm, and update the network parameters using gradient descent. Repeat the above forward propagation, loss calculation, and parameter update process on the training set until the total loss no longer decreases on the validation set, at which point training stops, resulting in the pre-trained evaluation network.

[0049] S06003. The integrity assessment branch calculates the mesh closure, hole ratio, and patch continuity of the optimized 3D model. After inputting the optimized 3D model into the pre-trained evaluation network, the integrity assessment branch processes the model. Mesh closure is the ratio of the number of all boundary edges on the surface of the 3D model to the total number of edges. A boundary edge is an edge possessed by only one patch. Mesh closure is obtained by counting the number of boundary edges in the mesh and dividing by the total number of edges. The closer the closure value is to zero, the more closed the model is. Hole ratio is the ratio of the area of ​​hole regions on the surface of the 3D model to the total surface area. The integrity assessment branch obtains the hole ratio by detecting areas in the mesh that are missing patches and calculating their areas. Patch continuity refers to the smoothness of the change in normal direction between adjacent patches. It is obtained by calculating the average of the angles between the normals of all adjacent patches and normalizing the average value to the range of zero to one. The larger the value, the more continuous the patch connection. The integrity assessment branch generates an integrity assessment score based on mesh closure, hole ratio, and patch continuity. Specifically, the complement of mesh closure, the complement of hole ratio, and patch continuity are multiplied by their respective weight coefficients and then summed to obtain the integrity assessment score. The integrity assessment score is a value between zero and one, with a higher value indicating better model integrity.

[0050] S06004. The structural rationality assessment branch calculates the structural deviation between the optimized 3D model's geometric structure and the target geometric structure constrained by the cultural attribute feature vector. The structural rationality assessment branch extracts geometric constraint components from the cultural attribute feature vector and obtains the target geometric structure parameters through an internal decoding network. It compares the current geometric structure of the optimized 3D model with the target geometric structure vertex-by-vertex, calculating the coordinate deviation and normal direction deviation of each vertex. The deviation values ​​of all vertices are averaged to obtain the structural deviation value. The structural rationality assessment branch generates a structural rationality assessment score based on the structural deviation. Specifically, the structural deviation value is converted into a score between zero and one using a nonlinear mapping function; the smaller the structural deviation, the higher the structural rationality assessment score.

[0051] S06005. The cultural consistency assessment branch calculates the semantic deviation between the texture style features and decorative morphological features of the optimized 3D model and the target cultural attribute constrained by the cultural attribute feature vector. The cultural consistency assessment branch extracts texture style features from the texture image of the optimized 3D model, including color distribution statistics, texture roughness, contrast, and directionality. It detects decorative morphological features from the mesh surface of the optimized 3D model, including the shape, size, repetition period, and spatial arrangement of the decorative areas. It extracts the target cultural attribute from the cultural attribute feature vector, including vessel type classification, age attribute, regional style label, and historical background association information. The branch compares the extracted texture style features and decorative morphological features with the standard features corresponding to the target cultural attribute, calculating the cosine distance and Euclidean distance between the feature vectors. The weighted sum of these two distances is used as the semantic deviation value. The cultural consistency assessment branch generates a cultural consistency assessment score based on the semantic deviation; a smaller semantic deviation results in a higher cultural consistency assessment score.

[0052] S0601. The integrity assessment score, structural rationality assessment score, and cultural consistency assessment score are weighted and summed according to preset evaluation weights to obtain a comprehensive assessment score. This comprehensive assessment score is used as the evaluation result. Specifically: the preset evaluation weights refer to the weight values ​​assigned to the integrity assessment score, structural rationality assessment score, and cultural consistency assessment score respectively. All three weight values ​​are positive real numbers, and their sum equals one. The integrity assessment score is multiplied by its corresponding weight value to obtain a weighted integrity score. The structural rationality assessment score is multiplied by its corresponding weight value to obtain a weighted structural rationality score. The cultural consistency assessment score is multiplied by its corresponding weight value to obtain a weighted cultural consistency score. The weighted integrity score, weighted structural rationality score, and weighted cultural consistency score are added together to obtain the comprehensive assessment score. The comprehensive assessment score is a scalar value between zero and one, and is used as the evaluation result output.

[0053] S061. When the evaluation result does not meet the preset conditions, adjust the latent space encoding vector of the generative 3D model, and re-input the adjusted latent space encoding vector into the generative 3D model to generate the 3D model until the evaluation result meets the preset conditions. The specific implementation process is as follows: S0610. The pre-trained evaluation network outputs a comprehensive evaluation score to the optimized 3D model. The comprehensive evaluation score is a scalar value between zero and one. The preset condition refers to the comprehensive evaluation score being greater than or equal to a preset evaluation threshold, which is a real number between zero and one, such as 0.85. The comprehensive evaluation score is compared with the evaluation threshold. When the comprehensive evaluation score is greater than or equal to the evaluation threshold, the evaluation result is determined to meet the preset condition, and the current optimized 3D model is directly output as the final 3D model without further adjustment steps. When the comprehensive evaluation score is less than the evaluation threshold, the evaluation result is determined to not meet the preset condition, and the latent space encoding vector adjustment step continues.

[0054] S0611. Obtain the deviations between the integrity assessment score, structural rationality assessment score, and cultural consistency assessment score in the evaluation results and their respective preset thresholds. The preset thresholds for the integrity assessment score, structural rationality assessment score, and cultural consistency assessment score are all independent, and are real numbers between zero and one. Subtract the integrity threshold from the integrity assessment score to obtain the integrity deviation. Subtract the structural rationality threshold from the structural rationality assessment score to obtain the structural rationality deviation. Subtract the cultural consistency threshold from the cultural consistency assessment score to obtain the cultural consistency deviation. When an assessment score is greater than or equal to its corresponding threshold, the deviation is positive or zero; when an assessment score is less than its corresponding threshold, the deviation is negative, indicating room for improvement.

[0055] S0612. Based on the deviation corresponding to the integrity assessment score, determine the adjustment step size of the first dimension direction in the latent space encoding vector related to mesh density and topology. The latent space encoding vector is a fixed-dimensional real-valued vector. Each dimension of the latent space encoding vector is pre-divided into three dimensional direction groups according to its function. The first dimension direction group includes dimensions in the latent space encoding vector related to controlling the mesh subdivision degree, mesh vertex density, and topological connectivity of the 3D model. When the integrity deviation is negative, it indicates insufficient geometric integrity of the model, requiring positive adjustment in the first dimension direction. Positive adjustment increases the value of the corresponding dimension to increase mesh density and strengthen the closure of the topology. When the integrity deviation is positive or zero, it indicates that the model integrity meets the requirements, and the adjustment step size in the first dimension direction is set to zero. The specific value of the adjustment step size is obtained by multiplying the absolute value of the deviation by a preset step size coefficient, which is a pre-set positive real number.

[0056] S0613. Based on the deviation amount corresponding to the structural rationality assessment score, determine the adjustment step size of the second dimension direction in the latent space encoding vector, which is related to the geometric shape and proportion. The second dimension direction group includes dimensions in the latent space encoding vector related to controlling the overall shape proportion, relative size of each part, and spatial positional relationship of the 3D model. When the structural rationality deviation amount is negative, it indicates that there is a deviation between the model's geometric shape and the target geometric structure, requiring positive or negative adjustment in the second dimension direction. Positive adjustment is used to increase the shape proportion, and negative adjustment is used to decrease the shape proportion. The specific adjustment direction is determined by a pre-trained shape direction discriminator, which outputs a positive or negative sign based on the deviation direction between the current 3D model's geometric structure and the target geometric structure. The absolute value of the adjustment step size is obtained by multiplying the absolute value of the structural rationality deviation amount by a preset step size coefficient.

[0057] S0614. Based on the deviation amount corresponding to the cultural consistency assessment score, determine the adjustment step size of the third dimension direction related to texture style and decorative features in the latent space encoding vector. The third dimension direction group includes dimensions in the latent space encoding vector related to controlling the surface texture style, color distribution, pattern type, and decorative form of the 3D model. When the cultural consistency deviation amount is negative, it indicates that the texture and decorative features of the model do not match the cultural attributes, and positive or negative adjustments need to be made in the third dimension direction. Positive adjustments are used to enhance the typicality of the texture style, and negative adjustments are used to weaken excessive texture features. The specific adjustment direction is determined by a pre-trained style direction discriminator, which outputs a positive or negative sign based on the deviation direction of the current 3D model's texture features from the target cultural attributes. The absolute value of the adjustment step size is obtained by multiplying the absolute value of the cultural consistency deviation amount by a preset step size coefficient.

[0058] S0615. Update the latent space coding vector's coding values ​​in the first, second, and third dimensions according to the adjustment step sizes for the first, second, and third dimensions, respectively, to obtain the adjusted latent space coding vector. For each dimension in the first dimension group, add the adjustment step size of the first dimension to the current coding value of the dimension. For each dimension in the second dimension group, add the product of the adjustment step size of the second dimension and the direction sign to the current coding value of the dimension. For each dimension in the third dimension group, add the product of the adjustment step size of the third dimension and the direction sign to the current coding value of the dimension. After completing the above update operations for all dimensions of the latent space coding vector, the adjusted latent space coding vector is obtained.

[0059] S0616. The adjusted latent space encoding vector is re-inputted into the generative 3D model to generate a 3D model. The adjusted latent space encoding vector and the cultural attribute feature vector are concatenated along the vector dimension to form a joint input vector. This joint input vector is then input into the generator network of the generative 3D model, which outputs a new 3D model. After iterative optimization of the geometric structure, the new 3D model is again input into the pre-trained evaluation network for evaluation. The comprehensive evaluation score is recalculated and compared with the evaluation threshold. If the recalculated comprehensive evaluation score is still less than the evaluation threshold, the above latent space encoding vector adjustment and regeneration process is repeated until the comprehensive evaluation score is greater than or equal to the evaluation threshold. At this point, the iteration stops, and the corresponding 3D model is output as the final 3D model.

[0060] Among these, the preset conditions refer to the pre-defined evaluation standards, specifically, a comprehensive evaluation score greater than or equal to a preset evaluation threshold. A comprehensive evaluation score below this threshold indicates that the generated 3D model does not meet the quality requirements. The latent space encoding vector refers to the vector representation located in the latent space layer of the generative 3D model. Adjusting the vector and re-inputting it into the generator network can affect the geometry, mesh density, shape proportions, texture style, and decorative features of the generated 3D model. The integrity deviation is the difference between the integrity evaluation score and the integrity threshold, measuring the degree of deficiency in mesh closure, hole conditions, and facet continuity. The structural rationality deviation is the difference between the structural rationality evaluation score and the structural rationality threshold, measuring the degree of deviation between the 3D model's geometry and the target geometry. The cultural consistency deviation is the difference between the cultural consistency evaluation score and the cultural consistency threshold, measuring the degree of deviation between the 3D model's texture style and decorative features and its cultural attributes. The first dimension refers to the set of dimensions in the latent space encoding vector that are related to the mesh density, topology, and closure control of the 3D model. Adjusting the encoding values ​​in this direction can change the geometric integrity of the model. The second dimension refers to the set of dimensions in the latent space encoding vector that are related to the overall shape proportions, size of parts, and spatial relationships of the 3D model. Adjusting the encoding values ​​in this direction can change the morphological rationality of the model. The third dimension refers to the set of dimensions in the latent space encoding vector that are related to the texture style, color distribution, pattern type, and decorative form control of the 3D model. Adjusting the encoding values ​​in this direction can change the cultural style expression of the model. The adjustment step size refers to the numerical value moved along each dimension direction during each iteration of the latent space encoding vector update. The step size is determined by multiplying the absolute value of the deviation by a preset step size coefficient, thus determining the magnitude of the adjustment.

[0061] Among them, adjusting the latent space encoding vector of the generative 3D model includes: S0610. When the evaluation results do not meet the preset conditions, obtain the deviation between the integrity evaluation score, structural rationality evaluation score, and cultural consistency evaluation score in the evaluation results and their respective preset thresholds. The specific implementation process is as follows: After determining that the evaluation results do not meet the preset conditions, the pre-set thresholds for integrity, structural rationality, and cultural consistency are obtained. The integrity threshold is a pre-set minimum passing score used to individually evaluate the geometric integrity of the 3D model. The structural rationality threshold is a pre-set minimum passing score used to individually evaluate the consistency between the 3D model's geometric form and the target geometric structure. The cultural consistency threshold is a pre-set minimum passing score used to individually evaluate the degree of matching between the 3D model's texture style and decorative features and its cultural attributes. All three thresholds are real numbers between zero and one, and are set independently of each other.

[0062] The integrity deviation is obtained by subtracting the integrity threshold from the integrity evaluation score output by the pre-trained evaluation network. Similarly, the structural rationality deviation is obtained by subtracting the structural rationality threshold from the structural rationality evaluation score. Finally, the cultural consistency deviation is obtained by subtracting the cultural consistency threshold from the cultural consistency evaluation score. All three deviations—integrity deviation, structural rationality deviation, and cultural consistency deviation—are real numbers. When an evaluation score is greater than or equal to its corresponding threshold, the deviation is positive or zero, indicating that the requirement is met. When an evaluation score is less than its corresponding threshold, the deviation is negative, indicating a deficiency that needs improvement through further adjustments.

[0063] S0611. Based on the deviation corresponding to the integrity assessment score, determine the adjustment step size of the first dimension direction in the latent space encoding vector that is related to the grid density and topology, specifically: After training to obtain the dimensionality information of the latent space encoding vectors, each dimension has automatically formed a functional differentiation through training. Some dimensions control the mesh density and topology of the 3D model, some control the geometry and proportions of the 3D model, and some control the texture style and decorative features of the 3D model. Principal component analysis is used to pre-label the functions of each dimension of the latent space encoding vectors, and the dimensions controlling the number of mesh vertices, mesh subdivision levels, patch density distribution, and the degree of vertex connection closure are assigned to the first dimension direction group.

[0064] When the integrity deviation is negative, the adjustment step size for the first dimension is set to a positive real number. The absolute value of the integrity deviation is multiplied by a preset step size coefficient to obtain the adjustment step size for the first dimension. The preset step size coefficient is a pre-defined positive real number used to control the magnitude of each adjustment. When the integrity deviation is positive or zero, it indicates that the geometric integrity of the 3D model has reached the single-item threshold requirement, and the adjustment step size for the first dimension is set to zero, meaning no adjustment is made to the first dimension. Positive adjustments to the first dimension increase the value of the corresponding dimension in the latent space encoding vector, thereby increasing the mesh density of the generated 3D model, strengthening the closure of the topology, and repairing holes and broken edges.

[0065] S0612. Based on the deviation amount corresponding to the structural rationality assessment score, determine the adjustment step size of the second dimension direction in the latent space encoding vector that is related to geometry and proportion, specifically: The dimensions controlling the overall shape proportions, relative sizes of parts, spatial relationships, and length-width-height ratios of the 3D model are grouped into the second dimension direction group. When the structural rationality deviation is negative, the encoded values ​​in the second dimension direction need to be adjusted to improve the geometry of the 3D model. Determining the adjustment step size for the second dimension direction consists of two sub-steps: the first sub-step multiplies the absolute value of the structural rationality deviation by a preset step size coefficient to obtain the absolute value of the adjustment step size; the second sub-step determines the adjustment direction of the second dimension direction using a pre-trained shape direction discriminator. The shape direction discriminator is a binary classification neural network. Its input is the deviation direction feature between the current 3D model's geometry and the target geometry constrained by the cultural attribute feature vector. The output is either positive or negative. A positive sign indicates a positive adjustment along the second dimension direction, i.e., increasing the encoded value of the shape proportion-related dimension; a negative sign indicates a negative adjustment along the second dimension direction, i.e., decreasing the encoded value of the shape proportion-related dimension. The absolute value of the adjustment step size is multiplied by the direction sign output by the shape direction discriminator to obtain the final value of the second dimension direction adjustment step size.

[0066] S0613. Based on the deviation corresponding to the cultural consistency assessment score, determine the adjustment step size of the third dimension direction related to texture style and decorative features in the latent space encoding vector, specifically: The dimensions controlling the surface texture style, color distribution, pattern type, decorative form, and cultural symbol presentation of the 3D model are categorized into the third dimension direction group. When the cultural consistency deviation is negative, the encoded values ​​in the third dimension direction need to be adjusted to improve the cultural style representation of the 3D model. Determining the adjustment step size for the third dimension direction consists of two sub-steps: the first sub-step multiplies the absolute value of the cultural consistency deviation by a preset step size coefficient to obtain the absolute value of the adjustment step size; the second sub-step determines the adjustment direction of the third dimension direction using a pre-trained style direction discriminator. The style direction discriminator is a binary classification neural network. Its input is the deviation direction feature between the current 3D model's texture style features and decorative form features and the target cultural attribute constrained by the cultural attribute feature vector. The output is either positive or negative. A positive sign indicates a positive adjustment along the third dimension direction, i.e., enhancing the typicality of the texture style and the vividness of the cultural symbols; a negative sign indicates a negative adjustment along the third dimension direction, i.e., weakening texture features that are too strong or do not conform to the cultural attribute. The absolute value of the adjustment step size is multiplied by the direction sign output by the style direction discriminator to obtain the final value of the third dimension direction adjustment step size.

[0067] S0614. According to the adjustment step size in the first dimension, the second dimension, and the third dimension, update the encoding values ​​of the latent space encoding vector in the first, second, and third dimensions respectively to obtain the adjusted latent space encoding vector, specifically: The latent space encoding vector is a fixed-dimensional real-valued vector with a pre-defined total number of dimensions. All dimensions of the latent space encoding vector are grouped into three groups: a first-dimensional direction group, a second-dimensional direction group, and a third-dimensional direction group. These three groups do not overlap and together cover all dimensions of the latent space encoding vector.

[0068] For each dimension belonging to the first dimension direction group, add the current encoded value of the dimension to the adjustment step size of the first dimension direction. Since the adjustment step size of the first dimension direction is a non-negative real number, the encoded values ​​of each dimension in the first dimension direction group will only increase or remain unchanged after the update, and will not decrease.

[0069] For each dimension belonging to the second-dimensional direction group, the current encoded value of the dimension is added to the adjustment step size of the second-dimensional direction. Since the adjustment step size of the second-dimensional direction can be positive or negative, the encoded values ​​of each dimension in the second-dimensional direction group may increase or decrease after the update, and the specific direction is determined by the output of the shape direction discriminator.

[0070] For each dimension belonging to the third-dimensional direction group, the current encoded value of the dimension is added to the adjustment step size of the third-dimensional direction. Since the adjustment step size of the third-dimensional direction can be positive or negative, the encoded values ​​of each dimension in the third-dimensional direction group may increase or decrease after the update, and the specific direction is determined by the output of the style direction discriminator.

[0071] After performing the above update operations on all dimensions of the latent space encoding vector, the adjusted latent space encoding vector is obtained. The adjusted latent space encoding vector has the same number of dimensions as the original latent space encoding vector, with only the values ​​of some dimensions changing. The adjusted latent space encoding vector is then used to re-input the generative 3D model for 3D model generation.

[0072] The integrity deviation is an algebraic difference between the integrity assessment score and the integrity threshold, used to quantify the deficiencies of the 3D model in terms of mesh closure, hole conditions, and surface continuity; a negative value indicates the need for improvement. The structural rationality deviation is also an algebraic difference between the structural rationality assessment score and the structural rationality threshold, used to quantify the deviation between the 3D model's geometry and the target geometry; a negative value indicates the need for improvement. The cultural consistency deviation is an algebraic difference between the cultural consistency assessment score and the cultural consistency threshold, used to quantify the inadequacy of the match between the 3D model's texture style and decorative forms and cultural attributes; a negative value indicates the need for improvement. The shape orientation discriminator is a pre-trained binary classification deep neural network. Its input is the feature vector of the deviation direction between the current geometry of the 3D model and the target geometry, and its output is positive or negative, used to determine the adjustment direction of the second dimension of the latent space encoding vector. The style orientation discriminator is a pre-trained binary classification deep neural network. Its input is a feature vector representing the deviation direction between the current texture style features and decorative morphology features of the 3D model and the target cultural attributes. The output is either positive or negative, used to determine the adjustment direction of the third dimension of the latent space encoding vector. In S6, determining the optimized 3D model as the final 3D model includes: determining the 3D model when the evaluation results meet the preset conditions as the final 3D model.

[0073] In the above embodiments, although the steps are numbered S1, S2, etc., they are only specific embodiments given by the present invention. Those skilled in the art can adjust the execution order of S1, S2, etc. according to the actual situation. The scheme after adjusting the order is also within the protection scope of the present invention. It can be understood that in some embodiments, some or all of the above embodiments may be included.

[0074] like Figure 2As shown, an embodiment of the present invention provides a generative 3D modeling system for cultural resources, comprising a data acquisition module, a data preprocessing module, an extraction and generation module, a model generation module, an optimization module, and a determination module. The data acquisition module is used to acquire multimodal data of cultural resources; The data preprocessing module is used to preprocess multimodal data to obtain structured multimodal data; The extraction and generation module is used to: extract the structural morphological features, texture features, and semantic information of cultural resources from structured multimodal data, jointly encode the structural morphological features, texture features, and semantic information, and generate cultural attribute feature vectors to constrain the 3D generation process; The model generation module is used to: input cultural attribute feature vectors into the generative 3D model to generate an initial 3D model; The optimization module is used to iteratively optimize the geometry and surface texture of the initial 3D model to obtain the optimized 3D model. The determination module is used to determine the optimized 3D model as the final 3D model.

[0075] Optionally, the above technical solution also includes an evaluation module, which is used for: Before determining the optimized 3D model as the final 3D model, an integrity assessment, a structural rationality assessment, and a cultural consistency assessment are conducted on the optimized 3D model to obtain the assessment results. When the evaluation result does not meet the preset conditions, the latent space encoding vector of the generative 3D model is adjusted, and the adjusted latent space encoding vector is re-input into the generative 3D model to generate the 3D model until the evaluation result meets the preset conditions. The determination module is specifically used to determine the 3D model that meets the preset conditions when the evaluation results are used as the final 3D model.

[0076] Optionally, in the above technical solution, the extraction and generation module is specifically used for: Three-dimensional spatial coordinate analysis is performed on image data and scan data in structured multimodal data to extract the geometric topological relationships and surface curvature distribution of cultural resources as structural morphological features; Color distribution statistics and surface texture recognition are performed on image and video data in structured multimodal data to extract color composition, decorative patterns and surface material information of cultural resources as texture features; Joint analysis is performed on the text description data and image data in structured multimodal data to extract the artifact type classification, age attribute, regional style label and cultural attribute information associated with historical background of cultural resources as semantic information; Structural features, texture features, and semantic information are mapped to the same feature space and then concatenated into vectors to generate cultural attribute feature vectors.

[0077] Optionally, in the above technical solution, the extraction and generation module is specifically used for: Structural features, texture features, and semantic information are input into three parallel feature encoding sub-networks. Each feature encoding sub-network performs normalization and dimension mapping on the input features to obtain structural morphology encoding vectors, texture encoding vectors, and semantic encoding vectors. The structural morphology encoding vector, texture encoding vector, and semantic encoding vector are weighted and fused according to preset weight coefficients to obtain the cultural attribute feature vector.

[0078] Optionally, in the above technical solution, the optimization module is specifically used for: Calculate the geometric difference between the initial 3D model's geometry and the target geometry constrained by the cultural attribute feature vector, and adjust the mesh vertex coordinates and surface normal directions of the initial 3D model according to the geometric difference to update the geometry of the initial 3D model; The texture difference between the surface texture of the updated 3D model and the target texture feature constrained by the cultural attribute feature vector is calculated, and the surface texture coordinates and texture pixel values ​​of the updated 3D model are adjusted according to the texture difference to update the surface texture of the 3D model. Repeatedly perform geometric structure adjustment and surface texture adjustment until the sum of geometric differences and texture differences is less than a preset threshold, then stop iterating and use the current 3D model as the optimized 3D model.

[0079] Optionally, in the above technical solution, the evaluation module is specifically used for: The optimized 3D model is input into a pre-trained evaluation network, which includes at least three branches: integrity evaluation, structural rationality evaluation, and cultural consistency evaluation. The integrity assessment branch calculates the mesh closure, hole ratio, and surface continuity of the optimized 3D model, and generates an integrity assessment score based on the mesh closure, hole ratio, and surface continuity. The structural rationality assessment branch calculates the structural deviation between the optimized 3D model's geometric structure and the target geometric structure constrained by the cultural attribute feature vector, and generates a structural rationality assessment score based on the structural deviation. The semantic deviation between the texture style features and decorative form features of the optimized 3D model and the target cultural attributes constrained by the cultural attribute feature vector is calculated through the cultural consistency assessment branch, and a cultural consistency assessment score is generated based on the semantic deviation. The integrity assessment score, structural rationality assessment score, and cultural consistency assessment score are weighted and summed according to the preset evaluation weights to obtain the comprehensive assessment score, which is used as the assessment result.

[0080] Optionally, in the above technical solution, the evaluation module is further specifically used for: When the evaluation results do not meet the preset conditions, obtain the deviation between the integrity evaluation score, structural rationality evaluation score, and cultural consistency evaluation score in the evaluation results and their respective preset thresholds; Based on the deviation corresponding to the integrity assessment score, determine the adjustment step size of the first dimension direction in the latent space encoding vector that is related to the grid density and topology; Based on the deviation amount corresponding to the structural rationality assessment score, determine the adjustment step size of the second dimension direction in the latent space encoding vector that is related to geometry and proportion; Based on the deviation corresponding to the cultural consistency assessment score, determine the adjustment step size of the third dimension direction related to texture style and decorative features in the latent space encoding vector; The latent space coding vector is updated according to the adjustment step size in the first dimension, the second dimension, and the third dimension, respectively, to obtain the adjusted latent space coding vector.

[0081] Optionally, in the above technical solution, the data preprocessing module is specifically used for: Standardized image data is obtained by unifying resolution and converting color space for image and video data in multimodal data. Standardized point cloud data is obtained by filtering point cloud data and removing outliers for scan data in multimodal data. Standardized text data is obtained by segmenting words and recognizing entities for text description data and extracting cultural resource attribute descriptive words. Standardized image data, standardized point cloud data and standardized text data are mapped to the same spatial coordinate system and then stored in association according to the spatial location relationship of cultural resources to generate structured multimodal data.

[0082] The technical solution of the present invention will be described through another embodiment. The system of this embodiment includes a cultural resource data acquisition and preprocessing module, a cultural semantic and style feature modeling module, a generative 3D modeling module, and a model evaluation and optimization module. The cultural resource data acquisition and preprocessing module acquires image, video, scanned, and textual description data, and performs preprocessing operations such as resolution unification, color space conversion, point cloud filtering, outlier removal, word segmentation, entity recognition, and spatial coordinate mapping to generate structured multimodal data, providing a unified input for subsequent modeling. The cultural semantic and style feature modeling module jointly encodes the structural morphological features, texture features, and semantic information of cultural resources to generate cultural attribute feature vectors that constrain the 3D generation process. The generative 3D modeling module automatically generates an initial 3D model based on these cultural attribute feature vectors and iteratively optimizes the geometry and surface texture of the initial 3D model to obtain an optimized 3D model. The model evaluation and optimization module evaluates the optimized 3D model for integrity, structural rationality, and cultural consistency. It outputs a comprehensive evaluation score through a pre-trained evaluation network and adjusts the latent space encoding vector of the generative 3D model if the evaluation results do not meet preset conditions. The adjusted latent space encoding vector is then re-inputted into the generative 3D model for 3D model generation until the evaluation results meet the preset conditions.

[0083] In the modeling input stage, the system receives multimodal data of cultural resources and completes data preprocessing and feature extraction operations. Specifically, this includes standardizing image data by unifying resolution and converting color space for image and video data; standardizing point cloud data by filtering point cloud data and removing outliers from scanned data; and standardizing text data by segmenting word data and recognizing entities. The standardized image data, standardized point cloud data, and standardized text data are then mapped to the same spatial coordinate system and stored in association according to the spatial relationships of the cultural resources, generating structured multimodal data. Subsequently, the cultural semantics and style feature modeling module jointly encodes the structural morphological features, texture features, and semantic information of the cultural resources to generate cultural attribute feature vectors that constrain the 3D generation process. In the 3D generation stage, the generative 3D modeling module generates an initial 3D model based on the cultural attribute feature vectors and optimizes the model's geometry and surface texture through multiple iterations to obtain an optimized 3D model. In the output phase, the system inputs the optimized 3D model into a pre-trained evaluation network for integrity evaluation, structural rationality evaluation, and cultural consistency evaluation, and obtains a comprehensive evaluation score as the evaluation result. When the evaluation result meets the preset conditions, the optimized 3D model is determined as the final 3D model; when the evaluation result does not meet the preset conditions, the latent space encoding vector of the generative 3D model is adjusted and the 3D model is regenerated until the evaluation result meets the preset conditions, and the final 3D model is output.

[0084] To more clearly demonstrate the technical solution of this invention, the following specific embodiments illustrate the implementation of generative 3D modeling technology for cultural resources in practical applications. These embodiments, focusing on the needs of cultural resource digitization, combine generative 3D modeling methods with cultural semantic and stylistic feature modeling methods to achieve high-quality automated 3D modeling of cultural objects, applicable to scenarios such as cultural heritage protection and digital display.

[0085] Example 1: A generative 3D modeling platform for cultural heritage digitization. This platform achieves automated 3D reconstruction and display of cultural relics and historical buildings through structured multimodal data modeling and generative 3D modeling modules. The platform system architecture includes a cultural resource data acquisition and preprocessing module, a cultural semantic and style feature modeling module, a generative 3D modeling module, and a model evaluation and optimization module. The cultural resource data acquisition and preprocessing module collects image data, video data, scanned data, and text description data of cultural resources. It performs preprocessing operations on the collected data, such as resolution unification, color space conversion, point cloud filtering, outlier removal, word segmentation, entity recognition, and spatial coordinate mapping, generating structured multimodal data to provide a unified input for subsequent feature modeling and 3D generation. The cultural semantics and style feature modeling module jointly encodes the structural morphological features, texture features, and semantic information of cultural resources. Through three parallel feature encoding sub-networks, it normalizes and maps the structural morphological features, texture features, and semantic information to obtain structural morphological encoding vectors, texture encoding vectors, and semantic encoding vectors. These are then weighted and fused according to preset weight coefficients to generate a cultural attribute feature vector. This vector constrains the generation process of the generative 3D model, ensuring that the generated 3D model meets cultural attribute requirements. The generative 3D modeling module, based on the cultural attribute feature vector, inputs it into the generative 3D model to automatically generate an initial 3D model. By calculating geometric and texture differences, it iteratively adjusts the mesh vertex coordinates, surface normal directions, surface texture coordinates, and texture pixel values ​​of the initial 3D model, achieving high-quality 3D reconstruction of the cultural object and obtaining an optimized 3D model. The model evaluation and optimization module inputs the optimized 3D model into a pre-trained evaluation network. Through integrity evaluation branches, structural rationality evaluation branches, and cultural consistency evaluation branches, it calculates integrity evaluation scores, structural rationality evaluation scores, and cultural consistency evaluation scores, respectively. These scores are then weighted and summed according to preset evaluation weights to obtain a comprehensive evaluation score as the evaluation result. When the evaluation results do not meet the preset conditions, the latent space encoding vector of the generative 3D model is adjusted, and the adjusted latent space encoding vector is re-input into the generative 3D model to generate the 3D model until the evaluation results meet the preset conditions and the final 3D model is output.

[0086] The operation process of the platform is as follows: ① The system receives image data, video data, scanned data, and text description data of cultural resources, and performs preprocessing such as resolution unification, color space conversion, point cloud filtering, outlier removal, word segmentation, entity recognition, and spatial coordinate mapping to generate structured multimodal data. ② The cultural semantic and style feature modeling module extracts the structural morphological features, texture features, and semantic information of cultural resources from the structured multimodal data. These three features are input into three parallel feature encoding sub-networks for normalization and dimensionality mapping to obtain structural morphological encoding vectors, texture encoding vectors, and semantic encoding vectors. These are then weighted and fused according to preset weight coefficients to generate cultural attribute feature vectors. ③ The generative 3D modeling module inputs the cultural attribute feature vectors into the generative 3D model to generate an initial 3D model. By calculating geometric and texture differences, iteratively adjusting the mesh vertex coordinates, surface normal directions, surface texture coordinates, and texture pixel values ​​of the initial 3D model, the optimized 3D model is obtained. ④ The model evaluation and optimization module inputs the optimized 3D model into the pre-trained evaluation network. It calculates the integrity evaluation score, structural rationality evaluation score, and cultural consistency evaluation score through the integrity evaluation branch, structural rationality evaluation branch, and cultural consistency evaluation branch, respectively. The comprehensive evaluation score is obtained by weighting and summing according to the preset evaluation weights. When the evaluation result meets the preset conditions, the 3D model is determined as the final 3D model. When the evaluation result does not meet the preset conditions, the latent space encoding vector of the generative 3D model is adjusted and the 3D model is regenerated until the evaluation result meets the preset conditions, and the final 3D model is output.

[0087] Example 2: The generative 3D modeling technology of this invention is applied to a digital museum scenario to achieve efficient 3D modeling and online display of cultural exhibits. The system architecture includes a generative 3D modeling module and a 3D display and interaction module. The generative 3D modeling module, based on generative 3D modeling technology, inputs cultural attribute feature vectors into the generative 3D model to quickly create 3D models of museum exhibits, generating initial 3D models. Through iterative optimization of geometric and texture differences, an optimized 3D model is obtained, reducing manual modeling workload while maintaining the accuracy of the exhibit's appearance and cultural style. The 3D display and interaction module integrates the optimized 3D model into the digital display system, supporting multi-angle browsing and interactive operations, providing users with an immersive cultural experience.

[0088] As can be seen from the above embodiments, the technical solution of the present invention can be flexibly deployed in different cultural resource application scenarios. Through the synergistic effect of generative 3D modeling and cultural attribute feature vectors, high-quality and resource-efficient 3D cultural modeling is achieved. This technology improves the automation level of cultural resource digitization and provides a reliable implementation path for large-scale cultural resource protection and digital display.

[0089] This invention combines generative 3D modeling technology with cultural resource feature modeling methods, effectively improving the intelligence level and overall efficiency of digital modeling of cultural resources. By introducing cultural attribute feature vectors to constrain the generative 3D model, the system can accurately reflect the structural morphological features, texture features, and semantic information of cultural objects during the automatic generation of 3D models. This allows the generated 3D models to not only possess visual realism but also accurately present the historical value and cultural connotations carried by the cultural resources. Furthermore, this invention enhances the system's robustness in the face of missing or quality-limited data through structured multimodal data co-modeling.

[0090] This invention reduces the need for manual modeling and high-precision data acquisition while achieving reasonable control over computing resources. This allows the system to operate stably under resource-constrained conditions, meeting the practical needs of large-scale cultural resource digitization. Compared to traditional modeling methods, this invention effectively alleviates the contradiction between modeling efficiency and quality by jointly training the encoder network, latent space layer, generator network, and discriminator network of a generative 3D model, as well as iteratively optimizing geometric and texture differences. This reduces modeling costs and time, and improves the overall efficiency of digital preservation and display of cultural resources.

[0091] Furthermore, this invention possesses the capability to evaluate and optimize generated results. Through a pre-trained evaluation network, it assesses the integrity, structural rationality, and cultural consistency of the optimized 3D model, outputting a comprehensive evaluation score. Based on model quality feedback, it continuously adjusts the latent space encoding vectors of the generative 3D model and regenerates the 3D model, ensuring that different types of cultural resources can obtain highly compatible and structurally complete 3D models. This enhances the utilization value of cultural resources and optimizes the allocation of technology and resources in digital culture construction. The generative 3D modeling technology of this invention is applicable to scenarios such as digital preservation of cultural heritage, construction of digital museums, restoration of historical scenes, and immersive cultural displays. It can efficiently construct high-quality 3D cultural resource models, improving the efficiency of digital utilization of cultural resources.

[0092] It should be noted that the beneficial effects of the generative 3D modeling system for cultural resources provided in the above embodiments are the same as those of the generative 3D modeling method for cultural resources described above, and will not be repeated here. Furthermore, the system provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the system can be divided into different functional modules according to the actual situation to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the method embodiments, and will not be repeated here.

[0093] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-mentioned generative 3D modeling methods for cultural resources.

[0094] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-mentioned generative 3D modeling methods for cultural resources.

[0095] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0096] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A generative 3D modeling method for cultural resources, characterized in that, include: Acquire multimodal data of cultural resources; The multimodal data is preprocessed to obtain structured multimodal data; The structural morphological features, texture features, and semantic information of cultural resources are extracted from the structured multimodal data. The structural morphological features, texture features, and semantic information are jointly encoded to generate cultural attribute feature vectors that constrain the 3D generation process. The cultural attribute feature vectors are input into the generative 3D model to generate an initial 3D model. The geometry and surface texture of the initial 3D model are iteratively optimized to obtain the optimized 3D model; The optimized 3D model is determined as the final 3D model.

2. The generative 3D modeling method for cultural resources according to claim 1, characterized in that, Before determining the optimized 3D model as the final 3D model, the following steps are also included: The optimized 3D model is subjected to integrity assessment, structural rationality assessment, and cultural consistency assessment to obtain the assessment results; When the evaluation result does not meet the preset conditions, the latent space encoding vector of the generative 3D model is adjusted, and the adjusted latent space encoding vector is re-inputted into the generative 3D model to generate the 3D model until the evaluation result meets the preset conditions. Determining the optimized 3D model as the final 3D model includes: determining the 3D model whose evaluation results meet preset conditions as the final 3D model.

3. The generative 3D modeling method for cultural resources according to claim 1, characterized in that, Extracting structural morphological features, texture features, and semantic information of cultural resources from the structured multimodal data includes: Three-dimensional spatial coordinate analysis is performed on the image data and scan data in the structured multimodal data to extract the geometric topological relationship and surface curvature distribution of the cultural resources as the structural morphological features; Color distribution statistics and surface texture recognition are performed on the image data and video data in the structured multimodal data to extract the color composition, decorative patterns and surface material information of the cultural resources as the texture features; The recognizable cultural symbols in the text description data and image data of the structured multimodal data are jointly analyzed to extract the object type classification, age attribute, regional style label and cultural attribute information associated with historical background of the cultural resources as the semantic information; The structural morphological features, texture features, and semantic information are mapped to the same feature space and then concatenated into vectors to generate the cultural attribute feature vector.

4. The generative 3D modeling method for cultural resources according to claim 3, characterized in that, The structural morphological features, texture features, and semantic information are jointly encoded to generate a cultural attribute feature vector that constrains the 3D generation process, including: The structural morphological features, texture features, and semantic information are respectively input into three parallel feature encoding sub-networks. Each feature encoding sub-network performs normalization and dimension mapping on the input features to obtain structural morphological encoding vector, texture encoding vector, and semantic encoding vector. The structural morphology encoding vector, the texture encoding vector, and the semantic encoding vector are weighted and fused according to preset weight coefficients to obtain the cultural attribute feature vector.

5. A generative 3D modeling system for cultural resources, characterized in that, It includes a data acquisition module, a data preprocessing module, an extraction and generation module, a model generation module, an optimization module, and a determination module; The data acquisition module is used to: acquire multimodal data of cultural resources; The data preprocessing module is used to: preprocess the multimodal data to obtain structured multimodal data; The extraction and generation module is used to: extract structural morphological features, texture features, and semantic information of cultural resources from the structured multimodal data; jointly encode the structural morphological features, texture features, and semantic information to generate a cultural attribute feature vector that constrains the three-dimensional generation process; The model generation module is used to: input the cultural attribute feature vector into the generative 3D model to generate an initial 3D model; The optimization module is used to iteratively optimize the geometry and surface texture of the initial 3D model to obtain an optimized 3D model. The determining module is used to: determine the optimized 3D model as the final 3D model.

6. A generative 3D modeling system for cultural resources according to claim 5, characterized in that, It also includes an evaluation module, which is used for: Before determining the optimized 3D model as the final 3D model, an integrity assessment, a structural rationality assessment, and a cultural consistency assessment are performed on the optimized 3D model to obtain the assessment results. When the evaluation result does not meet the preset conditions, the latent space encoding vector of the generative 3D model is adjusted, and the adjusted latent space encoding vector is re-inputted into the generative 3D model to generate the 3D model until the evaluation result meets the preset conditions. The determining module is specifically used to: determine the three-dimensional model when the evaluation results meet the preset conditions as the final three-dimensional model.

7. A generative 3D modeling system for cultural resources according to claim 5, characterized in that, The extraction and generation module is specifically used for: Three-dimensional spatial coordinate analysis is performed on the image data and scan data in the structured multimodal data to extract the geometric topological relationship and surface curvature distribution of the cultural resources as the structural morphological features; Color distribution statistics and surface texture recognition are performed on the image data and video data in the structured multimodal data to extract the color composition, decorative patterns and surface material information of the cultural resources as the texture features; The recognizable cultural symbols in the text description data and image data of the structured multimodal data are jointly analyzed to extract the object type classification, age attribute, regional style label and cultural attribute information associated with historical background of the cultural resources as the semantic information; The structural morphological features, texture features, and semantic information are mapped to the same feature space and then concatenated into vectors to generate the cultural attribute feature vector.

8. A generative 3D modeling system for cultural resources according to claim 7, characterized in that, The extraction and generation module is specifically used for: The structural morphological features, texture features, and semantic information are respectively input into three parallel feature encoding sub-networks. Each feature encoding sub-network performs normalization and dimension mapping on the input features to obtain structural morphological encoding vector, texture encoding vector, and semantic encoding vector. The structural morphology encoding vector, the texture encoding vector, and the semantic encoding vector are weighted and fused according to preset weight coefficients to obtain the cultural attribute feature vector.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a generative 3D modeling method for cultural resources as described in any one of claims 1 to 4.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the generative 3D modeling method for cultural resources as described in any one of claims 1 to 4.

Images