A method and system for constructing virtual exhibition halls based on intelligent assisted design

Abstract

This invention relates to the field of virtual exhibition hall technology, providing a method and system for constructing a virtual exhibition hall based on intelligent assisted design, including: S1. Requirements analysis and intelligent material processing: obtaining design requirement texts through interaction with clients, parsing and extracting key design elements and functional keywords; simultaneously collecting various materials, automatically classifying, tagging, and performing correlation analysis to generate a structured material library; S2. Intelligent generation and optimization of 3D scenes: automatically generating structural suggestions for a basic 3D model of the exhibition hall based on the extracted key design elements, outputting a high-quality scene; S3. Personalized arrangement and dynamic integration of interactive content: recommending initial placement positions for materials, arranging materials in the virtual exhibition hall and adding interactive hotspots; then associating with audio narration files, embedding interactive 3D model demonstration controls, and adapting background music. This invention achieves intelligent and automated construction of virtual exhibition halls.

Description

Technical Field

[0001] This invention relates to the field of virtual exhibition hall technology, and more specifically, to a method and system for constructing virtual exhibition halls based on intelligent assisted design. Background Technology

[0002] Virtual exhibition halls, as a product of the deep integration of digital technology with culture and commercial display, have been widely used in various fields such as online exhibitions, corporate promotion, cultural and museum displays, education and training, and product launches. They can break through the limitations of time and space, providing an immersive and interactive digital visiting experience, and are gradually becoming an important supplement to and even a replacement for physical exhibition halls.

[0003] However, the current construction process of virtual exhibition halls still faces a series of challenges, with efficiency, cost, and level of intelligence being the main bottlenecks. Traditional construction methods typically rely heavily on professional designers and developers for manual operation throughout the entire process, specifically in the following aspects: First, during the requirements analysis and content preparation phase, designers need to repeatedly communicate with clients to understand and transform vague, unstructured requirements documents. This process is time-consuming and prone to errors. Meanwhile, facing a massive amount of display materials such as images, videos, and text, the workload of manually classifying, tagging, and establishing relationships is enormous, inefficient, and makes it difficult to ensure the systematic organization and ease of retrieval of the materials.

[0004] Secondly, in the 3D scene construction and optimization phase, from conceptual design to 3D modeling, material and lighting adjustments, and high-quality rendering, the entire process heavily relies on the designer's personal experience and artistic skills. Especially in achieving realistic visual effects, tedious parameter adjustments and lengthy rendering calculations are required. Existing modeling tools lack the intelligent understanding and assisted generation capabilities of design intent, resulting in a high barrier to entry and a long cycle in the transformation from requirements to 3D models.

[0005] Furthermore, during the interactive content layout and integration phase, the arrangement of exhibits, hotspots, and multimedia elements in three-dimensional space relies heavily on the designer's subjective experience, making it difficult to quantify and evaluate the rationality, logic, and visual comfort of the layout. Simultaneously, the smooth loading and real-time rendering of complex 3D models in the exhibition hall poses a challenge to network and terminal performance, directly impacting user experience. The adaptation of background music, voice narration, and other content to scene themes and user preferences also typically depends on manual selection, lacking personalization and automation.

[0006] In recent years, artificial intelligence technologies, particularly natural language processing, computer vision, generative AI, and reinforcement learning, have made significant progress. These technologies offer potential pathways for automating and intelligently solving the aforementioned problems. However, existing technologies are mostly applied piecemeal, such as using a single algorithm for image recognition or simple model generation. A systematic approach and process that deeply integrates multiple AI technologies and spans the entire lifecycle of virtual exhibition hall construction has not yet been formed, failing to achieve end-to-end intelligent assistance from needs understanding, scene generation, optimization to interactive layout.

[0007] Therefore, there is an urgent need for an innovative construction method that can organically integrate advanced AI capabilities into the design and development process of virtual exhibition halls, so as to significantly reduce technical barriers, improve construction efficiency, optimize output quality, and achieve a higher degree of personalization and automation. Summary of the Invention

[0008] The present invention provides a method and system for constructing virtual exhibition halls based on intelligent assisted design, which can overcome some or all of the defects of the prior art.

[0009] According to a method for constructing a virtual exhibition hall based on intelligent assisted design according to the present invention, the method includes the following steps: S1. Requirements Analysis and Intelligent Material Processing: We obtain design style, functional architecture, and display content requirements text through interaction with clients; The requirement text was parsed using natural language processing algorithms to extract key design elements and functional keywords. Collect and upload video, audio, text, and graphic materials to the preprocessing platform; Image recognition and semantic analysis algorithms are used to automatically classify, tag, and analyze the correlations of the graphic and textual materials to generate a structured material library; S2. Intelligent generation and optimization of 3D scenes: Based on the key design elements extracted in step S1, a pre-trained generative 3D model-assisted algorithm is invoked to automatically generate structural suggestions for the basic exhibition hall 3D model. Use 3D modeling tools to perform refined modeling and model optimization based on the structural recommendations; The optimized scene model is imported into the rendering engine, and the physically based rendering algorithm and lighting model algorithm output a high-quality panoramic image. An automatic face baking algorithm is used to process the scene model, and the algorithm is used to automatically optimize the lighting and material parameters; S3. Personalized layout and dynamic integration of interactive content: Based on the structured material library and key design elements generated in step S1, the layout optimization algorithm recommends initial placement positions for audio, video, and graphic materials in the virtual exhibition hall scene; On mature mainstream technology platforms, the materials are arranged and interactive hotspots are added according to user-defined instructions or confirmed recommended positions; When setting up exhibits, an audio-text matching algorithm is used to automatically associate or suggest matching audio explanation files; A 3D model demonstration control that supports user interaction is embedded, and the control has a built-in dynamic loading and real-time rendering algorithm; Based on user profiles or scene themes, background music is adapted and added using a background music recommendation algorithm.

[0010] Preferably, in step S1, the process of parsing the requirement text and extracting key design elements and functional keywords using natural language processing algorithms specifically includes the following stages: 1.1) Construction of domain knowledge base; Construct a domain knowledge base that includes a style thesaurus, a functional module library, and design templates, where: The style thesaurus is constructed by manually defining core style tags and expanding their related words, color and material features; The functional module library is built by defining standardized interactive functional components; The design template is constructed by abstracting historical success cases into reusable templates that include fixed style parameters, layout references, and functional module combinations. 1.2) Real-time analysis of customer needs; Upon receiving a customer request text, the following process is executed: 1.21) Clean and segment the required text, and use the Named Entity Recognition (NER) model to extract the core display objects; 1.22) Using a BERT-based pre-trained language model, the pre-processed customer demand statements and the terms and template descriptions in the domain knowledge base are converted into high-dimensional semantic vectors, and the semantic similarity between them is calculated. 1.23) Based on the semantic similarity calculation results, combined with context analysis and degree word weights, customer needs are mapped to style tags, functional modules and design templates in the knowledge base, and reliability is assigned to each matching item; 1.24) Based on the mapping and allocation results, automatically generate a structured requirements analysis report, which includes at least a style profile with confidence level, a list of functions, and suggestions for content display format; 1.3) Feedback learning and optimization; The system uses designers' adoption or modification of recommendations as feedback data to update the semantic mapping model and recommendation weights through online learning or reinforcement learning mechanisms, thereby achieving closed-loop optimization.

[0011] Preferably, in step S1, the automatic classification, tagging, and correlation analysis of the planar and textual materials using image recognition and semantic analysis algorithms to generate a structured material library specifically includes the following steps: 2.1) Perform visual recognition and semantic analysis on the aforementioned planar materials: 2.11) Extract the visual feature vector of the planar material using a convolutional neural network; 2.12) Identify entity objects in the planar material using an object detection model, and generate semantic description tags for the planar material using a visual-language model; 2.13) Based on the visual feature vectors and semantic description tags, the graphic materials are automatically classified into predefined content categories; 2.2) Perform semantic understanding and structured extraction on the text material: 2.21) Use a named entity recognition model to extract the core entities from the text material; 2.22) The keyword and topic distribution of the text material are identified through keyword extraction algorithms and topic models; 2.23) Based on sentiment analysis and text classification models, identify the sentiment tendency and stylistic features of the text material; 2.3) Cross-modal correlation analysis and structured integration: 2.31) Calculate the similarity between the visual feature vector of the planar material and the semantic feature vector of the text material to establish a semantic relationship between the image and the text; 2.32) Perform cluster analysis on the feature vectors of all materials, group materials belonging to the same theme or display unit into the same material cluster, and construct a knowledge graph or association matrix representing the relationship between materials; 2.4) Generate a structured resource library: Output a structured digital resource library containing the following information: 2.41) Metadata of the material; 2.42) Visual or textual feature vectors of the material; 2.43) Automatically assigned classification results and multi-dimensional labels; 2.44) Links to other related materials or related weight information.

[0012] Preferably, in step S2, the step of calling the pre-trained generative 3D model-assisted algorithm to automatically generate structural suggestions for the basic exhibition hall 3D model specifically includes the following stages: 3.1) Model training data preparation and algorithm training; 3.11) Construct a 3D scene training corpus: Collect a large number of data pairs containing text descriptions and their corresponding 3D scene models. The text descriptions are formatted descriptions of scene layout, style and function. The 3D scene models are uniformly converted into standardized formats such as voxel meshes or point clouds. 3.12) Training a text-driven generative 3D model: The model is trained using a variant algorithm based on Neural Radiation Field (NeRF) or a 3D diffusion model, enabling the model to learn the mapping relationship from text descriptions to 3D geometric structures. 3.2) Text-driven 3D structure generation; 3.21) Input preprocessing: Receive the structured design requirements from step S1 and synthesize them into specific three-dimensional spatial description prompts; 3.22) Conditional Generative Reasoning: The prompt words are input into the pre-trained generative 3D model. The text encoder of the model converts the prompt words into semantic condition vectors and guides the generation module to output a preliminary 3D structure that is aligned with the semantics of the text. The preliminary 3D structure is a low-resolution voxel representation or a sparse point cloud. 3.23) Post-processing and format conversion: Perform geometric normalization on the generated preliminary 3D structure and automatically convert it into an intermediate file format that can be recognized and edited by the target 3D modeling software; 3.3) Human-machine collaborative optimization and model iteration; 3.31) Design Collaboration: Designers review, refine, modify, and add details to the preliminary 3D structure in 3D modeling software; 3.32) Feedback learning: The final model modified by the designer and the original design requirement text are used as new training data pairs and fed back into the training corpus to perform incremental fine-tuning of the generative 3D model and achieve closed-loop optimization.

[0013] Preferably, in step S2, the optimized scene model is imported into the rendering engine, and a high-quality panoramic image is output using a physically based rendering algorithm and a lighting model algorithm. This specifically includes the following steps: 4.1) Data preparation and import: The 3D scene model data, including geometric structure, optimized material parameters and optimized lighting scheme, will be automatically configured and imported into the physically based rendering engine; 4.2) Physical rendering and global illumination calculation: The physical shader and global illumination algorithm of the rendering engine are called to simulate and render the material representation, direct lighting, indirect lighting and shadows of all surfaces in the scene based on micro-surface theory, energy conservation principle and light propagation physical model. 4.3) Automatic deployment and rendering of panoramic viewpoints: Based on the location of the scene's entrance and core exhibition area, one or more panoramic rendering cameras are automatically deployed; for each camera point, the rendering engine generates equidistant rectangular projection images covering six directions: front, back, left, right, up, and down. 4.4) Panoramic Image Synthesis and Standardized Output: The generated six directional images are automatically stitched together into a complete 2:1 spherical panoramic image, and output as a high-quality panoramic image file with a specified resolution, color space and file format according to preset standards.

[0014] Preferably, in step S2, the automatic optimization of lighting and material parameters specifically involves: 5.1) Constructing a reinforcement learning optimization environment and agent: Define a reinforcement learning framework that uses a 3D rendering engine as the simulation environment, where: The status includes the encoding of current lighting and material parameters, simplified geometric features of the scene, and image features and rendering time of the previous rendering result; The action involves adjusting the continuous values ​​of light source intensity, color, position, orientation, or material reflectivity and roughness parameters; The reward function is a weighted sum of visual realism reward and rendering efficiency reward. The visual realism reward is calculated by comparing the structural similarity index or perceptual similarity index between the rendered image and the reference image. The rendering efficiency reward is negatively correlated with the rendering time. 5.2) Offline training phase: On a virtual exhibition hall scene library containing various layouts and styles, the agent is trained using proximal policy optimization or deep deterministic policy gradient algorithms, enabling it to learn how to adjust parameters to maximize long-term cumulative rewards through iterative interaction. 5.3) Online optimization execution phase: For the newly modeled virtual exhibition hall scene, load the trained agent and execute the following loop until the termination condition is met: 5.31) The agent adjusts its actions based on the output parameters of the current state; 5.32) The rendering engine performs fast rendering based on the adjusted parameters and provides feedback on the new status and rewards; 5.33) The agent fine-tunes its strategy in the scenario based on the reward; 5.4) Output and Final Rendering: Select the set of lighting and material parameters that yielded the highest reward during the optimization process, apply them to the scene, and use this set of parameters to execute a high-quality panoramic rendering output.

[0015] Preferably, in step S3, the layout optimization algorithm is a multi-objective optimization algorithm based on genetic algorithm or particle swarm optimization algorithm, specifically including the following steps: 6.1) Define the input and decision variables for the optimization problem: Input includes: 6.11) The feasible region representation S extracted from the 3D scene; 6.12) A set C of content items to be arranged, where each content item c i Includes type, preset size, and logical labels; 6.13) The association matrix R represents the strength of logical relationships between content items; 6.14) The decision variable is the layout scheme L, which is expressed as: L = {(x i , y i , θ i ) | i = 1,…,N} Among them, (x i , y i ) is the content item c i The position coordinates θ in the scene i Its orientation angle; 6.2) Constructing a multi-objective fitness function: The fitness function F_total is a weighted sum of the following four cost functions: F total = w1·F1 + w2·F2 + w3·F3 + w4·F4 in: F1 is the visual balance cost, calculated as follows: The Euclidean distance between the visual quality center and the scene geometric center; F2 is the hotspot spacing conflict cost, calculated as follows: If the distance between content items is less than the minimum comfortable spacing or greater than the maximum related spacing, a penalty is applied; F3 is the cost of content logical relatedness, calculated as follows: F3 = Σ i Σ j (r) ij · d ij ) Where r ij For the correlation strength, d ij Distance between content items; F4 represents the cost of user preference deviation, which calculates the degree of deviation based on hard constraints or soft preferences specified by the user. w1, w2, w3, and w4 are user-preset weight coefficients; 6.3) Perform intelligent optimization search: Genetic algorithms or particle swarm optimization algorithms are used to achieve F totalThe fitness evaluation function is used to iteratively optimize the layout scheme L until the termination condition is met. 6.4) Output layout scheme and support human-computer interaction: It outputs one or more optimal layout schemes, allowing designers to select and fine-tune the schemes through an interactive interface, and automatically maps the final layout scheme to the virtual exhibition hall platform via API to complete the content arrangement.

[0016] Preferably, in step S3, the dynamic loading and real-time rendering algorithm built into the 3D model demonstration control specifically includes the following steps: 7.1) Data preprocessing and multi-level LOD model chain construction: The original high-precision 3D model is processed offline to generate a chain of LOD models with multiple levels of detail, where each LOD level corresponds to a different number of faces and texture precision, and the LOD models are spatially aligned. For ultra-large 3D models, octrees or KD trees are used for spatial partitioning to form multiple logical blocks, each of which independently possesses the LOD model chain; 7.2) Runtime intelligent monitoring and dynamic scheduling decision-making: 7.21) Real-time collection of user viewpoint parameters, client performance parameters, and network status parameters; 7.22) Based on the viewpoint parameters, calculate the theoretically required target LOD level for each model or block in the field of view, and generate a loading request queue; 7.23) Based on network bandwidth and client performance, the request queue is prioritized and traffic controlled, wherein the priority is determined based on the distance between the block and the center of view and its contribution to the image. 7.3) Asynchronous streaming loading and progressive rendering: 7.31) Send the priority-sorted requests to the cloud server to obtain the corresponding block and LOD-level compressed model data package; 7.32) The client asynchronously receives, decodes, and buffers the data packets in the background; 7.33) The renderer retrieves available data from the cache according to priority and performs real-time drawing. When switching between high and low LOD, it uses geometric deformation or fade-in / fade-out techniques to achieve a smooth visual transition. 7.4) Real-time feedback and adaptive adjustment: Monitor the rendering frame rate. If the frame rate is lower than the target threshold, actively reduce the LOD level of the global or non-core areas, or reduce the loading requests of blocks outside the field of view. If the frame rate remains stable and resources are sufficient, the LOD level in the center of the field of view will be gradually increased to optimize visual quality.

[0017] Preferably, in step S3, the background music recommendation algorithm is specifically as follows: 8.1) Multi-dimensional feature extraction and representation: Extract the semantic vector V of the exhibition hall theme text text Emotional feature vector V of visual materials visual and user historical preference vector V user ; Wherein, the visual emotion feature vector V visual It is obtained by fusing color sentiment analysis, scene semantic recognition and visual sentiment prediction models; 8.2) Hybrid Recommendation Model Inference: 8.21) Initial screening based on content filtering: Constructing the joint feature vector V of the exhibition hall exhibition = α·V text + β·V visual , where α and β are adjustable weights; Calculate V exhibition With the audio feature vector V of each song in the music library music Based on cross-modal similarity, the Top-K most similar musical pieces were selected as the initial candidate set; 8.22) Reordering based on collaborative filtering: The initial candidate set, V exhibition and V user Input a ranking model, predict the adoption rate score for each piece of music, and output a ranked list of recommended background music. 8.3) Interactive Feedback and Closed-Loop Optimization: Record the designer's selection or rejection of recommended music and assign the corresponding V exhibition V music Data pairs are added to the training set; The ranking model and feature encoder are retrained periodically with new data to achieve continuous optimization of the recommendation system.

[0018] This invention provides a virtual exhibition hall construction system based on intelligent assisted design, which adopts the above-mentioned virtual exhibition hall construction method based on intelligent assisted design.

[0019] The beneficial effects of this invention are as follows: This invention significantly improves construction efficiency and automation while reducing costs. It automatically parses client requirements using natural language processing algorithms and accurately maps them using a pre-built domain knowledge base, rapidly transforming unstructured text into structured design guidelines. This greatly reduces the time and errors associated with repeated communication and interpretation of requirements between designers and clients. Furthermore, it utilizes image recognition and semantic analysis algorithms to automatically classify, label, and analyze the associations of materials, generating a structured material library. This replaces the inefficient work of manually organizing massive amounts of material, laying a high-efficiency data foundation for subsequent content layout. Finally, through a pre-trained generative 3D model, it can automatically generate structural suggestions for a basic exhibition hall based on text descriptions, providing designers with a high-quality starting point for modeling and shortening the cycle from concept to 3D sketch.

[0020] This invention uses a reinforcement learning-based algorithm to automatically optimize lighting and material parameters. It can efficiently find the best balance between rendering quality and computational cost while ensuring visual realism, reducing the tedious manual debugging burden on designers and outputting higher quality and more physically realistic panoramic images.

[0021] The dynamic loading and real-time rendering algorithms (such as multi-level LOD and asynchronous streaming loading) built into the 3D model demonstration control of this invention ensure smooth loading and rendering of complex models on terminals with different performance levels, thereby improving the stability and immersion of the user experience.

[0022] This invention employs a multi-objective layout optimization algorithm based on genetic algorithms or particle swarm optimization algorithms. It comprehensively considers various factors such as visual balance, logical association, and user preferences to recommend the optimal placement scheme for exhibits and hotspots, making the exhibition hall layout more logical and visually comfortable, surpassing the design that relies solely on human experience.

[0023] This invention achieves intelligent matching and recommendation of audio explanations, background music, exhibit content, exhibition hall theme, and even user profiles through audio-text matching algorithms and background music recommendation algorithms, thereby enhancing the relevance of exhibition hall content and the personalized emotional experience of visitors.

[0024] This invention provides algorithm-based quantitative suggestions and confidence assessments (such as structured requirement reports and multiple layout schemes) in multiple stages, including requirement analysis, scene generation, and layout planning. It provides designers with strong data support and decision assistance, rather than completely replacing them, thus achieving efficient human-computer collaboration.

[0025] This invention utilizes a feedback learning mechanism (such as online learning and incremental fine-tuning) to feed back designers' modifications and user interaction data, continuously optimizing the requirement analysis model, 3D generation model, and recommendation algorithm. This allows the entire system to learn and improve continuously during use, becoming increasingly intelligent with each use. Attached Figure Description

[0026] Figure 1 This is a flowchart of a virtual exhibition hall construction method based on intelligent assisted design, as shown in the embodiment. Detailed Implementation

[0027] To further understand the content of this invention, a detailed description of the invention will be provided in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention. Example

[0028] like Figure 1 As shown, this embodiment provides a method for constructing a virtual exhibition hall based on intelligent assisted design, which includes the following steps: S1. Requirements Analysis and Intelligent Material Processing: We obtain design style, functional architecture, and display content requirements text through interaction with clients; The requirement text was parsed using natural language processing algorithms to extract key design elements and functional keywords. Collect and upload video, audio, text, and graphic materials to the preprocessing platform; Image recognition and semantic analysis algorithms are used to automatically classify, tag, and analyze the correlations of the graphic and textual materials to generate a structured material library; S2. Intelligent generation and optimization of 3D scenes: Based on the key design elements extracted in step S1, a pre-trained generative 3D model-assisted algorithm is invoked to automatically generate structural suggestions for the basic exhibition hall 3D model. Use 3D modeling tools (SketchUp or 3ds Max) to perform detailed modeling and model optimization based on the structural suggestions; The optimized scene model is imported into the rendering engine, and the physically based rendering algorithm and lighting model algorithm output a high-quality panoramic image. An automatic face baking algorithm is used to process the scene model, and the algorithm is used to automatically optimize the lighting and material parameters; S3. Personalized layout and dynamic integration of interactive content: Based on the structured material library and key design elements generated in step S1, the layout optimization algorithm recommends initial placement positions for audio, video, and graphic materials in the virtual exhibition hall scene; On mature mainstream technology platforms, the materials are arranged and interactive hotspots are added according to user-defined instructions or confirmed recommended positions; When setting up exhibits, an audio-text matching algorithm is used to automatically associate or suggest matching audio explanation files; A 3D model demonstration control that supports user interaction is embedded, and the control has a built-in dynamic loading and real-time rendering algorithm; Based on user profiles or scene themes, background music is adapted and added using a background music recommendation algorithm.

[0029] This embodiment first performs requirements analysis and intelligent material processing, then intelligently generates and optimizes the 3D scene, and finally personalizes and dynamically integrates interactive content. Through data transfer and algorithmic collaboration between the preceding and following steps (e.g., using the "key design elements" extracted in S1 directly in S2 to generate "structural suggestions for the basic exhibition hall 3D model," and using the "structured material library" generated in S1 for "recommended initial placement locations" in S3), an end-to-end intelligent production pipeline is constructed. This achieves automated assembly line operations, significantly reducing the overall time cycle from initial requirements to an interactive exhibition hall.

[0030] By using "natural language processing algorithms to parse requirement text" and "generative 3D model-assisted algorithms to automatically generate structural suggestions for basic exhibition hall 3D models," even planners without extensive experience in 3D spatial design can quickly obtain a professional 3D scene prototype that meets their intentions through text descriptions, greatly reducing the professional threshold and start-up cost of 3D scene creation.

[0031] This embodiment can automatically generate layout schemes that conform to both visual aesthetics and content logic, realize intelligent association of audio, visuals and text, and ensure that various resources (especially high-precision 3D models) can be loaded and interacted with smoothly on various terminals, thereby systematically improving the interactive smoothness, content logic and visitor immersion of the final virtual exhibition hall.

[0032] In step S1, the process of parsing the requirement text and extracting key design elements and functional keywords using natural language processing algorithms specifically includes the following stages: 1.1) Construction of domain knowledge base; Construct a domain knowledge base that includes a style thesaurus, a functional module library, and design templates, where: The style thesaurus is constructed by manually defining core style tags and expanding their related words, color and material features; The functional module library is built by defining standardized interactive functional components; The design template is constructed by abstracting historical success cases into reusable templates that include fixed style parameters, layout references, and functional module combinations. 1.2) Real-time analysis of customer needs; Upon receiving a customer request text, the following process is executed: 1.21) Clean and segment the required text, and use the Named Entity Recognition (NER) model to extract the core display objects; 1.22) Using a BERT-based pre-trained language model, the pre-processed customer demand statements and the terms and template descriptions in the domain knowledge base are converted into high-dimensional semantic vectors, and the semantic similarity between them is calculated. 1.23) Based on the semantic similarity calculation results, combined with context analysis and degree word weights, customer needs are mapped to style tags, functional modules and design templates in the knowledge base, and reliability is assigned to each matching item; 1.24) Based on the mapping and allocation results, automatically generate a structured requirements analysis report, which includes at least a style profile with confidence level, a list of functions, and suggestions for content display format; 1.3) Feedback learning and optimization; The system uses designers' adoption or modification of recommendations as feedback data to update the semantic mapping model and recommendation weights through online learning or reinforcement learning mechanisms, thereby achieving closed-loop optimization.

[0033] Step 1.1) constructs a structured domain knowledge base (including a style thesaurus, functional module library, and design templates). This is a structured knowledge system that deeply integrates professional knowledge of virtual showroom design. When the system analyzes customer requirements (Step 1.2), it can understand and match them within a professional context. For example, if a customer describes "wanting the showroom to have a technological feel," the system can accurately map this to specific executable design parameters and reference cases such as "cool colors," "linear lighting," and "metallic materials" based on the knowledge base. This avoids ambiguity or superficiality that may arise from general semantic understanding, and outputs an accurate structured requirements report.

[0034] The process described in detail in step 1.2) (cleaning, NER extraction, BERT-based semantic vector matching, and confidence level assignment) achieves a key transformation: converting the client's subjective and vague natural language descriptions (such as "more sophisticated" or "easy to use") into quantifiable style tags, feature lists, and content format suggestions with confidence level assessments. The direct result is a well-structured, clearly defined requirements analysis report with matching confidence levels. This provides designers and clients with an objective and efficient basis for communication and decision-making, significantly reducing repeated revisions due to misunderstandings and making the goal consensus process during project initiation more scientific and efficient.

[0035] In step 1.23), the output not only provides recommendations but also assigns them confidence levels. This makes the system's decision-making process more transparent and interpretable. Designers can quickly focus on high-confidence recommendations for confirmation, while paying close attention to low-confidence or conflicting items for manual review or adjustment. This design organically combines the computing power and data breadth of artificial intelligence with the experience and aesthetic judgment of designers, improving efficiency while ensuring the professionalism and creativity of the final solution, and enhancing the reliability and acceptability of the human-machine collaboration model.

[0036] Step 1.3) of feedback learning and optimization explicitly uses the designers' adoption or modification results as feedback data to update the model and weights. This allows the system to become increasingly accurate in understanding the preferences of specific customer groups (such as a certain industry) and specific designer teams as the number of uses increases, and the confidence and accuracy of recommendations will continue to improve, thereby achieving a personalized service capability that becomes smarter and more tailored with use.

[0037] In step S1, the automatic classification, tagging, and correlation analysis of the planar and textual materials using image recognition and semantic analysis algorithms to generate a structured material library specifically includes the following steps: 2.1) Perform visual recognition and semantic analysis on the aforementioned planar materials: 2.11) Extract the visual feature vector of the planar material using a convolutional neural network; 2.12) Identify entity objects in the planar material using an object detection model, and generate semantic description tags for the planar material using a visual-language model; 2.13) Based on the visual feature vectors and semantic description tags, the graphic materials are automatically classified into predefined content categories; 2.2) Perform semantic understanding and structured extraction on the text material: 2.21) Use a named entity recognition model to extract the core entities from the text material; 2.22) The keyword and topic distribution of the text material are identified through keyword extraction algorithms and topic models; 2.23) Based on sentiment analysis and text classification models, identify the sentiment tendency and stylistic features of the text material; 2.3) Cross-modal correlation analysis and structured integration: 2.31) Calculate the similarity between the visual feature vector of the planar material and the semantic feature vector of the text material to establish a semantic relationship between the image and the text; 2.32) Perform cluster analysis on the feature vectors of all materials, group materials belonging to the same theme or display unit into the same material cluster, and construct a knowledge graph or association matrix representing the relationship between materials; 2.4) Generate a structured resource library: Output a structured digital resource library containing the following information: 2.41) Metadata of the material; 2.42) Visual or textual feature vectors of the material; 2.43) Automatically assigned classification results and multi-dimensional labels; 2.44) Links to other related materials or related weight information.

[0038] Steps 2.1 and 2.2 utilize convolutional neural networks, object detection models, named entity recognition, and topic models to perform deep content analysis on image and text materials. The beneficial effect is that it can automatically complete the tasks of viewing, summarizing, tagging, and abstracting that previously required significant manual effort. The system can not only identify objects in images and generate semantic descriptions, but also extract core entities and themes from text, and even analyze sentiment and style. This not only completely liberates designers from tedious material organization work, but more importantly, the machine-executed classification and tagging has the advantages of objectivity, consistency, and completeness, fundamentally avoiding errors caused by human processing, subjective bias, and fatigue, and significantly improving the quality and usability of the underlying material data.

[0039] Steps 2.3 (Cross-modal association analysis) and 2.4 (Generating a structured resource library) automatically discover inherent, semantic-level connections between images and text, between images themselves, and between texts by calculating the similarity of image and text feature vectors, performing cluster analysis, and constructing association matrices or knowledge graphs. For example, it automatically groups a product's design drawings, promotional copy, and user reviews into the same "resource cluster." This directly generates a "structured digital resource library" with rich contextual relationships. This allows designers or subsequent algorithms to quickly retrieve and call resources by topic and logical relationship, as if in an interconnected knowledge network, providing a solid data foundation for subsequent intelligent display (such as placing strongly related resources in adjacent positions).

[0040] Through in-depth understanding of text and image content, cross-modal semantic association, and standardized data structured output, not only has the efficiency and accuracy of the material preparation stage been greatly improved, but more importantly, a high-quality data foundation has been laid for the intelligent construction of the entire virtual exhibition hall, and intelligent curation and personalized experience have been directly empowered.

[0041] In step S2, the step of calling the pre-trained generative 3D model-assisted algorithm to automatically generate structural suggestions for the basic exhibition hall 3D model specifically includes the following stages: 3.1) Model training data preparation and algorithm training; 3.11) Construct a 3D scene training corpus: Collect a large number of data pairs containing text descriptions and their corresponding 3D scene models. The text descriptions are formatted descriptions of scene layout, style and function. The 3D scene models are uniformly converted into standardized formats such as voxel meshes or point clouds. 3.12) Training a text-driven generative 3D model: The model is trained using a variant algorithm based on Neural Radiation Field (NeRF) or a 3D diffusion model, enabling the model to learn the mapping relationship from text descriptions to 3D geometric structures. 3.2) Text-driven 3D structure generation; 3.21) Input preprocessing: Receive the structured design requirements from step S1 and synthesize them into specific three-dimensional spatial description prompts; 3.22) Conditional Generative Reasoning: The prompt words are input into the pre-trained generative 3D model. The text encoder of the model converts the prompt words into semantic condition vectors and guides the generation module to output a preliminary 3D structure that is aligned with the semantics of the text. The preliminary 3D structure is a low-resolution voxel representation or a sparse point cloud. 3.23) Post-processing and format conversion: Perform geometric normalization on the generated preliminary 3D structure and automatically convert it into an intermediate file format that can be recognized and edited by the target 3D modeling software; 3.3) Human-machine collaborative optimization and model iteration; 3.31) Design Collaboration: Designers review, refine, modify, and add details to the preliminary 3D structure in 3D modeling software; 3.32) Feedback learning: The final model modified by the designer and the original design requirement text are used as new training data pairs and fed back into the training corpus to perform incremental fine-tuning of the generative 3D model and achieve closed-loop optimization.

[0042] By utilizing pre-trained generative algorithms based on NeRF or 3D diffusion models, the "structured design requirements" (text prompts) from S1 are directly mapped to "preliminary 3D structures" (voxel meshes or point clouds). This automates and intelligently transforms abstract design requirements into concrete 3D spatial structures, greatly liberating designers' creative productivity and allowing them to focus on refinement and optimization rather than building from scratch. Through the application of cutting-edge generative models, the provision of editable intermediate results, and the establishment of a human-computer feedback loop, not only is the low efficiency in the early stages of 3D creation solved, but the collaborative relationship between designers and digital tools is also redefined at a deeper level.

[0043] In step S2, the optimized scene model is imported into the rendering engine, and the physically based rendering algorithm and lighting model algorithm output a high-quality panoramic image, specifically including the following steps: 4.1) Data preparation and import: The 3D scene model data, including geometric structure, optimized material parameters and optimized lighting scheme, will be automatically configured and imported into the physically based rendering engine; 4.2) Physical rendering and global illumination calculation: The physical shader and global illumination algorithm of the rendering engine are called to simulate and render the material representation, direct lighting, indirect lighting and shadows of all surfaces in the scene based on micro-surface theory, energy conservation principle and light propagation physical model. 4.3) Automatic deployment and rendering of panoramic viewpoints: Based on the location of the scene's entrance and core exhibition area, one or more panoramic rendering cameras are automatically deployed; for each camera point, the rendering engine generates equidistant rectangular projection images covering six directions: front, back, left, right, up, and down. 4.4) Panoramic Image Synthesis and Standardized Output: The generated six directional images are automatically stitched together into a complete 2:1 spherical panoramic image, and output as a high-quality panoramic image file with a specified resolution, color space and file format according to preset standards.

[0044] This embodiment abandons traditional simplified or subjective rendering methods. By physically simulating the interaction between light and materials, it achieves near-photorealistic quality in the generated panoramic images of the exhibition hall, with excellent material texture, light and shadow transitions, global illumination (including indirect diffuse reflection), and shadow details. This provides end users with a highly realistic and deeply immersive visual environment. This embodiment completely automates the complex and error-prone process of manually selecting camera positions, setting six-directional rendering, and post-processing stitching. It can intelligently determine the best viewing points based on the scene's entrance and core exhibition area logic, and automatically complete the seamless image acquisition and compositing covering the entire spherical space. This not only greatly improves production efficiency but, more importantly, completely eliminates problems such as viewpoint omissions, stitching misalignments, or inconsistent brightness that may be caused by human operation, ensuring that every output panoramic image is a standardized product with complete structure and visual uniformity. Through physically realistic rendering technology, intelligent automatic shooting and compositing processes, and strict output standardization, this embodiment efficiently and reliably solves the core requirement of high-quality visual content production in the construction of virtual exhibition halls.

[0045] In step S2, the automatic optimization of lighting and material parameters specifically involves: 5.1) Constructing a reinforcement learning optimization environment and agent: Define a reinforcement learning framework that uses a 3D rendering engine as the simulation environment, where: The status includes the encoding of current lighting and material parameters, simplified geometric features of the scene, and image features and rendering time of the previous rendering result; The action involves adjusting the continuous values ​​of light source intensity, color, position, orientation, or material reflectivity and roughness parameters; The reward function is a weighted sum of visual realism reward and rendering efficiency reward. The visual realism reward is calculated by comparing the structural similarity index or perceptual similarity index between the rendered image and the reference image. The rendering efficiency reward is negatively correlated with the rendering time. 5.2) Offline training phase: On a virtual exhibition hall scene library containing various layouts and styles, the agent is trained using proximal policy optimization or deep deterministic policy gradient algorithms, enabling it to learn how to adjust parameters to maximize long-term cumulative rewards through iterative interaction. 5.3) Online optimization execution phase: For the newly modeled virtual exhibition hall scene, load the trained agent and execute the following loop until the termination condition is met: 5.31) The agent adjusts its actions based on the output parameters of the current state; 5.32) The rendering engine performs rapid rendering based on the adjusted parameters and provides feedback on the new state and rewards; 5.33) The agent fine-tunes its strategy in the scene based on the rewards; 5.4) Output and Final Rendering: Select the set of lighting and material parameters that yielded the highest reward during the optimization process, apply them to the scene, and use this set of parameters to execute a high-quality panoramic rendering output.

[0046] This embodiment models the overall optimization of dozens or even hundreds of continuously interacting parameters of light (intensity, color, position, etc.) and materials (reflectivity, roughness, etc.) as a sequential decision-making process. Through continuous interaction and trial and error between the agent and the environment (step 5.3), the system can autonomously explore a vast parameter space and automatically learn parameter combination strategies that approximate the optimal one. This completely changes the traditional mode that relies on manual work and experience by designers, solving the problem that manual debugging is difficult to achieve global optimization and is prone to getting trapped in local optima.

[0047] This embodiment guides the agent to find the optimal set of parameters that achieves the best visual effect within an acceptable rendering time. The final output parameters (step 5.4) represent a solution that achieves an intelligent balance between quality and efficiency. For example, it may learn to simulate complex global illumination effects by cleverly arranging a few efficient light sources, thereby achieving near-perfect visual quality with a shorter rendering time, achieving the optimal "cost-effectiveness" that is difficult to achieve with traditional methods.

[0048] In step S3, the layout optimization algorithm is a multi-objective optimization algorithm based on genetic algorithm or particle swarm optimization algorithm, specifically including the following steps: 6.1) Define the input and decision variables for the optimization problem: Input includes: 6.11) The feasible region representation S extracted from the 3D scene; 6.12) A set C of content items to be arranged, where each content item c i Includes type, preset size, and logical labels; 6.13) The association matrix R represents the strength of logical relationships between content items; 6.14) The decision variable is the layout scheme L, which is expressed as: L = {(x i , y i , θ i ) | i = 1,…,N} Among them, (x i , y i ) is the content item c i The position coordinates θ in the scene i Its orientation angle; 6.2) Constructing a multi-objective fitness function: The fitness function F_total is a weighted sum of the following four cost functions: F total = w1·F1 + w2·F2 + w3·F3 + w4·F4 in: F1 is the visual balance cost, calculated as follows: The Euclidean distance between the visual quality center and the scene geometric center; F2 is the hotspot spacing conflict cost, calculated as follows: If the distance between content items is less than the minimum comfortable spacing or greater than the maximum related spacing, a penalty is applied; F3 is the cost of content logical relatedness, calculated as follows: F3 = Σ i Σ j (r) ij · d ij ) Where r ij For the correlation strength, d ij Distance between content items; F4 represents the cost of user preference deviation, which calculates the degree of deviation based on hard constraints or soft preferences specified by the user. w1, w2, w3, and w4 are user-preset weight coefficients; 6.3) Perform intelligent optimization search: Genetic algorithms or particle swarm optimization algorithms are used to achieve F total The fitness evaluation function is used to iteratively optimize the layout scheme L until the termination condition is met. 6.4) Output layout scheme and support human-computer interaction: It outputs one or more optimal layout schemes, allowing designers to select and fine-tune the schemes through an interactive interface, and automatically maps the final layout scheme to the virtual exhibition hall platform via API to complete the content arrangement.

[0049] By constructing a quantified fitness function encompassing four dimensions (visual balance F1, spacing conflict F2, logical association F3, and user preference F4) and employing a genetic algorithm or particle swarm optimization algorithm for intelligent search, the layout problem is transformed into a well-defined, multi-objective mathematical optimization problem. This embodiment simultaneously considers avoiding visual top-heavy layouts, ensuring comfortable browsing distances between hotspots, placing content-related exhibits close together, and respecting specific user requirements. This solves the problem of difficulty in overall coordination and neglect of details in manual layout, automatically exploring and recommending layout schemes that achieve comprehensive optimization across multiple indicators, making the exhibition space both aesthetically pleasing and functional.

[0050] This embodiment allows users to preset the weight coefficients (w1, w2, w3, w4) of four cost functions, providing designers with an advanced strategy control panel. For example, if the design emphasizes narrative coherence, the designer can increase the logical connection weight w3; if emphasizing a grand visual effect, the visual balance weight w1 can be increased. This allows designers to directly input their design strategies and priorities to guide the algorithm, resulting in more personalized and strategically consistent solutions. Simultaneously, the optimization process and results are based on explicit mathematical calculations, making the basis for layout decisions transparent and interpretable, facilitating internal team communication and solution review.

[0051] In step S3, the dynamic loading and real-time rendering algorithm built into the 3D model demonstration control specifically includes the following steps: 7.1) Data preprocessing and multi-level LOD model chain construction: The original high-precision 3D model is processed offline to generate a chain of LOD models with multiple levels of detail, where each LOD level corresponds to a different number of faces and texture precision, and the LOD models are spatially aligned. For ultra-large 3D models, octrees or KD trees are used for spatial partitioning to form multiple logical blocks, each of which independently possesses the LOD model chain; 7.2) Runtime intelligent monitoring and dynamic scheduling decision-making: 7.21) Real-time collection of user viewpoint parameters, client performance parameters, and network status parameters; 7.22) Based on the viewpoint parameters, calculate the theoretically required target LOD level for each model or block in the field of view, and generate a loading request queue; 7.23) Based on network bandwidth and client performance, the request queue is prioritized and traffic controlled, wherein the priority is determined based on the distance between the block and the center of view and its contribution to the image. 7.3) Asynchronous streaming loading and progressive rendering: 7.31) Send the priority-sorted requests to the cloud server to obtain the corresponding block and LOD-level compressed model data package; 7.32) The client asynchronously receives, decodes, and buffers the data packets in the background; 7.33) The renderer retrieves available data from the cache according to priority and performs real-time drawing. When switching between high and low LOD, it uses geometric deformation or fade-in / fade-out techniques to achieve a smooth visual transition. 7.4) Real-time feedback and adaptive adjustment: Monitor the rendering frame rate. If the frame rate is lower than the target threshold, actively reduce the LOD level of the global or non-core areas, or reduce the loading requests of blocks outside the field of view. If the frame rate remains stable and resources are sufficient, the LOD level in the center of the field of view will be gradually increased to optimize visual quality.

[0052] By constructing a multi-level LOD model chain and spatial block division (step 7.1), and combining it with intelligent scheduling based on viewpoint and performance (step 7.2), precise resource allocation is achieved. This allows for real-time dynamic loading of model data with the most appropriate level of detail based on the user's viewing distance and angle, while using low-detail representation for areas outside the field of view or in the distance. This enables smooth browsing of complex high-poly models, including large architectural sculptures and precision industrial equipment, even on ordinary web browsers, mobile devices, or with limited bandwidth. It breaks the dependence of such content on high-end graphics workstations, allowing high-quality virtual exhibition halls to be popularized on a wider range of consumer devices.

[0053] This embodiment employs geometric deformation or fade-in / fade-out techniques when switching between high and low LOD levels. This eliminates visual jumps or flickering caused by sudden changes in model details during dynamic loading. This smooth transition ensures that despite frequent data scheduling behind the scenes, the visuals perceived by the user remain continuous and stable. This is crucial for maintaining the realism of the virtual environment and the user's deep immersion, ensuring that performance optimization does not come at the expense of visual quality.

[0054] The real-time feedback and adaptive adjustment mechanism described in step 7.4 brings the beneficial effect of dynamic adaptation. The system continuously monitors the rendering frame rate and proactively adjusts the LOD level or loading strategy. This means that whether faced with sudden performance fluctuations (such as background programs consuming resources) or aggressive operations like rapid perspective switching by the user, the system can automatically and in real-time adjust the rendering load to ensure that the interactive frame rate does not fall below an acceptable threshold. This provides users with a robust and smooth experience, ensuring that the virtual exhibition hall can run stably under different hardware and different operational intensities.

[0055] In step S3, the background music recommendation algorithm specifically includes: 8.1) Multi-dimensional feature extraction and representation: Extract the semantic vector V of the exhibition hall theme text text Emotional feature vector V of visual materials visual and user historical preference vector V user ; Wherein, the visual emotion feature vector V visual It is obtained by fusing color sentiment analysis, scene semantic recognition and visual sentiment prediction models; 8.2) Hybrid Recommendation Model Inference: 8.21) Initial screening based on content filtering: Constructing the joint feature vector V of the exhibition hall exhibition = α·V text + β·V visual , where α and β are adjustable weights; Calculate V exhibition With the audio feature vector V of each song in the music library music Based on cross-modal similarity, the Top-K most similar musical pieces were selected as the initial candidate set; 8.22) Reordering based on collaborative filtering: The initial candidate set, V exhibition and V user Input a ranking model, predict the adoption rate score for each piece of music, and output a ranked list of recommended background music. 8.3) Interactive Feedback and Closed-Loop Optimization: Record the designer's selection or rejection of recommended music and assign the corresponding V exhibition V music Data pairs are added to the training set; The ranking model and feature encoder are retrained periodically with new data to achieve continuous optimization of the recommendation system.

[0056] By deeply integrating multi-dimensional features, employing hybrid recommendation strategies, and optimizing closed-loop feedback, the core issue of personalized environmental sound effects adaptation in virtual exhibition halls has been intelligently addressed. This has enabled cross-modal deep emotional matching, taking into account both content relevance and individual preferences, and possessing continuous self-evolution capabilities, thereby improving music selection efficiency and brand consistency.

[0057] This embodiment provides a virtual exhibition hall construction system based on intelligent assisted design, which adopts the above-mentioned virtual exhibition hall construction method based on intelligent assisted design.

[0058] This embodiment effectively solves the core pain points of traditional virtual exhibition hall construction, such as low efficiency, over-reliance on human experience, lack of personalization, and inconsistent quality, by constructing a full-chain intelligent assisted design system that integrates intelligent demand analysis, intelligent material processing, intelligent scene generation and optimization, and intelligent layout and integration of interactive content. It achieves quality improvement, efficiency enhancement, cost reduction and experience upgrade in the construction process, and has significant practical value and industrial application prospects.

[0059] The present invention and its embodiments have been described above illustratively. This description is not restrictive, and the figures shown are only one embodiment of the present invention; the actual structure is not limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for constructing a virtual exhibition hall based on intelligent assisted design, characterized in that: Includes the following steps: S1. Requirements Analysis and Intelligent Material Processing: We obtain design style, functional architecture, and display content requirements text through interaction with clients; The requirement text was parsed using natural language processing algorithms to extract key design elements and functional keywords. Collect and upload video, audio, text, and graphic materials to the preprocessing platform; Image recognition and semantic analysis algorithms are used to automatically classify, tag, and analyze the correlations of the graphic and textual materials to generate a structured material library; S2. Intelligent generation and optimization of 3D scenes: Based on the key design elements extracted in step S1, a pre-trained generative 3D model-assisted algorithm is invoked to automatically generate structural suggestions for the basic exhibition hall 3D model. Use 3D modeling tools to perform refined modeling and model optimization based on the structural recommendations; The optimized scene model is imported into the rendering engine, and the physically based rendering algorithm and lighting model algorithm output a high-quality panoramic image. An automatic face baking algorithm is used to process the scene model, and the algorithm is used to automatically optimize the lighting and material parameters; S3. Personalized layout and dynamic integration of interactive content: Based on the structured material library and key design elements generated in step S1, the layout optimization algorithm recommends initial placement positions for audio, video, and graphic materials in the virtual exhibition hall scene; On mature mainstream technology platforms, the materials are arranged and interactive hotspots are added according to user-defined instructions or confirmed recommended positions; When setting up exhibits, an audio-text matching algorithm is used to automatically associate or suggest matching audio explanation files; A 3D model demonstration control that supports user interaction is embedded, and the control has a built-in dynamic loading and real-time rendering algorithm; Based on user profiles or scene themes, background music is adapted and added using a background music recommendation algorithm.

2. The method for constructing a virtual exhibition hall based on intelligent assisted design according to claim 1, characterized in that: In step S1, the requirement text is parsed using a natural language processing algorithm to extract key design elements and functional keywords. Includes the following stages: 1.1) Construction of domain knowledge base; Construct a domain knowledge base that includes a style thesaurus, a functional module library, and design templates, where: The style thesaurus is constructed by manually defining core style tags and expanding their related words, color and material features; The functional module library is built by defining standardized interactive functional components; The design template is constructed by abstracting historical success cases into reusable templates that include fixed style parameters, layout references, and functional module combinations. 1.2) Real-time analysis of customer needs; Upon receiving a customer request text, the following process is executed: 1.21) Clean and segment the required text, and use the Named Entity Recognition (NER) model to extract the core display objects; 1.22) Using a BERT-based pre-trained language model, the pre-processed customer demand statements and the terms and template descriptions in the domain knowledge base are converted into high-dimensional semantic vectors, and the semantic similarity between them is calculated. 1.23) Based on the semantic similarity calculation results, combined with context analysis and degree word weights, customer needs are mapped to style tags, functional modules and design templates in the knowledge base, and reliability is assigned to each matching item; 1.24) Based on the mapping and allocation results, automatically generate a structured requirements analysis report, which includes at least a style profile with confidence level, a list of functions, and suggestions for content display format; 1.3) Feedback learning and optimization; The system uses designers' adoption or modification of recommendations as feedback data to update the semantic mapping model and recommendation weights through online learning or reinforcement learning mechanisms, thereby achieving closed-loop optimization.

3. The method for constructing a virtual exhibition hall based on intelligent assisted design according to claim 2, characterized in that: In step S1, the automatic classification, tagging, and correlation analysis of the planar and textual materials using image recognition and semantic analysis algorithms to generate a structured material library specifically includes the following steps: 2.1) Perform visual recognition and semantic analysis on the aforementioned planar materials: 2.11) Extract the visual feature vector of the planar material using a convolutional neural network; 2.12) Identify entity objects in the planar material using an object detection model, and generate semantic description tags for the planar material using a visual-language model; 2.13) Based on the visual feature vectors and semantic description tags, the graphic materials are automatically classified into predefined content categories; 2.2) Perform semantic understanding and structured extraction on the text material: 2.21) Use a named entity recognition model to extract the core entities from the text material; 2.22) The keyword and topic distribution of the text material are identified through keyword extraction algorithms and topic models; 2.23) Based on sentiment analysis and text classification models, identify the sentiment tendency and stylistic features of the text material; 2.3) Cross-modal correlation analysis and structured integration: 2.31) Calculate the similarity between the visual feature vector of the planar material and the semantic feature vector of the text material to establish a semantic relationship between the image and the text; 2.32) Perform cluster analysis on the feature vectors of all materials, group materials belonging to the same theme or display unit into the same material cluster, and construct a knowledge graph or association matrix representing the relationship between materials; 2.4) Generate a structured resource library: Output a structured digital resource library containing the following information: 2.41) Metadata of the material; 2.42) Visual or textual feature vectors of the material; 2.43) Automatically assigned classification results and multi-dimensional labels; 2.44) Links to other related materials or related weight information.

4. The method for constructing a virtual exhibition hall based on intelligent assisted design according to claim 3, characterized in that: In step S2, the pre-trained generative 3D model-assisted algorithm is invoked to automatically generate structural suggestions for the basic exhibition hall 3D model. Specifically... Includes the following stages: 3.1) Model training data preparation and algorithm training; 3.11) Construct a 3D scene training corpus: Collect a large number of data pairs containing text descriptions and their corresponding 3D scene models. The text descriptions are formatted descriptions of scene layout, style and function. The 3D scene models are uniformly converted into standardized formats such as voxel meshes or point clouds. 3.12) Training a text-driven generative 3D model: The model is trained using a variant algorithm based on Neural Radiation Field (NeRF) or a 3D diffusion model, enabling the model to learn the mapping relationship from text descriptions to 3D geometric structures. 3.2) Text-driven 3D structure generation; 3.21) Input preprocessing: Receive the structured design requirements from step S1 and synthesize them into specific three-dimensional spatial description prompts; 3.22) Conditional Generative Reasoning: The prompt words are input into the pre-trained generative 3D model. The text encoder of the model converts the prompt words into semantic condition vectors and guides the generation module to output a preliminary 3D structure that is aligned with the semantics of the text. The preliminary 3D structure is a low-resolution voxel representation or a sparse point cloud. 3.23) Post-processing and format conversion: Perform geometric normalization on the generated preliminary 3D structure and automatically convert it into an intermediate file format that can be recognized and edited by the target 3D modeling software; 3.3) Human-machine collaborative optimization and model iteration; 3.31) Design Collaboration: Designers review, refine, modify, and add details to the preliminary 3D structure in 3D modeling software; 3.32) Feedback learning: The final model modified by the designer and the original design requirement text are used as new training data pairs and fed back into the training corpus to perform incremental fine-tuning of the generative 3D model and achieve closed-loop optimization.

5. The method for constructing a virtual exhibition hall based on intelligent assisted design according to claim 4, characterized in that: In step S2, the optimized scene model is imported into the rendering engine, and the physically based rendering algorithm and lighting model algorithm output a high-quality panoramic image, specifically including the following steps: 4.1) Data preparation and import: The 3D scene model data, including geometric structure, optimized material parameters and optimized lighting scheme, will be automatically configured and imported into the physically based rendering engine; 4.2) Physical rendering and global illumination calculation: The physical shader and global illumination algorithm of the rendering engine are called to simulate and render the material representation, direct lighting, indirect lighting and shadows of all surfaces in the scene based on micro-surface theory, energy conservation principle and light propagation physical model. 4.3) Automatic deployment and rendering of panoramic viewpoints: Based on the location of the scene's entrance and core exhibition area, one or more panoramic rendering cameras are automatically deployed; for each camera point, the rendering engine generates equidistant rectangular projection images covering six directions: front, back, left, right, up, and down. 4.4) Panoramic Image Synthesis and Standardized Output: The generated six directional images are automatically stitched together into a complete 2:1 spherical panoramic image, and output as a high-quality panoramic image file with a specified resolution, color space and file format according to preset standards.

6. The method for constructing a virtual exhibition hall based on intelligent assisted design according to claim 5, characterized in that: In step S2, the automatic optimization of lighting and material parameters specifically involves: 5.1) Constructing a reinforcement learning optimization environment and agent: Define a reinforcement learning framework that uses a 3D rendering engine as the simulation environment, where: The status includes the encoding of current lighting and material parameters, simplified geometric features of the scene, and image features and rendering time of the previous rendering result; The action involves adjusting the continuous values ​​of light source intensity, color, position, orientation, or material reflectivity and roughness parameters; The reward function is a weighted sum of visual realism reward and rendering efficiency reward. The visual realism reward is calculated by comparing the structural similarity index or perceptual similarity index between the rendered image and the reference image. The rendering efficiency reward is negatively correlated with the rendering time. 5.2) Offline training phase: On a virtual exhibition hall scene library containing various layouts and styles, the agent is trained using proximal policy optimization or deep deterministic policy gradient algorithm, so that it learns how to adjust parameters to maximize long-term cumulative rewards through iterative interaction. 5.3) Online optimization execution phase: For the newly modeled virtual exhibition hall scene, load the trained agent and execute the following loop until the termination condition is met: 5.31) The agent adjusts its actions based on the output parameters of the current state; 5.32) The rendering engine performs fast rendering based on the adjusted parameters and provides feedback on the new status and rewards; 5.33) The agent fine-tunes its strategy in the scenario based on the reward; 5.4) Output and final rendering: Select the set of lighting and material parameters that yielded the highest reward during the optimization process, apply them to the scene, and use this set of parameters to perform a high-quality panoramic rendering output.

7. The method for constructing a virtual exhibition hall based on intelligent assisted design according to claim 6, characterized in that: In step S3, the layout optimization algorithm is a multi-objective optimization algorithm based on genetic algorithm or particle swarm optimization algorithm, specifically including the following steps: 6.1) Define the input and decision variables for the optimization problem: Input includes: 6.11) The feasible region representation S extracted from the 3D scene; 6.12) A set C of content items to be arranged, where each content item c i Includes type, preset size, and logical labels; 6.13) The association matrix R represents the strength of logical relationships between content items; 6.14) The decision variable is the layout scheme L, which is expressed as: L = {(x i , y i , θ i ) | i = 1,…,N} Among them, (x i , y i ) is the content item c i The position coordinates θ in the scene i Its orientation angle; 6.2) Constructing a multi-objective fitness function: The fitness function F_total is a weighted sum of the following four cost functions: F total = w1·F1 + w2·F2 + w3·F3 + w4·F4 in: F1 is the visual balance cost, calculated as follows: The Euclidean distance between the visual quality center and the scene geometric center; F2 is the hotspot spacing conflict cost, calculated as follows: If the distance between content items is less than the minimum comfortable spacing or greater than the maximum related spacing, a penalty is applied; F3 is the cost of content logical relatedness, calculated as follows: F3 = S i S j ( r ij · d ij ) Where r ij For the correlation strength, d ij This represents the distance between content items; F4 represents the cost of user preference deviation, which calculates the degree of deviation based on hard constraints or soft preferences specified by the user. w1, w2, w3, and w4 are user-preset weight coefficients; 6.3) Perform intelligent optimization search: Genetic algorithms or particle swarm optimization algorithms are used to achieve F total The fitness evaluation function is used to iteratively optimize the layout scheme L until the termination condition is met. 6.4) Output layout scheme and support human-computer interaction: It outputs one or more optimal layout schemes, allowing designers to select and fine-tune the schemes through an interactive interface, and automatically maps the final layout scheme to the virtual showroom platform via API to complete the content arrangement.

8. The method for constructing a virtual exhibition hall based on intelligent assisted design according to claim 7, characterized in that: In step S3, the dynamic loading and real-time rendering algorithm built into the 3D model demonstration control specifically includes the following steps: 7.1) Data preprocessing and multi-level LOD model chain construction: The original high-precision 3D model is processed offline to generate a chain of LOD models with multiple levels of detail, where each LOD level corresponds to a different number of faces and texture precision, and the LOD models are spatially aligned. For ultra-large 3D models, octrees or KD trees are used for spatial partitioning to form multiple logical blocks, each of which independently possesses the LOD model chain; 7.2) Runtime intelligent monitoring and dynamic scheduling decision-making: 7.21) Real-time collection of user viewpoint parameters, client performance parameters, and network status parameters; 7.22) Based on the viewpoint parameters, calculate the theoretically required target LOD level for each model or block in the field of view, and generate a loading request queue; 7.23) Based on network bandwidth and client performance, the request queue is prioritized and traffic controlled, wherein the priority is determined based on the distance between the block and the center of view and its contribution to the image. 7.3) Asynchronous streaming loading and progressive rendering: 7.31) Send the priority-sorted requests to the cloud server to obtain the corresponding block and LOD-level compressed model data package; 7.32) The client asynchronously receives, decodes, and buffers the data packets in the background; 7.33) The renderer retrieves available data from the cache according to priority and performs real-time drawing. When switching between high and low LOD, it uses geometric deformation or fade-in / fade-out techniques to achieve a smooth visual transition. 7.4) Real-time feedback and adaptive adjustment: Monitor the rendering frame rate. If the frame rate is lower than the target threshold, actively reduce the LOD level of the global or non-core areas, or reduce the loading requests of blocks outside the field of view. If the frame rate remains stable and resources are sufficient, the LOD level in the center of the field of view will be gradually increased to optimize visual quality.

9. A method for constructing a virtual exhibition hall based on intelligent assisted design according to claim 8, characterized in that: In step S3, the background music recommendation algorithm specifically includes: 8.1) Multi-dimensional feature extraction and representation: Extract the semantic vector V of the exhibition hall theme text text Emotional feature vector V of visual materials visual and user historical preference vector V user ; Wherein, the visual emotion feature vector V visual It is obtained by fusing color sentiment analysis, scene semantic recognition and visual sentiment prediction models; 8.2) Hybrid Recommendation Model Inference: 8.21) Initial screening based on content filtering: Constructing the joint feature vector V of the exhibition hall exhibition = α·V text + β·V visual , where α and β are adjustable weights; Calculate V exhibition With the audio feature vector V of each song in the music library music Based on cross-modal similarity, the Top-K most similar musical pieces were selected as the initial candidate set; 8.22) Reordering based on collaborative filtering: The initial candidate set, V exhibition and V user Input a ranking model, predict the adoption rate score for each piece of music, and output a ranked list of recommended background music. 8.3) Interactive Feedback and Closed-Loop Optimization: Record the designer's selection or rejection of recommended music and assign the corresponding V exhibition V music Data pairs are added to the training set; The ranking model and feature encoder are retrained periodically with new data to achieve continuous optimization of the recommendation system.

10. A virtual exhibition hall construction system based on intelligent assisted design, characterized in that: The virtual exhibition hall construction method based on intelligent assisted design as described in any one of claims 1-9 is adopted.

Images