Method for mining graphical modal design knowledge based on three-dimensional model of product parts

Abstract

The application discloses a kind of based on product parts three-dimensional model graphic modal design knowledge mining method, comprising the following steps: step one: three-dimensional graphics preprocessing: 11) by 3D format conversion tool, three-dimensional graphic data is converted into Obj format;12) using PCL point cloud library is uniformly sampled, obtains the three-dimensional point cloud set corresponding to each three-dimensional graphic;Step two: construct three-dimensional target detection module to extract the initial latent feature of candidate object;Step three: using the feature enhancement module of attribute and relationship perception to strengthen the initial latent feature expression;Step four: construct lightweight prediction network, generate the text description of three-dimensional object;Step five: combined with text modal knowledge graph generation algorithm generates graphic modal knowledge graph.The application based on product parts three-dimensional model graphic modal design knowledge mining method can effectively improve the accuracy of three-dimensional object recognition and associated mining, enhance the semantic analysis capability of three-dimensional graphic data.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial knowledge representation, and specifically provides a method for mining graphic modal design knowledge based on three-dimensional models of product parts. Background Art

[0002] With the rapid development of digital technology and computer vision, 3D models have been widely used in fields such as computer-aided design, medical diagnosis, bioinformatics, 3D printing, medical imaging, and digital entertainment. 3D models not only improve the accuracy of design and manufacturing but also provide more intuitive and rich visual information. The application of 3D graphic data has become particularly important throughout the entire life cycle of industrial products, from design to manufacturing to maintenance. They are not only the foundation of product design but also play a key role in product manufacturing and maintenance.

[0003] In the industrial sector, designers can achieve more intuitive product design by building 3D models. This is where Model-Based Definition (MBD) technology comes in. MBD aims to integrate information from 2D engineering drawings into 3D models, detailing the product's manufacturing methods and characteristics. However, in traditional MBD methods, the design basis is often not integrated into the 3D model. This basis includes both explicit knowledge in the design specifications and implicit knowledge held within the designer's mind. Integrating this knowledge into the 3D model helps designers better understand the product, improving design quality and efficiency.

[0004] With the increasing popularity of 3D model applications, understanding and processing 3D model retrieval on large-scale data has become a hot topic. In recent years, advances in deep learning have brought about numerous 3D model retrieval methods. For example, MVCNN generates compact 3D shape descriptors by extracting information from multiple 2D image combinations. PointNet and PointNet++ use dense occupancy grids to represent input data, recursively feeding them into a hierarchical neural network to obtain a 3D shape representation.

[0005] Although existing methods have achieved certain results in 3D model retrieval, they mainly focus on describing global structural information and ignore the local structural information and correlation of the model. This makes it impossible to fully express the intrinsic characteristics and detailed information of the model when processing complex 3D models of industrial products. In the industrial knowledge system of the entire product life cycle, knowledge mining and extraction of graphical modalities, combined with knowledge graph technology, can greatly improve the consistency of knowledge representation and the accuracy of mining. Knowledge graphs not only provide a guiding framework for label generation, but also ensure the unified expression and consistency of knowledge in various modalities. In addition, the introduction of the cross-modal learning framework enables knowledge under different modalities to be uniformly expressed and integrated in the text space, ultimately generating a multimodal industrial knowledge graph with high consistency, coherence and completeness. Through this method, complex industrial knowledge can be effectively parsed, the efficiency of knowledge sharing and utilization can be improved, and the development of intelligent manufacturing can be promoted.

[0006] The widespread application of intelligent manufacturing systems today involves a vast amount of graphical modal data throughout the lifecycle of industrial products, including 3D assemblies during the design phase and 3D scenes within manufacturing twin systems. This data typically consists of multiple graphical objects with complex interrelationships. Effectively parsing the semantics of this complex and heterogeneous 3D graphical data is a key challenge in intelligent manufacturing. Summary of the Invention

[0007] In view of this, the purpose of the present invention is to provide a graphic modal design knowledge mining method based on the three-dimensional model of product parts, which can effectively improve the accuracy of three-dimensional object recognition and association mining, and enhance the semantic parsing ability of three-dimensional graphic data.

[0008] In order to achieve the above object, the present invention provides the following technical solutions:

[0009] A method for mining graphical modal design knowledge based on three-dimensional models of product parts includes the following steps:

[0010] Step 1: 3D graphics preprocessing

[0011] 11) Convert the 3D graphics data into Obj format using a 3D format conversion tool;

[0012] 12) Use PCL point cloud library for uniform sampling to obtain the 3D point cloud set corresponding to each 3D figure Where: N is the number of coordinate points, K is the auxiliary feature dimension;

[0013] Step 2: Build a 3D object detection module to extract the initial potential features of the candidate objects;

[0014] Step 3: Use a feature enhancement module that combines attribute and relationship perception to enhance the original latent feature expression;

[0015] Step 4: Build a lightweight prediction network to generate a text description of the 3D object;

[0016] Step 5: Combine the text modal knowledge graph generation algorithm to generate a graphic modal knowledge graph.

[0017] Furthermore, in step 2, the method steps for extracting initial potential features of candidate objects using a three-dimensional object detection module are as follows:

[0018] 21) Construct an object detection module based on VoteNet to achieve the initial encoding of the potential features of candidate objects;

[0019] 22) Use PointNet++ to adaptively extract point cloud features and obtain seed point sets Each seed point consists of spatial coordinates and hidden layer features; where M is the number of seed points and C is the dimension of hidden layer features; any seed point s in S i Expressed as: s i =[x i ;f i ], where x i represents the spatial coordinates, f i Represents the hidden layer features;

[0020] 23) Using the MLP-based point cloud voting module, transform the seed point in all dimensions, learn the feature offset of the seed point to the candidate target center point, and output the voting point And each point v in the voting point V i Represented as v i =[y i ;g i ], and y i =x i +Δx i , g i =f i +Δf, where Δx i and Δf i Represent the predicted coordinates and hidden feature offsets respectively;

[0021] 24) Based on spatial location similarity and farthest point sampling, the voting points are clustered to obtain several sample clusters and the voting points in each sample cluster wi = [z i ;h i ], where: z i represents the coordinates of the cluster center point, h i Represents the cluster radius; coordinate normalization is performed within each sample cluster to fully utilize the local geometric information within the cluster;

[0022] 25) A proposal extraction module consisting of Max Pooling and multiple MLP layers is used to generate proposals for each sample cluster, and the initial potential features of the candidate graphic objects covering spatial boundary information are obtained.

[0023] Furthermore, in step 3, the method steps of using the feature enhancement module to enhance the expression of the original latent features are as follows:

[0024] 31) Use a fully connected layer to map the multi-view RGB features and local position features reflecting the attributes to the same dimension as the initial latent features of the candidate object, and add the two to embed the fine-grained object attribute features;

[0025] 32) Encode the pairwise Euclidean distances between the center points of different candidate objects and the component distances in the three coordinate axes of the Cartesian coordinate system, and construct a spatial adjacency matrix group to express the positional relationship between the candidate objects;

[0026] 33) Encode through the fully connected layer to obtain the relational embedding features Where: H is the number of attention heads of the feature enhancement module;

[0027] Furthermore, in step 4, the method steps for generating a text description of the three-dimensional object are as follows:

[0028] 41) Construct a 3D description text generation network based on a single-layer multi-head Transformer module to achieve information interaction between the enhanced latent features of the target candidate and the initial latent features of other candidate objects;

[0029] 42) Select the target candidate, perform linear transformation on its potential features, and generate the query matrix, key matrix and value matrix;

[0030] 43) Using the KNN strategy to screen out the initial potential features of the K candidate objects closest to the center of the target candidate object;

[0031] 44) A post-processing module consisting of a fully connected layer and a word prediction layer is used to iteratively generate each word of the description text to obtain a text description of the three-dimensional object.

[0032] Furthermore, in step 5, the steps of the method for generating the graphical modality knowledge graph are as follows:

[0033] 51) Obtain text descriptions of scenes and objects from graphic data using a three-dimensional description model;

[0034] 52) Use the text modality entity relationship extraction algorithm to deeply mine text knowledge, extract the text into triples, and generate a graphic modality knowledge graph that conforms to the ontology system framework.

[0035] The beneficial effects of the present invention are:

[0036] The present invention is based on a graphic modal design knowledge mining method for three-dimensional models of product parts. First, the three-dimensional graphic data is unified into the Obj format through a 3D format conversion tool, and the PCL point cloud library is used for point cloud processing to obtain a three-dimensional point cloud set; then, a target detection module is constructed to extract the initial potential features of the candidate objects, and a feature enhancement module combining attribute and relationship perception is used to strengthen the expression of these features; then, a lightweight prediction network is trained to generate a text description of the three-dimensional object; finally, a graphic modal knowledge graph is generated in combination with a text modal knowledge graph generation algorithm; in summary, the present invention is based on a graphic modal design knowledge mining method for three-dimensional models of product parts, which can effectively improve the accuracy of three-dimensional object recognition and association mining, enhance the semantic parsing ability of three-dimensional graphic data, and realize efficient parsing and knowledge expression of complex heterogeneous three-dimensional data; it not only contributes to knowledge management and sharing throughout the life cycle of industrial products, but also significantly improves the efficiency and quality of the design and manufacturing process.

[0037] The present invention also has the following advantages:

[0038] (1) Unified data format and point cloud processing. Preprocess the massive and heterogeneous 3D graphics data, unify its expression, and convert it into the standard Obj format to ensure data standardization and consistency. Subsequently, the 3D graphics are uniformly sampled through the PCL point cloud library to generate a 3D point cloud set. This not only captures the geometric structure of the graphics, but also reflects the core intrinsic properties of the 3D object through the encoding of auxiliary features (such as multi-view appearance information, global coordinate information, normal vector information, etc.), laying the foundation for subsequent object recognition and association mining.

[0039] (2) Efficient 3D target detection. This paper constructs a 3D target detection module based on VoteNet. PointNet++ is used to adaptively extract point cloud features, generate a seed point set, and learn feature offsets through the MLP point cloud voting module to output voting points. The voting points are clustered and coordinate normalized to generate sample clusters, and the initial potential features of the candidate graphic objects are then obtained through the Proposal extraction module. This process efficiently captures the target object features in 3D space, providing a solid foundation for subsequent feature enhancement.

[0040] (3) Feature enhancement of attribute and relationship perception. In the feature enhancement module, the present invention introduces attribute encoding and relationship encoding. The attribute encoding module maps attribute features such as color and material to the same dimension as the initial potential features of the candidate object through a fully connected layer, and performs feature fusion, thereby enhancing the fine-grained attribute expression of the object. The relationship encoding module constructs a spatial adjacency matrix group by encoding the spatial relationship between objects, and uses the Transformer encoder to fuse attribute and relationship features, thereby enhancing the expression of the association characteristics between objects. In this way, the potential features of the candidate object are expressed more accurately and richly.

[0041] (4) Training a lightweight prediction network. This paper designs a prediction network based on a lightweight Transformer architecture to generate text descriptions of 3D objects. The description text generation network generates a description text that comprehensively reflects the attributes and relationships of the 3D object by interacting with the features of the target object and other objects.

[0042] (5) Comprehensive knowledge representation of text modalities and three-dimensional modalities. When generating a graphic modal knowledge graph, the present invention utilizes a three-dimensional description model to achieve a comprehensive textual description of graphic data scenes and objects. Through a text modal entity relationship extraction algorithm, a graphic modal knowledge graph that conforms to the ontology system framework is generated.

[0043] In summary, this method, by fully mining and expressing 3D object attribute information and associated information, not only improves the ability to parse complex 3D graphics data, but also achieves greater accuracy and comprehensiveness in knowledge representation. This approach has important applications in intelligent manufacturing, 3D design, and industrial product lifecycle management, greatly facilitating subsequent knowledge graph generation and mining tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to make the purpose, technical solutions and beneficial effects of the present invention more clear, the present invention provides the following drawings for illustration:

[0045] Figure 1 This is a flow chart of the method for mining knowledge of graphical modal design based on three-dimensional models of product parts according to the present invention;

[0046] Figure 2 It is a flow chart of the three-dimensional graphics preprocessing method;

[0047] Figure 3 Schematic diagram for combining attribute- and relation-aware feature embedding and enhancement;

[0048] Figure 4 A schematic diagram for generating graphical modal knowledge graphs;

[0049] Figure 5 Flowchart for designing and building the training framework;

[0050] Figure 6 Schematic diagram of the knowledge mining method. DETAILED DESCRIPTION

[0051] The present invention will be further described below with reference to the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments are not intended to limit the present invention.

[0052] 1. Graphical modal design knowledge mining method based on 3D models of product components

[0053] like Figure 1 As shown, the present embodiment provides a method for mining knowledge of graphical modal design based on three-dimensional models of product parts, including the following steps.

[0054] Step 1: 3D graphics preprocessing

[0055] like Figure 2 As shown, in this embodiment, the method steps for preprocessing three-dimensional graphics are:

[0056] 11) Convert the 3D graphics data into Obj format using the 3D format conversion tool.

[0057] 12) Use PCL point cloud library for uniform sampling to obtain the 3D point cloud set corresponding to each 3D figure Where N is the number of coordinate points, and K is the auxiliary feature dimension. Specifically, auxiliary features are encoded based on the multi-view appearance information, global coordinate information, and normal vector information of the coordinate points. They can reflect the core intrinsic properties of 3D graphic objects and facilitate subsequent tasks such as object recognition and association mining.

[0058] Step 2: Build a 3D object detection module to extract the initial potential features of candidate objects.

[0059] In this embodiment, the method steps for extracting initial potential features of candidate objects using the 3D object detection module are as follows:

[0060] 21) Construct an object detection module based on VoteNet to achieve the initial encoding of the potential features of candidate objects.

[0061] 22) Use PointNet++ to adaptively extract point cloud features and obtain seed point sets Each seed point consists of spatial coordinates and hidden layer features; where M is the number of seed points and C is the hidden layer feature dimension; any seed point si in S is represented by: s i =[x i ;f i ], where x i represents the spatial coordinates, fi represents the hidden layer feature.

[0062] 23) Adopting the point cloud voting module based on MLP, transforming in all dimensions of the seed point, learning the feature offset of the seed point to the candidate target center point, and outputting the voting point and each point v in the voting point V i is represented as v i = [y i ; g i ], and y i = x i + Δx i , g i = f i + Δf, wherein Δx i and Δf i respectively represent the predicted coordinates and the hidden layer feature offset.

[0063] 24) Clustering the voting points based on spatial position similarity and farthest point sampling to obtain a plurality of sample clusters and voting points w i = [z i ; h i ] in each sample cluster, wherein: z i represents the cluster center point coordinates, and h i represents the cluster radius; coordinate normalization is performed inside each sample cluster to fully utilize the local geometric information inside the cluster.

[0064] 25) Adopting the proposal extraction module composed of Max Pooling and a plurality of MLP layers to generate the proposals of each sample cluster, and solving to obtain the initial latent features of the candidate graphic objects covering the spatial boundary information.

[0065] Step three: using the feature enhancement module combining attribute and relationship perception to strengthen the initial latent feature expression.

[0066] Specifically, in the embodiment, the main part of the feature enhancement module adopts the Transformer encoder architecture, wherein the attribute encoding module and the relationship encoding module are fused, which is used to realize the enhanced expression of the candidate object latent features, and output more accurate spatial positioning results and feature representations integrating internal attribute information and multi-object association information.

[0067] In the embodiment, the method steps of using the feature enhancement module to strengthen the initial latent feature expression are as follows:

[0068] 31) Using the fully connected layer to map the multi-view RGB features reflecting the attributes and the local position features to the same dimension as the initial latent features of the candidate objects, and embedding the fine-grained object attribute features after adding the two.

[0069] 32) Encode the pairwise Euclidean distances between the center points of different candidate objects and the component distances in the three coordinate axes of the Cartesian coordinate system, and construct the spatial adjacency matrix group CD x , D y , D z , Dist) to express the positional relationship between candidate objects. x , D y and D z They are the component distances in the three coordinate axis directions of the Cartesian coordinate system; Dist is the pairwise Euclidean distance between the center points of the candidate objects.

[0070] 33) Encode through the fully connected layer to obtain the relational embedding features Where: H is the number of attention heads of the feature enhancement module.

[0071] Step 4: Build a lightweight prediction network to generate text descriptions of three-dimensional objects.

[0072] like Figure 3 As shown, in this embodiment, the method steps for generating a text description of a three-dimensional object are:

[0073] 41) A 3D description text generation network is constructed based on a single-layer multi-head Transformer module to achieve information interaction between the enhanced latent features of the target candidate object and the initial latent features of other candidate objects.

[0074] 42) The target candidate is selected through a specific strategy, and its potential features are linearly transformed to generate the Query matrix, Key and Value matrix.

[0075] 43) The initial potential features of the K candidate objects closest to the center of the target candidate object are screened using the KNN strategy.

[0076] 44) A post-processing module consisting of a fully connected layer and a word prediction layer is used to iteratively generate each word of the description text to obtain a text description of the three-dimensional object.

[0077] Step 5: Combine the text modal knowledge graph generation algorithm to generate a graphic modal knowledge graph.

[0078] like Figure 4 As shown, in this embodiment, the steps of the method for generating the graphical modality knowledge graph are as follows:

[0079] 51) With the help of the three-dimensional description model, obtain the text description of the scene and objects of the graphic data.

[0080] 52) Use the text modality entity relationship extraction algorithm to deeply mine text knowledge, extract the text into triples, and generate a graphic modality knowledge graph that conforms to the ontology system framework.

[0081] 2. Knowledge graph generation technology based on three-dimensional description

[0082] In this embodiment, graphical modal knowledge graph generation involves automatically extracting semantically related triples from massive amounts of graphical data using computer algorithms. These triples can effectively describe key information such as a product's geometry, functional characteristics, and assembly relationships. For example, during the product design phase, triple generation technology can automatically extract key structural and functional components from design drawings, enabling structured storage and intelligent retrieval of design information.

[0083] 3D representation technology uses 3D models to visually display a product's geometry and internal structure. This technology can be used not only for visual analysis during product design and manufacturing, but also for fault diagnosis and performance optimization during product use and maintenance. For example, during the manufacturing phase, 3D representation technology can visually demonstrate the product's assembly process and the relationships between components, thereby improving assembly efficiency and precision.

[0084] By analyzing and processing graphic modal data, it is possible to identify and describe various graphic objects throughout the lifecycle of industrial products, providing fundamental support for intelligent manufacturing. Graphic modal knowledge graph generation technology can automatically extract triplet knowledge, including entities, attributes, and relationships, by analyzing and mining graphic data, thereby enabling semantic description and understanding of graphic objects. Three-dimensional description technology, on the other hand, is dedicated to providing more detailed and accurate descriptions of graphic objects, including geometric shapes, material properties, functional characteristics, and more, providing richer information support for intelligent manufacturing processes.

[0085] 3. Training framework design and construction

[0086] The training framework aims to semantically parse the complex, heterogeneous, and diverse 3D graphics data found in industrial scenarios. It uses 3D description training to generate textual descriptions of 3D objects. The framework includes five main modules: preprocessing, 3D object detection, feature enhancement, description text generation, and knowledge graph generation.

[0087] The loss function of the entire training framework is derived from the descriptive text loss. Thanks to the framework's careful attention to 3D object attribute and association information, the resulting descriptive text comprehensively reflects the entity and relationship information within the 3D graphics data. The relationship between the text and the 3D objects is clearly and reliably expressed, greatly facilitating subsequent knowledge graph generation tasks. Furthermore, the large language model can be used to fuse the descriptive information of individual objects within the same 3D graphic to generate a global description of the entire graphic, further enriching the generated text features.

[0088] The design and construction of the training framework aims to solve the semantic parsing problem of complex, heterogeneous and diverse 3D graphic data in industrial scenarios, and to generate text descriptions of 3D objects through 3D description training methods. The specific process of training framework design and construction is as follows: Figure 5 As shown in the figure, the process of designing and constructing a training framework includes the following key steps. First, clarify the needs and goals, and determine the functional and performance requirements of the training framework. Next, design the system architecture and select appropriate technology stacks and tools, such as deep learning frameworks (such as TensorFlow, PyTorch) and hardware resources (such as GPU, TPU). Then, modularly design each component, including data processing modules, model building modules, training management modules, and evaluation modules. Next, perform code implementation and integration testing to ensure that each module works together. Finally, perform performance optimization and iterative improvements, and adjust parameters and algorithms based on actual training results to achieve the best results.

[0089] 4. Knowledge Mining Methods

[0090] Knowledge mining refers to obtaining entities, new entity links, new association rules and other information from given data, which is of great significance for the automatic construction of large-scale knowledge graphs. Knowledge mining system is a comprehensive tool or platform used to extract valuable information and knowledge from large amounts of data to support decision-making and action. It integrates multiple functions such as data preprocessing, data mining, model evaluation and selection, knowledge expression and knowledge application. Knowledge mining methods and systems such as Figure 6 shown.

[0091] The above embodiments are merely preferred embodiments for the purpose of fully illustrating the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are within the scope of protection of the present invention. The scope of protection of the present invention shall be subject to the claims.

Claims

1. A method for mining graphical modal design knowledge based on three-dimensional models of product parts, characterized by: The steps include: Step 1: 3D graphics preprocessing 11) Convert 3D graphics data into Obj format using 3D format conversion tools; 12) Use PCL point cloud library for uniform sampling to obtain the 3D point cloud set corresponding to each 3D figure ,in: is the number of coordinate points, is the auxiliary feature dimension; Step 2: Build a 3D object detection module to extract the initial potential features of the candidate objects; Step 3: Use a feature enhancement module that combines attribute and relationship perception to enhance the original latent feature expression; Step 4: Build a lightweight prediction network to generate a text description of the 3D object; Step 5: Generate a graphic modal knowledge graph by combining the text modal knowledge graph generation algorithm; In step 2, the method steps for extracting the initial potential features of the candidate object using the three-dimensional object detection module are as follows: 21) Construct an object detection module based on VoteNet to achieve initial encoding of the potential features of candidate objects; 22) Use PointNet++ to adaptively extract point cloud features and obtain seed point sets , each seed point consists of spatial coordinates and hidden layer features; among them, is the number of seed points, is the hidden layer feature dimension; Any seed point in Expressed as: ,in, represents the spatial coordinates, Represents the hidden layer features; 23) Using the MLP-based point cloud voting module, transform the seed point in all dimensions, learn the feature offset of the seed point to the candidate target center point, and output the voting point ; and polling place Each point Expressed as ,and ,in and Represent the predicted coordinates and hidden feature offsets respectively; 24) Cluster voting points based on spatial location similarity and farthest point sampling to obtain several sample clusters and voting points in each sample cluster ,in: represents the coordinates of the cluster center point, Represents the cluster radius; coordinate normalization is performed within each sample cluster to fully utilize the local geometric information within the cluster; 25) A proposal extraction module consisting of Max Pooling and multiple MLP layers is used to generate proposals for each sample cluster, and the initial latent features of the candidate graphic objects that include spatial boundary information are obtained; In step 3, the method steps of using the feature enhancement module to enhance the initial latent feature expression are as follows: 31) Use a fully connected layer to map the multi-view RGB features and local position features reflecting the attributes to the same dimension as the initial latent features of the candidate object. After adding the two, the fine-grained object attribute features are embedded; 32) Encode the pairwise Euclidean distances between the center points of different candidate objects and the component distances in the three coordinate axes of the Cartesian coordinate system, and construct a spatial adjacency matrix group to express the positional relationship between the candidate objects; 33) Encode through the fully connected layer to obtain the relational embedding features ,in: is the number of attention heads of the feature enhancement module.

2. The method for mining graphical modal design knowledge based on a three-dimensional model of a product component according to claim 1, characterized in that: In step 4, the method steps for generating a text description of the three-dimensional object are: 41) A 3D description text generation network is constructed based on a single-layer multi-head Transformer module to achieve information interaction between the enhanced latent features of the target candidate and the initial latent features of other candidate objects; 42) Select the target candidate, perform linear transformation on its potential features, and generate the query matrix, key matrix, and value matrix; 43) Initial potential features of the K candidate objects closest to the center of the target candidate object screened using the KNN strategy; 44) A post-processing module consisting of a fully connected layer and a word prediction layer is used to iteratively generate each word of the description text to obtain a text description of the three-dimensional object.

3. The method for mining graphical modal design knowledge based on a three-dimensional model of a product component according to claim 1, characterized in that: In step 5, the steps of the method for generating the graphical modality knowledge graph are as follows: 51) Obtain text descriptions of scenes and objects from graphic data using a 3D description model; 52) Use the text modality entity relationship extraction algorithm to conduct in-depth mining of text knowledge, extract the text into triples, and generate a graphical modality knowledge graph that conforms to the ontology system framework.

Images