Three-dimensional model intelligent generation and optimization method and system based on AI

Abstract

The invention discloses a three-dimensional model intelligent generation and optimization method and system based on AI. The method comprises the steps that S1, target information is acquired, and system parameters are initialized; s2, virtual and real data fusion processing is executed, and a mixed training data set is generated; s3, driving generation of a three-dimensional model based on the mixed training data set; s4, evaluating the model performance in the parallel system and generating an optimization instruction; s5, executing closed-loop iterative optimization based on the optimization instruction; and S6, outputting a final model and completing deployment. Through virtual and real data double-path fusion acquisition and generation, on one hand, a virtual data generation technology is utilized to efficiently generate massive synthetic data with accurate labels at low cost, and various long-tail scenes and extreme conditions are covered; and on the other hand, the reality sense and the physical consistency of the model are ensured by utilizing a real data acquisition technology, the defect of a single data source is effectively made up through the fusion of the two, a high-quality mixed training data set is constructed, and the generalization performance and the robustness of the generated model are enhanced.

Description

Technical Field

[0001] This application belongs to the field of 3D model generation technology, specifically involving AI-based intelligent generation and optimization methods and systems for 3D models. Background Technology

[0002] With the rapid development of fields such as virtual reality, digital twins, autonomous driving, intelligent manufacturing, and cultural creativity, the demand for 3D models, as the core digital assets for describing and constructing the physical and virtual worlds, is growing, placing higher demands on the efficiency, accuracy, realism, and scene adaptability of model generation.

[0003] Currently, the acquisition and generation of 3D models mainly rely on the following methods:

[0004] Traditional manual modeling relies on professional designers to create models manually using 3D modeling software. While the models are of high quality, they are time-consuming, labor-intensive, and costly, making it difficult to meet the needs of large-scale, personalized generation.

[0005] Modeling based on real-world data acquisition: This method uses techniques such as oblique photogrammetry and laser scanning to obtain 3D data of real-world scenes. While this approach generates highly realistic models, it is limited by equipment costs, environmental conditions (such as weather and occlusion), complex data processing, and insufficient robustness to incomplete or noisy data. Furthermore, the acquired models often lack precise semantic annotations, making them difficult to use directly for downstream AI tasks.

[0006] AI modeling based on purely virtual data generation utilizes computer graphics to generate large amounts of labeled synthetic data to drive AI model training. This method can generate arbitrary scenes and has low data annotation costs, but the "Sim2Real" gap is significant, meaning that models trained on virtual data tend to have degraded generalization performance in real-world scenarios.

[0007] Existing AI generative models, including some deep learning-based 3D generative models (such as those based on point clouds, voxels, or implicit representations), can generate 3D content from inputs such as images and text. However, they typically face the following limitations: (a) limited ability to understand multimodal and ambiguous user intent and conditional generation; (b) a lack of systematic closed-loop evaluation and optimization mechanisms for the quality and stability of the generative models, often resulting in "one-time generation, no change for life"; and (c) limited evaluation methods, usually testing only on a single dataset or in a simulation environment, failing to combine virtual extreme testing with real high-fidelity verification for comprehensive diagnosis.

[0008] In summary, existing technologies generally suffer from problems such as a single data source, a disconnect between the generation process and evaluation / optimization, and a lack of adaptive iterative capabilities. This makes it difficult to achieve an optimal balance between efficiency, accuracy, robustness, and scene adaptability in the generated 3D models. Therefore, there is an urgent need for an automated method and system that can integrate multi-source data, achieve intelligent generation, and possess self-evaluation and continuous optimization capabilities. Summary of the Invention

[0009] This application provides an AI-based intelligent generation and optimization method and system for 3D models, aiming to solve the problems existing in the prior art.

[0010] Firstly, an AI-based intelligent generation and optimization method for 3D models, the method comprising:

[0011] S1: Obtain target information and initialize system parameters. Input target description information and optional initial conditions. Parse and process the target description information to generate quantifiable technical parameters and a task list. Based on the task list, call the corresponding initial data acquisition strategy, model generation strategy and performance evaluation strategy from the preset strategy rule library to complete the system initialization configuration.

[0012] S2: Perform virtual and real data fusion processing to generate a hybrid training dataset;

[0013] S3: Based on the hybrid training dataset, drive the generation of the 3D model by inputting the hybrid training dataset generated in S2 into the conditional 3D generation model;

[0014] S4: Evaluate the model performance and generate optimization instructions in the parallel system. Input the initial 3D model generated in S3 into the parallel evaluation subsystem. Evaluate the performance of the initial 3D model in the virtual test environment and the real verification environment, identify its performance gap and locate the weak link, and generate optimization instructions based on the location result.

[0015] S5: Execute closed-loop iterative optimization based on optimization instructions. The system decision module receives the optimization instructions generated in step S4 and determines whether iterative optimization is triggered based on preset performance thresholds and cost rules. If it is determined to be triggered, the data fusion acquisition step and the three-dimensional model generation step are dynamically scheduled to perform targeted optimization, forming a closed-loop iterative process until the output three-dimensional model meets the preset conditions.

[0016] S6: Output the final model and complete the deployment. When the termination condition of step S5 is met, perform format conversion and lightweight processing on the final 3D model and output it in the specified application format.

[0017] Optionally, step S2 includes the following steps:

[0018] Generate precisely labeled virtual 3D data by at least one of the following methods: parametric scene simulation, structured 2D upscaling, or image depth backprojection.

[0019] Obtain real 3D data through oblique photogrammetry, active 3D scanning, or by calling at least one of the existing model libraries;

[0020] The generated virtual 3D data is spatially registered with the acquired real 3D data, and foreground and background are synthesized and consistency is optimized to generate the hybrid training dataset.

[0021] Optionally, S3 includes the following steps:

[0022] Multimodal feature encoding is performed on the input point cloud, grid, image data, and user condition information;

[0023] The encoded multimodal features are aligned and fused using a cross-attention mechanism to obtain a unified multimodal feature representation.

[0024] Based on the unified multimodal feature representation, 3D model data is generated through a 3D generative network;

[0025] Convert the generated 3D model data into a specified general 3D format and output it.

[0026] Optionally, S4 includes the following steps:

[0027] The downstream task performance metrics of the initial 3D model are evaluated in a series of parameterized virtual test scenarios.

[0028] The initial 3D model is compared with high-precision verification data of the real scene to calculate its geometric accuracy and texture quality index.

[0029] Correlation analysis of the evaluation results of virtual testing and real verification can pinpoint the specific weak links that cause performance degradation;

[0030] Based on the pre-set policy rule base and the characteristics of the weak links, optimization suggestions are generated, including data augmentation instructions, model adjustment instructions, or training strategy instructions.

[0031] Optionally, S5 includes the following steps:

[0032] The initial 3D model is controlled to perform large-scale automated calculation experiments in a virtual artificial system, and simultaneously a small-scale high-fidelity verification is performed in a real physical system.

[0033] Collect and integrate performance data, evaluation reports, and external feedback from parallel experiments to generate a comprehensive system status report;

[0034] Analyze the system status panoramic report, make decisions and generate scheduling instructions to drive the data fusion and acquisition steps to generate specific enhanced data, and drive the 3D model generation steps to initiate targeted model optimization;

[0035] Monitor the performance of the optimized model and decide whether to start a new round of iterations based on the new evaluation feedback, until the iteration termination conditions are met.

[0036] Secondly, an AI-based intelligent generation and optimization system for 3D models, the system comprising:

[0037] The dual-channel data fusion acquisition and generation module is used to generate virtual 3D data and acquire real 3D data in parallel, and then merge the two into a hybrid training dataset.

[0038] The 3D model intelligent generation module is connected to the dual-channel data fusion acquisition and generation module. It is used to receive the hybrid training dataset and user condition input, and output the initial 3D model through a multimodal fusion deep generation network.

[0039] The model optimization and adaptive evaluation module is connected to the intelligent generation module of the 3D model and is used to evaluate the performance of the initial 3D model in a virtual test environment and a real verification environment, diagnose its weaknesses and generate optimization suggestions.

[0040] The parallel execution and feedback closed-loop management module is connected to the dual-channel data fusion acquisition and generation module, the three-dimensional model intelligent generation module, and the model optimization and adaptive evaluation module, respectively. It is used to deploy and test the initial three-dimensional model in parallel in the virtual artificial system and the real physical system, aggregate feedback information, and dynamically schedule each module of the system to perform closed-loop iterative optimization according to the optimization suggestions.

[0041] Optionally, the dual-channel data fusion acquisition and generation module includes:

[0042] The virtual data generation unit generates precisely labeled virtual 3D data using a software-defined method, based on task requirements.

[0043] The real data acquisition unit acquires 3D data of real scenes from the physical world;

[0044] The virtual-real fusion processing unit is connected to the virtual data generation unit and the real data acquisition unit, and is used to generate mixed reality training data.

[0045] Optionally, the intelligent 3D model generation module includes:

[0046] A multimodal data encoder is used to extract and align features from input data.

[0047] A multimodal feature fusion unit, connected to the multimodal data encoder, aligns and fuses the extracted multimodal features through a cross-attention mechanism;

[0048] A 3D model generator is connected to the multimodal feature fusion unit and generates 3D model data based on the fused features.

[0049] An output adapter, connected to the 3D model generator, is used to convert the generated 3D model data into a specified output format.

[0050] Compared with the prior art, this application has at least the following beneficial effects:

[0051] This application utilizes a dual-path fusion of virtual and real data acquisition and generation. On the one hand, it leverages virtual data generation technology to efficiently and cost-effectively generate massive amounts of precisely labeled synthetic data, covering various long-tail scenarios and extreme conditions. On the other hand, it employs real data acquisition technology to ensure the realism and physical consistency of the model. The fusion of these two approaches effectively compensates for the deficiencies of a single data source, constructs a high-quality hybrid training dataset, and enhances the generalization performance and robustness of the generated model.

[0052] This application also introduces a parallel system evaluation mechanism to conduct large-scale, automated, and parameterized extreme scenario tests in a virtual artificial system, while simultaneously conducting small-scale high-fidelity verification in a real physical system. This enables a comprehensive and three-dimensional diagnosis of the model's performance under different environments and tasks, and accurately identifies specific weaknesses in geometry, texture, structure, and downstream task levels that are difficult to detect using traditional single evaluation methods.

[0053] Based on the evaluation and diagnosis results, the system can automatically generate targeted optimization instructions, and the decision module can dynamically schedule the data acquisition and model generation steps for targeted optimization. This closed-loop feedback mechanism enables the 3D model to continuously evolve in the iterative cycle of generation-evaluation-optimization until it meets the preset performance requirements, thereby improving the quality and adaptability of the final output model and reducing the cost of manual intervention. Attached Figure Description

[0054] Figure 1 A flowchart illustrating an AI-based intelligent generation and optimization method for 3D models provided in one embodiment of this application;

[0055] Figure 2 This is a schematic diagram of the module connections of an AI-based intelligent generation and optimization system for 3D models provided in one embodiment of this application. Detailed Implementation

[0056] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figure 1-2 The present application will be further described in detail with reference to the embodiments.

[0057] The AI-based intelligent generation and optimization method for 3D models provided in this application includes the following steps:

[0058] S1: Obtain target information and initialize system parameters. Input target description information and optional initial conditions. Parse and process the target description information to generate quantifiable technical parameters and a task list. Based on the task list, call the corresponding initial data acquisition strategy, model generation strategy and performance evaluation strategy from the preset strategy rule library to complete the system initialization configuration.

[0059] S2: Perform virtual-real data fusion processing to generate a hybrid training dataset. Based on the initial data acquisition strategy determined in S1, perform the following operations in parallel:

[0060] Real-world data acquisition and preprocessing: Based on target area information, plan and control data acquisition equipment to execute acquisition tasks and obtain real-world scene data; perform automated preprocessing on the real-world scene data to obtain first three-dimensional data;

[0061] Virtual data generation and rendering: Based on the target area information and technical parameters, a corresponding virtual scene is constructed in a virtual simulation environment; according to the task list, key virtual objects are configured and rendered in the virtual scene; and a second three-dimensional data with precise annotations is generated by simulating the real acquisition process through parameterized virtual sensors.

[0062] The virtual and real data fusion process involves spatially registering the first 3D data with the second 3D data; synthesizing the virtual objects in the second 3D data with the real background in the first 3D data; and making consistent adjustments to lighting, occlusion, and scale to generate hybrid 3D data and corresponding component labels, thus forming a hybrid training dataset.

[0063] S3: Based on a hybrid training dataset, a 3D model is generated. The hybrid training dataset generated in S2 is input into the conditional 3D generation model. If the task list corresponds to a general task, the pre-trained generation model is called for fine-tuning; if it corresponds to a customized task, the training of the multimodal fusion generation network is initiated. Through the processing of the generation model, an initial 3D model for the target region is output.

[0064] S4: Evaluate model performance and generate optimization instructions in a parallel system. Input the initial 3D model generated in S3 into the parallel evaluation subsystem, which includes a virtual artificial system and a real physical system.

[0065] In the virtual artificial intelligence system, various preset extreme test scenarios are constructed, the initial 3D model is imported and its performance in downstream tasks is tested, performance data is recorded and failure cases are located.

[0066] In a real physical system, high-precision verification data is collected over a small area. The initial 3D model is then quantitatively compared and analyzed with the high-precision verification data to calculate geometric and texture error indices.

[0067] By comprehensively analyzing the evaluation results from the virtual artificial system and the real physical system, the root cause of the model's insufficient performance is determined through the defect diagnosis module. Based on the root cause, specific optimization instructions are generated, which include at least the enhanced data generation instructions for the data preparation stage and / or the network parameter adjustment instructions for the model generation stage.

[0068] S5: Execute closed-loop iterative optimization based on the optimization instructions. The system decision module receives the optimization instructions generated in step S4 and determines whether iterative optimization is triggered based on preset performance thresholds and cost rules. If it is determined to be triggered, then execute:

[0069] The enhanced data generation instruction is sent to the virtual and real data generation and fusion module of S2, driving it to generate supplementary data that meets the requirements of the instruction.

[0070] The network parameter adjustment command is sent to the model generation module of S3, driving it to optimize the initial 3D model through incremental learning or targeted fine-tuning.

[0071] The optimized 3D model is then sent back to S4 for evaluation, forming a closed loop of "evaluation-diagnosis-optimization" until the model performance meets the preset requirements or reaches the upper limit of the number of iterations.

[0072] S6: Output the final model and complete the deployment. When the termination condition of step S5 is met, perform format conversion and lightweight processing on the final 3D model and output it in the specified application format. At the same time, the system archives the entire process configuration, strategy and results of this task to the experience knowledge base for optimizing the initialization strategy of subsequent tasks.

[0073] In one embodiment, an AI-based intelligent generation and optimization system for 3D models is provided, comprising:

[0074] The dual-channel data fusion acquisition and generation module is used to execute S1 and S2;

[0075] Specifically, the dual-path data fusion acquisition and generation module includes a virtual data generation unit, a real data acquisition unit, and a virtual-real fusion processing unit, which are used to generate and acquire multi-source heterogeneous 3D data in parallel, and output a hybrid training dataset after fusion.

[0076] The virtual data generation unit, used to generate precisely labeled virtual 3D data through software definition according to task requirements, includes:

[0077] The parametric scene simulation engine, based on a pre-built 3D model library and scene rule library, responds to system commands and automatically configures the environmental parameters, object layout, and material properties of the virtual scene; through the built-in parametric virtual sensor model, it simulates the sampling characteristics of real physical sensors, scans or renders the virtual scene, and outputs virtual point cloud data or mesh models with corresponding truth labels.

[0078] The structured 2D upscaling engine receives input 2D planar images, sketches, or a limited number of scene images; uses a trained deep convolutional neural network to identify structural feature points and topological relationships in the input; connects the feature points into a set of line segments representing the boundaries of the 3D structure according to preset structured rules; and generates a regular 3D structural model based on the set of line segments through geometric extrusion and construction methods.

[0079] The image depth back-projection engine receives monocular or binocular image input; calculates the depth information of each pixel in the image using a depth estimation model; and back-projects the image pixels and their depth values ​​into three-dimensional space based on the camera's intrinsic and extrinsic parameter matrices to generate pseudo-point cloud data.

[0080] The real data acquisition unit acquires 3D data of real scenes from the physical world, including:

[0081] The oblique photogrammetry subunit controls the multi-view camera array carried by the UAV to take pictures of the target area according to the predetermined route; through aerial triangulation and dense matching processing of multi-view images, a real-scene 3D mesh model is generated.

[0082] The active 3D scanning subunit transmits and receives optical signals through LiDAR or depth camera to directly acquire high-precision 3D point cloud data of the target scene;

[0083] Existing data interfaces are used to access and preprocess existing third-party 3D model library data;

[0084] The virtual-real fusion processing unit, connected to the virtual data generation unit and the real data acquisition unit, is used to generate mixed reality training data and performs the following steps:

[0085] Data alignment involves spatially registering the virtual object data output by the virtual data generation unit with the real scene data output by the real data acquisition unit in the same coordinate system.

[0086] Foreground-background compositing involves embedding virtual objects as the foreground into a background composed of real-world scene data to generate a preliminary composite scene.

[0087] Consistency optimization processing involves intelligent processing of the composite scene, including: simulating real occlusion relationships based on ray detection and depth comparison, erasing or making transparent the parts of the virtual foreground that are occluded by the real background; dynamically adjusting the texture brightness, hue, and shadow of the virtual foreground according to the lighting estimation information of the real background to make it consistent with the background lighting conditions; and fine-tuning the scale of the virtual foreground to ensure that its proportion is coordinated with the real scene.

[0088] The dataset is generated, and the composite 3D data that has undergone consistency optimization is output as mixed training samples. At the same time, auxiliary labels that identify the virtual and real components are generated.

[0089] The intelligent 3D model generation module is connected to the dual-channel data fusion acquisition and generation module and is used to execute S3.

[0090] Specifically, the 3D model intelligent generation module is a deep learning-based generation module used to receive and process diverse input data from the data fusion module and generate high-quality 3D models. The specific structure and workflow of the 3D model intelligent generation module include:

[0091] A multimodal data encoder is used for feature extraction and alignment of input data. This encoder includes multiple dedicated coding subnetworks in parallel:

[0092] The point cloud encoding subnetwork receives a point cloud sequence as input, extracts features through hierarchical features, and outputs a feature representation of the point cloud.

[0093] The mesh coding subnetwork receives 3D mesh data input, extracts its geometric and topological features through graph convolution operations, and outputs the feature representation of the mesh.

[0094] The image coding subnetwork receives a two-dimensional image input, extracts its visual features through a convolutional neural network, and outputs a feature representation of the image.

[0095] A condition encoder receives condition information provided by a user, the condition information including at least one of text description, two-dimensional bounding box, or low-precision reference model; and encodes the condition information to generate a condition feature vector.

[0096] A multimodal feature fusion unit is connected to the multimodal data encoder. This unit calculates the correlation weights between different modal features by introducing a cross-attention mechanism, thereby achieving adaptive alignment and information complementarity between different modal features and outputting a unified, fused multimodal feature representation.

[0097] A 3D model generator is connected to the multimodal feature fusion unit. This generator takes the unified multimodal feature representation as input and generates corresponding 3D model data through a deconvolutional network, a coordinate generation network, or a differentiable renderer. During training, the generator's learning objective is subject to multiple constraints, including: geometric similarity constraints between the generated model and real data, realistic constraints on surface texture, and reasonable constraints on physical structure.

[0098] An output adapter, connected to the 3D model generator, is used to convert the internal data format output by the generator into a specified general 3D format and output it, the format including at least one of point cloud, polygon mesh or voxel mesh;

[0099] The model optimization and adaptive evaluation module is connected to the three-dimensional model intelligent generation module and is used to execute S4;

[0100] Specifically, the model optimization and adaptive evaluation module is used to systematically evaluate the generated 3D model, diagnose defects, and generate optimization strategies. It includes a performance evaluation unit, a gap analysis unit, and an optimization instruction generation unit.

[0101] The performance evaluation unit, used to quantitatively evaluate the input 3D model in both virtual and real environments, includes:

[0102] The virtual environment testing platform has a built-in parametric scene generator that can automatically generate a series of virtual test scenes covering different weather conditions, light intensity, observation angle and noise level according to predefined parameter combinations. The 3D model to be evaluated is placed into the test scene as input for downstream tasks (such as object detection and semantic segmentation) and its task performance indicators are recorded.

[0103] The real-world verification interface receives high-precision verification data collected from a small area in the real world; it performs spatial registration and comparison between the 3D model to be evaluated and the high-precision verification data, and calculates its geometric accuracy index, texture quality index and structural integrity index.

[0104] The comprehensive evaluation report generator summarizes and formats the quantitative results from the virtual environment testing platform and the real environment verification interface to generate a structured comprehensive evaluation report;

[0105] The gap analysis unit, connected to the performance evaluation unit, is used to diagnose the root causes of model defects. This unit performs the following steps:

[0106] Data association involves mapping the performance of the same model in the virtual test scenario in the comprehensive evaluation report to its error performance on real verification data.

[0107] Weak link identification involves analyzing the correlation between performance degradation and specific error patterns, and combining this with visualization analysis of the model's internal features to pinpoint the specific weak links that cause performance degradation. These weak links include, but are not limited to: insufficient perception of targets within a specific distance range, failure to reconstruct sparse point cloud data, or distortion in texture generation for specific material surfaces.

[0108] Generate a diagnostic report, outputting a report that includes the identified weaknesses and their associated environmental or data characteristics;

[0109] An optimization instruction generation unit, connected to the gap analysis unit, is used to automatically generate executable optimization instructions based on the diagnostic conclusions. This unit includes a pre-configured optimization strategy rule base, which matches and triggers corresponding optimization strategies based on the characteristics of the weak links in the diagnostic report, generating an instruction set including at least one of the following:

[0110] The data augmentation command is directed to the data fusion acquisition and generation module. The command content includes the precise specifications of the virtual data to be generated, including target type, spatial distribution, environmental conditions, and data volume.

[0111] The model adjustment command points to the 3D model intelligent generation module. The command content includes suggestions for adjusting the weights of the network loss function, modification schemes for specific network layer structures, or commands to start the domain adaptive training process.

[0112] Evaluation focus adjustment instructions: These are directed to the performance evaluation unit and are used to dynamically adjust the distribution of test scenarios or the weight of verification metrics in subsequent evaluation cycles, in order to focus on the identified weaknesses.

[0113] The parallel execution and feedback closed-loop management module is connected to the dual-channel data fusion acquisition and generation module, the three-dimensional model intelligent generation module, and the model optimization and adaptive evaluation module, respectively, and is used to execute parallel execution and S5;

[0114] Specifically, the parallel execution and feedback closed-loop management module is used to coordinate the execution of each module and realize closed-loop optimization based on the ACP (Artificial System, Computational Experiment, Parallel Execution) theory. This module specifically includes a parallel execution controller, a feedback information aggregator, and an adaptive strategy scheduler.

[0115] The parallel execution controller is used for the collaborative deployment and experimentation of 3D models in both virtual artificial systems and real physical systems. It performs the following operations:

[0116] Virtual system deployment and experimentation: The 3D model to be evaluated is instantiated in a virtual simulation environment, and a parametric scene generator is driven to build a set of test scenarios that cover a wide range of edge cases in batches; large-scale, automated computational experiments are run in the virtual system, and model performance data and abnormal event records are collected;

[0117] Real system deployment and verification: Deploy the same 3D model on a small scale in a controlled real physical environment, collect its operating data through real sensors, and perform high-fidelity verification and performance recording;

[0118] Execution status management: Monitor and synchronize the experimental processes of the two parallel systems mentioned above, and manage the start and stop of experimental tasks and resource allocation;

[0119] A feedback information aggregator, connected to the parallel execution controller and the model optimization and adaptive evaluation module respectively, is used to collect, clean, and fuse feedback information from multiple sources. The aggregated information includes:

[0120] Raw performance logs and exception event reports from the virtual and physical systems from the parallel execution controller;

[0121] Quantitative evaluation reports and diagnostic conclusions from the model optimization and adaptive evaluation module;

[0122] External human corrections or annotations;

[0123] The aggregator performs time alignment, format standardization, and correlation analysis on the aforementioned multi-source and heterogeneous feedback information to generate a unified system status overview report.

[0124] The adaptive policy scheduler, connected to the feedback information aggregator, serves as the core of closed-loop decision-making, and its workflow is as follows:

[0125] Status analysis and decision-making involves parsing the system status overview report, identifying systemic defects or performance bottlenecks in the current 3D model, and making decisions based on predefined optimization objectives and resource constraints.

[0126] Generate and issue scheduling instructions. Based on the decision results, generate specific, executable sets of scheduling instructions and distribute them to other functional modules of the system. For example:

[0127] If the analysis shows that the model's performance index is below the threshold under "heavy rain" conditions, an instruction is sent to the dual-channel data fusion acquisition and generation module, requiring it to generate a specified number of virtual augmented data containing heavy rain meteorological conditions.

[0128] Simultaneously, an instruction is sent to the intelligent generation module of the 3D model to initiate a targeted incremental training or fine-tuning process with the enhanced data as input;

[0129] Closed-loop iterative management: Monitor the execution status of scheduling instructions, and after a new round of model optimization and evaluation is completed, trigger the re-analysis and decision-making of feedback information, thereby forming a closed-loop iterative process of evaluation-decision-optimization-re-evaluation until the system's preset termination conditions are met.

[0130] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. An AI-based intelligent generation and optimization method for 3D models, characterized in that, The method includes: S1: Obtain target information and initialize system parameters. Input target description information and optional initial conditions. Parse and process the target description information to generate quantifiable technical parameters and a task list. Based on the task list, call the corresponding initial data acquisition strategy, model generation strategy and performance evaluation strategy from the preset strategy rule library to complete the system initialization configuration. S2: Perform virtual and real data fusion processing to generate a hybrid training dataset; S3: Based on the hybrid training dataset, drive the generation of the 3D model by inputting the hybrid training dataset generated in S2 into the conditional 3D generation model; S4: Evaluate the model performance and generate optimization instructions in the parallel system. Input the initial 3D model generated in S3 into the parallel evaluation subsystem. Evaluate the performance of the initial 3D model in the virtual test environment and the real verification environment, identify its performance gap and locate the weak link, and generate optimization instructions based on the location result. S5: Execute closed-loop iterative optimization based on optimization instructions. The system decision module receives the optimization instructions generated in step S4 and determines whether iterative optimization is triggered based on preset performance thresholds and cost rules. If it is determined to be triggered, the data fusion acquisition step and the three-dimensional model generation step are dynamically scheduled to perform targeted optimization, forming a closed-loop iterative process until the output three-dimensional model meets the preset conditions. S6: Output the final model and complete the deployment. When the termination condition of step S5 is met, perform format conversion and lightweight processing on the final 3D model and output it in the specified application format.

2. The AI-based intelligent generation and optimization method for 3D models according to claim 1, characterized in that, S2 includes the following steps: Generate precisely labeled virtual 3D data by at least one of the following methods: parametric scene simulation, structured 2D upscaling, or image depth backprojection. Obtain real 3D data through oblique photogrammetry, active 3D scanning, or by calling at least one of the existing model libraries; The generated virtual 3D data is spatially registered with the acquired real 3D data, and foreground and background are synthesized and consistency is optimized to generate the hybrid training dataset.

3. The AI-based intelligent generation and optimization method for 3D models according to claim 1, characterized in that, S3 includes the following steps: Multimodal feature encoding is performed on the input point cloud, mesh, image data, and user condition information; The encoded multimodal features are aligned and fused using a cross-attention mechanism to obtain a unified multimodal feature representation. Based on the unified multimodal feature representation, 3D model data is generated through a 3D generative network; Convert the generated 3D model data into the specified general 3D format and output it.

4. The AI-based intelligent generation and optimization method for 3D models according to claim 1, characterized in that, S4 includes the following steps: The downstream task performance metrics of the initial 3D model are evaluated in a series of parameterized virtual test scenarios. The initial 3D model is compared with high-precision verification data of the real scene to calculate its geometric accuracy and texture quality index. Correlation analysis of the evaluation results of virtual testing and real verification can pinpoint the specific weak links that cause performance degradation; Based on the pre-set policy rule base and the characteristics of the weak links, optimization suggestions are generated, including data augmentation instructions, model adjustment instructions, or training strategy instructions.

5. The AI-based intelligent generation and optimization method for 3D models according to claim 1, characterized in that, S5 includes the following steps: The initial 3D model is controlled to perform large-scale automated calculation experiments in a virtual artificial system, and simultaneously a small-scale high-fidelity verification is performed in a real physical system. Collect and integrate performance data, evaluation reports, and external feedback from parallel experiments to generate a comprehensive system status report; Analyze the system status panoramic report, make decisions and generate scheduling instructions to drive the data fusion and acquisition steps to generate specific enhanced data, and drive the 3D model generation steps to initiate targeted model optimization; Monitor the performance of the optimized model and decide whether to start a new round of iterations based on the new evaluation feedback, until the iteration termination conditions are met.

6. An AI-based intelligent generation and optimization system for 3D models, characterized in that, The system for implementing the AI-based intelligent generation and optimization method for 3D models as described in any one of claims 1 to 5, the system comprising: The dual-channel data fusion acquisition and generation module is used to generate virtual 3D data and acquire real 3D data in parallel, and then merge the two into a hybrid training dataset. The 3D model intelligent generation module is connected to the dual-channel data fusion acquisition and generation module. It is used to receive the hybrid training dataset and user condition input, and output the initial 3D model through a multimodal fusion deep generation network. The model optimization and adaptive evaluation module is connected to the intelligent generation module of the 3D model and is used to evaluate the performance of the initial 3D model in a virtual test environment and a real verification environment, diagnose its weaknesses and generate optimization suggestions. The parallel execution and feedback closed-loop management module is connected to the dual-channel data fusion acquisition and generation module, the three-dimensional model intelligent generation module, and the model optimization and adaptive evaluation module, respectively. It is used to deploy and test the initial three-dimensional model in parallel in the virtual artificial system and the real physical system, aggregate feedback information, and dynamically schedule each module of the system to perform closed-loop iterative optimization according to the optimization suggestions.

7. The AI-based intelligent generation and optimization system for 3D models according to claim 6, characterized in that, The dual-channel data fusion acquisition and generation module includes: The virtual data generation unit generates precisely labeled virtual 3D data using a software-defined method, based on task requirements. The real data acquisition unit acquires 3D data of real scenes from the physical world; The virtual-real fusion processing unit is connected to the virtual data generation unit and the real data acquisition unit, and is used to generate mixed reality training data.

8. The AI-based intelligent generation and optimization system for 3D models according to claim 6, characterized in that, The intelligent 3D model generation module includes: A multimodal data encoder is used for feature extraction and alignment of input data; A multimodal feature fusion unit, connected to the multimodal data encoder, aligns and fuses the extracted multimodal features through a cross-attention mechanism; A 3D model generator is connected to the multimodal feature fusion unit and generates 3D model data based on the fused features. An output adapter, connected to the 3D model generator, is used to convert the generated 3D model data into a specified output format.

Images