A method, system and medium for adjusting a three-dimensional model

Abstract

This invention relates to the field of 3D model adjustment technology, and more particularly to a method, system, and medium for 3D model adjustment. The method includes the following steps: acquiring multi-angle images of the original model and performing multi-view normalization to generate a normalized view image set to extract feature points of the original model; performing parallax correction on these feature points to reconstruct a simulated 3D model; collecting basic usage data of the model; mapping these data to ideal requirements to determine the ideal 3D model structure; dividing the reconstructed model into core regions based on the basic usage data to obtain key region model slices; comparing these slices with the ideal model structure by highlighting the key regions; analyzing model differences; determining the structural adjustment range based on the comparison results; and performing cyclical fine-tuning and correction on the key region model slices until they match the ideal 3D model structure, thereby generating an optimized 3D model. This invention achieves efficient and accurate 3D model adjustment and optimization.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional model adjustment technology, and in particular to a three-dimensional model adjustment method, system and medium. Background Technology

[0002] Traditional model adjustment methods often rely on manual operation and experience-based judgment, resulting in insufficient model accuracy and low efficiency. When acquiring multi-angle images, existing technologies often fail to fully capture model details, and the normalization process is not accurate enough, affecting feature point extraction and subsequent reconstruction. The parallax problem in the model reconstruction process is often not effectively resolved, leading to discrepancies between the final model and the actual object, which affects the reliability of subsequent applications. In addition, the basic data acquisition and ideal requirement mapping process of the model lacks systematicity, the model design fails to closely revolve around the actual application requirements, and the determination of the ideal model structure often relies on subjective judgment, resulting in unclear design goals. The segmentation and difference comparison of key areas are complex, making it impossible to quickly identify and correct model defects, which affects the efficiency and effectiveness of the overall design. Existing methods often fail to achieve fine-grained optimization of details when dealing with complex models. Summary of the Invention

[0003] Therefore, it is necessary to provide a three-dimensional model adjustment method, system, and medium to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a three-dimensional model adjustment method includes the following steps:

[0005] Step S1: Obtain multi-angle images of the original model; perform multi-view normalization on the multi-angle images of the original model to obtain a normalized view image set; extract feature points of the original model based on the normalized view image set;

[0006] Step S2: Perform disparity correction on the feature points of the original model to obtain the corrected model feature points; reconstruct the simulation 3D model based on the corrected model feature points;

[0007] Step S3: Collect basic model usage data; map ideal requirements using the basic model usage data to generate ideal requirement data; determine the ideal 3D model structure using the ideal requirement data;

[0008] Step S4: Based on the basic application data of the model, the core region of the reconstructed simulation 3D model is segmented to obtain key region model slices; the highlighted regions are compared with the key region model slices based on the ideal 3D model structure, and the direction of model differences is analyzed based on the comparison results of the highlighted regions;

[0009] Step S5: Determine the structural adjustment range based on the comparison results of the highlighted areas; perform interval cyclic fine-tuning and correction on the key area model slices according to the structural adjustment range and the direction of model differences, until the key area model slices are consistent with the ideal 3D model structure, thereby generating an adjusted and optimized 3D model.

[0010] This invention acquires multi-angle images of the original model, enabling the system to comprehensively capture every detail of the model and ensuring the richness and accuracy of the data. The generation of a normalized viewpoint image set makes subsequent feature extraction more efficient and accurate. The extracted feature points provide a solid foundation for model reconstruction. The application of disparity correction technology effectively eliminates disparity errors in the images, ensuring the accuracy of feature points and thus improving the quality and credibility of the simulated 3D model. The 3D model reconstructed based on corrected model feature points can realistically reflect the geometric features of the original model. The process of collecting basic application data ensures the goal-oriented nature of the model design. The implementation of ideal requirement mapping allows the design process to closely revolve around practical application needs. The generated ideal requirement data provides a clear basis for determining the ideal 3D model structure. The segmentation of the core region makes the analysis of key model slices more focused, effectively improving the efficiency and accuracy of subsequent comparisons. By comparing the ideal 3D model structure with the highlighted areas of the key region model slices, the differences and deficiencies of the model can be quickly identified. Analyzing the direction of model differences further clarifies the specific areas that need optimization. The structural adjustment range determined based on the highlighted area comparison results provides a scientific basis for fine-tuning. Combined with interval cyclical fine-tuning based on the direction of model differences, the consistency between the key region model slices and the ideal 3D model structure is ensured. The generated adjusted and optimized 3D model not only improves the overall performance of the model, but also provides greater adaptability for subsequent applications. The entire system effectively integrates image processing, data analysis, and model optimization technologies, significantly improving the efficiency and accuracy of 3D model adjustment, providing strong technical support for various application scenarios, promoting the further development of 3D modeling technology, enhancing the functionality and reliability of models in practical applications, and ultimately realizing the intelligent and automated design and adjustment of 3D models.

[0011] The present invention also provides a three-dimensional model adjustment system for performing the three-dimensional model adjustment method described above, the three-dimensional model adjustment system comprising:

[0012] The image acquisition module is used to acquire multi-angle images of the original model; perform multi-view normalization on the multi-angle images of the original model to obtain a normalized view image set; and extract feature points of the original model based on the normalized view image set.

[0013] The model reconstruction module is used to perform disparity correction on the feature points of the original model, thereby obtaining the corrected model feature points; and to reconstruct the simulation 3D model based on the corrected model feature points.

[0014] The requirement mapping module is used to collect basic model usage data; to perform ideal requirement mapping based on the basic model usage data to generate ideal requirement data; and to determine the ideal 3D model structure based on the ideal requirement data.

[0015] The difference comparison module is used to segment the core region of the reconstructed simulation 3D model based on the model's basic application data, thereby obtaining key region model slices; it compares the highlighted regions with the ideal 3D model structure and the key region model slices, and analyzes the direction of model differences based on the highlighted region comparison results;

[0016] The cyclic adjustment module is used to determine the range of structural adjustment based on the comparison results of the highlighted areas; according to the range of structural adjustment and the direction of model differences, the key area model slices are cyclically fine-tuned and corrected until the key area model slices are consistent with the ideal 3D model structure, thereby generating an adjusted and optimized 3D model.

[0017] This invention, through the application of an image acquisition module, enables the system to comprehensively collect multi-angle images of the original model, ensuring data diversity and integrity. The generation of a normalized viewpoint image set improves the accuracy of subsequent feature point extraction. The extracted feature points of the original model provide a solid foundation for subsequent model reconstruction. The implementation of parallax correction technology effectively eliminates errors in the images, ensuring the accuracy of feature points, thereby improving the quality and realism of the reconstructed 3D model. The reconstructed simulation 3D model can realistically reflect the geometric features of the original model. The requirement mapping module, by collecting basic model usage data, ensures the clarity of design objectives. The realization of ideal requirement mapping allows the design process to closely revolve around practical application requirements. The generated ideal requirement data provides a clear basis for determining the ideal 3D model structure. The segmentation of the core region makes the analysis of key region model slices more focused, improving the efficiency and accuracy of subsequent comparisons. The difference comparison module compares the ideal 3D model structure with the key region model slices in highlighted areas. Yes, it can quickly identify model differences and deficiencies, analyze the direction of model differences to clarify the specific areas for optimization, and the cyclic adjustment module determines the structural adjustment range based on the comparison results of highlighted areas, providing a scientific basis for fine-tuning. Combined with the cyclic fine-tuning of the model difference direction, it ensures the consistency of key area model slices with the ideal 3D model structure. The generated adjusted and optimized 3D model not only improves the overall performance of the model, but also provides greater adaptability for subsequent applications. The entire system effectively integrates image processing, data analysis, and model optimization technologies, significantly improving the efficiency and accuracy of 3D model adjustment, providing strong technical support for various application scenarios, promoting the further development of 3D modeling technology, enhancing the functionality and reliability of models in practical applications, and ultimately realizing the intelligent and automated design and adjustment of 3D models. This promotes technological progress and innovation in related industries, ensures the efficient application of models in complex and ever-changing environments, and enhances product competitiveness and market responsiveness.

[0018] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed, implements the three-dimensional model adjustment method as described in any of the above claims.

[0019] This invention provides a highly efficient 3D model adjustment capability through a computer program stored on a computer-readable storage medium, ensuring the accuracy and consistency of model data. It integrates image acquisition and processing technologies, improving the accuracy of model feature extraction, optimizing the reconstruction process, and enhancing the realism of the simulated 3D model. Based on the requirement mapping function, it clarifies the design objectives, ensuring that the model structure meets the actual application requirements. It uses difference comparison technology to quickly identify model defects, promoting targeted optimization. The cyclic adjustment function enables fine-grained correction of model details, enhancing the model's adaptability and performance. The digital and automated characteristics of the entire system promote the improvement of design efficiency and the rational use of resources. Attached Figure Description

[0020] Figure 1 A flowchart illustrating the steps of a three-dimensional model adjustment method;

[0021] Figure 2 This is a detailed flowchart illustrating the implementation steps of step S3;

[0022] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0023] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0024] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0025] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0026] To achieve the above objectives, please refer to Figures 1 to 2 A method for adjusting a three-dimensional model includes the following steps:

[0027] Step S1: Obtain multi-angle images of the original model; perform multi-view normalization on the multi-angle images of the original model to obtain a normalized view image set; extract feature points of the original model based on the normalized view image set;

[0028] Step S2: Perform disparity correction on the feature points of the original model to obtain the corrected model feature points; reconstruct the simulation 3D model based on the corrected model feature points;

[0029] Step S3: Collect basic model usage data; map ideal requirements using the basic model usage data to generate ideal requirement data; determine the ideal 3D model structure using the ideal requirement data;

[0030] Step S4: Based on the basic application data of the model, the core region of the reconstructed simulation 3D model is segmented to obtain key region model slices; the highlighted regions are compared with the key region model slices based on the ideal 3D model structure, and the direction of model differences is analyzed based on the comparison results of the highlighted regions;

[0031] Step S5: Determine the structural adjustment range based on the comparison results of the highlighted areas; perform interval cyclic fine-tuning and correction on the key area model slices according to the structural adjustment range and the direction of model differences, until the key area model slices are consistent with the ideal 3D model structure, thereby generating an adjusted and optimized 3D model.

[0032] This invention acquires multi-angle images of the original model, enabling the system to comprehensively capture every detail of the model and ensuring the richness and accuracy of the data. The generation of a normalized viewpoint image set makes subsequent feature extraction more efficient and accurate. The extracted feature points provide a solid foundation for model reconstruction. The application of disparity correction technology effectively eliminates disparity errors in the images, ensuring the accuracy of feature points and thus improving the quality and credibility of the simulated 3D model. The 3D model reconstructed based on corrected model feature points can realistically reflect the geometric features of the original model. The process of collecting basic application data ensures the goal-oriented nature of the model design. The implementation of ideal requirement mapping allows the design process to closely revolve around practical application needs. The generated ideal requirement data provides a clear basis for determining the ideal 3D model structure. The segmentation of the core region makes the analysis of key model slices more focused, effectively improving the efficiency and accuracy of subsequent comparisons. By comparing the ideal 3D model structure with the highlighted areas of the key region model slices, the differences and deficiencies of the model can be quickly identified. Analyzing the direction of model differences further clarifies the specific areas that need optimization. The structural adjustment range determined based on the highlighted area comparison results provides a scientific basis for fine-tuning. Combined with interval cyclical fine-tuning based on the direction of model differences, the consistency between the key region model slices and the ideal 3D model structure is ensured. The generated adjusted and optimized 3D model not only improves the overall performance of the model, but also provides greater adaptability for subsequent applications. The entire system effectively integrates image processing, data analysis, and model optimization technologies, significantly improving the efficiency and accuracy of 3D model adjustment, providing strong technical support for various application scenarios, promoting the further development of 3D modeling technology, enhancing the functionality and reliability of models in practical applications, and ultimately realizing the intelligent and automated design and adjustment of 3D models.

[0033] In this embodiment of the invention, the three-dimensional model adjustment method includes the following steps:

[0034] Step S1: Obtain multi-angle images of the original model; perform multi-view normalization on the multi-angle images of the original model to obtain a normalized view image set; extract feature points of the original model based on the normalized view image set;

[0035] In this embodiment, a multi-axis rotation platform is used to acquire images of the static original 3D model from 0° to 360° in 2° increments. A 50-megapixel industrial-grade camera is used with a linearly polarized light source for uniform illumination. The resolution of each frame of the acquired image is maintained at 8128×6136 pixels. The images are then uniformly converted to PNG uncompressed format and archived. All acquired images are normalized for color space using a common color model conversion (e.g., sRGB to CIELAB). The images are then segmented into blocks based on the gradient edge direction. The segmentation results are then processed by pixel-level histogram equalization and scale normalization to construct a normalized viewpoint image set. In the normalization operation, the normalization scale is limited to a target image width of no more than 1024 pixels and a consistent aspect ratio. Subsequently, the SIFT (Scale Invariant Feature Transform) algorithm is used to extract feature points from each frame of the normalized image set. Low-response points (threshold less than 0.02) are removed, and the stability of the descriptor is further enhanced by SURF (Accelerated Robust Features). The initial set of feature points of the original model is then generated, and the output is a 3D feature point array data in a unified coordinate system, in millimeters.

[0036] Step S2: Perform disparity correction on the feature points of the original model to obtain the corrected model feature points; reconstruct the simulation 3D model based on the corrected model feature points;

[0037] In this embodiment, a three-dimensional feature point array is input into a multi-disparity fusion network structure based on depth disparity tensor estimation. This network uses a four-way epipolar geometry relationship for local disparity backprojection, constructs a disparity residual map through a depth multi-scale optical flow correction function, and performs point-by-point correction within an error propagation threshold of 0.5 pixels. The corrected model feature point coordinate set is output. After feature point correction, the Marching Cubes algorithm is used to generate a voxel mesh. The voxel side length is set to 0.2 mm, and bilinear interpolation is applied to generate a reconstructed simulation 3D model. After generating the mesh, curvature continuity constraints are constructed based on B-spline surfaces, and secondary fitting is performed on the mesh boundary surfaces. The final generated simulation 3D model is exported in PLY (polygonal file format) format to ensure that subsequent structural comparison and adjustment can be performed.

[0038] Step S3: Collect basic model usage data; map ideal requirements using the basic model usage data to generate ideal requirement data; determine the ideal 3D model structure using the ideal requirement data;

[0039] In this embodiment, based on the specific engineering project specification document and the relevant industrial scenario design requirement specification document, a natural language processing module is used to perform syntactic dependency parsing and semantic role labeling on the model purpose text data. A keyword extraction submodule based on the Transformer architecture generates nested functional intent vector groups, each with a dimension of 512. Ideal requirement data is constructed by structured matching between the nested word vectors and the purpose feature vectors, where the feature selection threshold is set to a similarity greater than 0.75. After construction, the ideal requirement data is embedded into a multi-layer neural graph encoding structure. Combined with the defined component label dictionary and 3D topology template library, the ideal 3D model structure is automatically output. The generated structure output is a JSON data package with structure hierarchy numbers, where each structure number contains component category, constraint rules, and 3D configuration parameters.

[0040] Step S4: Based on the basic application data of the model, the core region of the reconstructed simulation 3D model is segmented to obtain key region model slices; the highlighted regions are compared with the key region model slices based on the ideal 3D model structure, and the direction of model differences is analyzed based on the comparison results of the highlighted regions;

[0041] In this embodiment, the core functional segment data of the model is extracted based on the purpose and function positioning record table archived in the same project document. The component label field filtering module calls the existing component-purpose correspondence matrix. Based on the mapping density weight in the matrix, the corresponding component set in the simulation 3D model is extracted. The spatial region is divided in the reconstructed 3D model with the component geometric centroid as the reference. The division operation is completed using an adaptive octree structure. The minimum unit side length of the spatial division is set to 1 mm. After the division is completed, the high-density component region in each octree unit is extracted using a patch recognition algorithm. The region with a weight greater than 0.6 in the extraction result is marked as the key region model slice. After the region slice is completed, the shape context matching algorithm is used to compare each segment with the standard structural components in the ideal 3D model structure. During the comparison process, the shape matching accuracy index (Chamfer distance) is used as the basis for similarity judgment. The segment with a distance greater than 2 mm in the comparison result is marked as the structural difference part. The overall model difference direction is determined according to the statistical distribution of the difference displacement vector direction.

[0042] Step S5: Determine the structural adjustment range based on the comparison results of the highlighted areas; perform interval cyclic fine-tuning and correction on the key area model slices according to the structural adjustment range and the direction of model differences, until the key area model slices are consistent with the ideal 3D model structure, thereby generating an adjusted and optimized 3D model.

[0043] In this embodiment, when constructing the structural adjustment range, the local deviation value obtained from the comparison is used as the adjustment benchmark. The curvature adjustment function is established with the three-dimensional local curvature change as the benchmark parameter. The upper limit of the adjustment range is set to not exceed 10 mm and the step size of each fine adjustment is not greater than 1 mm. A cyclic fine adjustment mechanism is established with the difference direction as the gradient update direction. A structural adjustment priority matrix is ​​constructed according to the spatial relative position. Priority is given to processing the edge segments of the model in the difference segment. In each round of adjustment, the node control points are translated by B-spline local interpolation. The translation range of the control points is within 1 mm of the target fitting path to minimize the interpolation error. After each round of fine adjustment, the region comparison is re-performed and the residual is calculated. If the residual distance is all less than 0.5 mm, the fine adjustment loop is exited. Finally, the corrected model is exported and exported in STL (triangular mesh) format to form an adjusted and optimized three-dimensional model. The whole process uses a unified ID to identify the comparison block and the target block to form a mapping index to ensure that there is a one-to-one correspondence between the adjusted structure and the target structure without missing parts and that the position matching is stable.

[0044] Preferably, step S1 includes the following steps:

[0045] Step S11: Obtain multi-angle images of the original model; perform spectral layering processing on the multi-angle images of the original model to obtain layered view images, wherein the number of spectral layering bands is limited to 7-11 principal component channels;

[0046] Step S12: Perform color gamut normalization on the layered view images to generate a normalized view image set, wherein the maximum view offset angle between adjacent image frames does not exceed 15°.

[0047] Step S13: Perform dynamic viewing angle registration on the color gamut normalized image to generate a registered viewing angle image; perform global illumination correction based on the registered viewing angle image to generate an illumination-corrected image.

[0048] Step S14: Extract multidimensional response features from the illumination-corrected image, wherein the number of response channels is limited to no less than 128 feature channels;

[0049] Step S15: Aggregate spatial feature points based on multidimensional response features to obtain the original model feature points, wherein the aggregation radius is limited to a Euclidean neighborhood within the range of 0.8 to 1.5 units in length.

[0050] In this embodiment, when acquiring multi-angle images of the original model, a six-axis collaborative industrial camera arm equipped with a high-resolution spectroscopic image acquisition device is used. While maintaining a constant distance between the camera and the model's center, images of the model are captured along a polar coordinate trajectory. The shooting angle covers 25 evenly distributed points within a range of approximately 120° from top to bottom. The horizontal rotation step is 15°. One frame of spectroscopic image data is acquired at each angle point. The image resolution is set to 4096×3072 pixels, and the exposure time is 1 / 100 second. After image acquisition, the images are uploaded to the image preprocessing module using a 10Gbps high-speed Ethernet. A multispectral filter group is used to separate each frame of image into nine independent band layers, with the band centers at 410nm, 470nm, and 53nm respectively. The wavelengths 0nm, 590nm, 650nm, 710nm, 770nm, 830nm, and 890nm are set with corresponding bandwidths of ±10nm. The layered images are stored in a three-dimensional matrix with dimensions of 9×4096×3072, and each pixel is a 16-bit grayscale value. After processing, the output is a layered viewpoint image. When performing color gamut normalization on the above layered viewpoint images, the Lab color space mapping function is called to convert the original RGB pixel vectors into Lab three-channel vectors. After conversion, a principal component mean normalization strategy is used, subtracting the channel mean from the pixel intensity of each channel and then dividing by the channel standard deviation to ensure that the distribution center of each channel is consistent and converges near 0. To eliminate interference from differences between viewpoints, only the phase... Image pairs with a viewing angle offset of less than or equal to 15° from adjacent frames proceed to the next step of processing. The offset angle is calculated using gyroscope angle feedback information combined with camera calibration parameters. After normalization, the image output is a three-dimensional structure containing three types of attributes: image matrix, channel statistical parameters, and shooting viewing angle parameters. During dynamic viewing angle registration, an optical flow field matching combined with corner tracking method is used. First, the Harris corner detection algorithm is used to obtain the corner positions and record the corner intensities within each frame. Then, the Lucas-Kanade dense optical flow method is used to calculate the inter-corner displacement vector field between normalized image pairs. Based on the initial registration results, the RANSAC algorithm is used to remove incorrect matches and re-estimate the transformation matrix, ultimately outputting a 3×3 homography. A transformation matrix is ​​used for image reprojection to obtain a registered viewpoint image. After registration, global illumination correction is performed. By analyzing the brightness gradient distribution in multiple regions of the image, a global average illumination offset matrix is ​​constructed. This matrix is ​​then used to perform pixel-level inverse offset correction on the image's grayscale space, outputting an illumination-corrected image matrix. When extracting multi-dimensional response features from the illumination-corrected image, a shallow feature extraction network based on the ResNet50 network structure is first constructed. The number of input channels is expanded to the number of band channels in the image, and the output feature channel dimension is set to 128. All convolutional layers use a 3×3 kernel size, a stride of 1, and padding of 1 to maintain the spatial dimensions. Batch normalization is inserted between every two convolutional layers, and the ReLU activation function is used.Each image patch is processed to output a feature tensor of size 128×64×64. All feature tensors are pooled to a uniform size of 128×1×1 through max pooling, and then normalized and encoded to output a multidimensional response feature vector of length 128. An image feature set is constructed in the form of a vector list. When using the multidimensional response feature vector set to aggregate spatial feature points, a fast neighborhood lookup operation is performed based on the KD-Tree structure. First, all feature vectors are projected onto three-dimensional space according to image coordinates to establish a spatial coordinate index structure. For each feature point, a Euclidean spherical neighborhood aggregation operation is performed with that point as the center, and the aggregation radius r is limited to... Between 0.8 and 1.5, the unit length is based on the mean side length in the image pixel space, typically set to r = 1.2. During aggregation, weak response points with response intensities below a set threshold (the threshold is the top 20% quantile of the overall feature intensity) are removed. Local principal direction analysis is performed on each aggregation cluster, and the feature centroid of that local region is output as the original model feature point. All original model feature points are output in the format (x, y, z, f), where f represents a 64-bit description vector composed of the local maximum response channel number and intensity value of that feature point. Finally, the original model feature point extraction operation is completed, and the feature point set is output for subsequent processing.

[0051] Preferably, step S2, which involves disparity correction of the feature points of the original model, includes:

[0052] Determine the angle direction of each image in the multi-angle image of the original model based on the feature points of the original model;

[0053] Based on the angle direction of each image, the original model multi-angle images are mapped to the same space, thereby obtaining the model projection space data for each angle.

[0054] Based on the spatial data projected from each angle model, the original model's multi-angle images are compared with corresponding three-dimensional images to obtain the difference data of the three-dimensional images at each angle.

[0055] Based on the difference data of three-dimensional images from various angles, the projection space data of the model from various angles is mapped to viewpoint differences, thereby generating parallax data of the projection model;

[0056] By projecting disparity data, the feature points of the original model are disparity corrected, thereby obtaining the corrected model feature points.

[0057] In this embodiment, a viewpoint direction resolution operation is performed on each frame of the original model's multi-angle images. First, based on the external camera parameter matrix generated during image generation, which contains three-dimensional position vectors (such as spatial coordinates in the x, y, and z directions) and orientation rotation matrices (such as a 3×3 rotation matrix transformed based on Rodrigues rotation vectors), the viewpoint direction vector of each frame is calculated. The direction vector is then normalized and stored in an angle direction index table. This index table records the index number of each image frame and its corresponding spatial viewpoint vector, maintaining a floating-point precision of no less than 0.0001 to ensure the accuracy of direction recognition. A same-space projection operation is performed on all original model multi-angle images. Dense point cloud fusion technology is used to map each frame of image to the same spatial coordinate system, with the world coordinate system bound to the initial frame image serving as the unified spatial reference. A depth-guided voxel reconstruction method is used to reconstruct the projected spatial data. Specifically, a depth-guided voxel integration algorithm is used to combine RGB images with depth maps to construct a voxel mesh, where the voxel side length is set to 0.003 units. Truncated voxel fusion is performed using Truncated Voxel Reconstruction. The Signed Distance Function (SDM) is used to update the fusion value of each spatial voxel. A weighted average method is used during the fusion value update, with weights calculated based on depth map accuracy and camera signal-to-noise ratio. The projection result of each frame is a local 3D point cloud dataset, which is then uniformly merged into the global voxel model. For each angle, the projected spatial data is compared with the 3D model data generated from other angles. First, based on the angle direction index table, all other image frames with an angle within 15 degrees of the current angle are selected, and the voxel projection results of the corresponding frames are extracted into point cloud models. Iterative Closed Pixel Plots (ICPs) are then used to update the fusion value of each spatial voxel. The Point algorithm performs iterative comparison between the current angle model and the target angle model. During the comparison, rigid registration is performed with a fixed number of 100 iterations. In each iteration, the transformation matrix is ​​calculated using the minimum mean square error criterion, and registration results with residual errors below 0.00005 are stored in the difference comparison data table. Finally, the 3D image difference data between each angle is obtained. A difference mapping operation is performed on the voxel data of each angle to construct a viewpoint difference mapping model. First, based on the difference matrix output after comparison, the difference values ​​are projected back to the spatial voxels of each angle using a 3D resampling method. The voxel weight update algorithm is used to assign an angle response value to each voxel node. The weight distribution is based on the cosine value of the included angle in the original angle direction to construct a weighted mask matrix, and the difference value is higher than 0.A voxel of length 0.2 is labeled as a disparity anomalous node. All voxel disparity anomalous nodes at each angle are recorded to generate a high-dimensional disparity tensor with a dimensional structure of N×X×Y×Z, where N is the number of angles, and X, Y, and Z are the voxel space dimensions. The tensor values ​​are recorded as three-dimensional error offset vectors. The original model feature points are mapped to their corresponding positions in the tensor based on their coordinates in the unified projection space. Based on the three-dimensional error offsets recorded in the tensor, the coordinate values ​​of each feature point in the X, Y, and Z directions are adjusted. The adjustment amount is the original coordinates plus the error vector value at the corresponding position in the tensor. This adjustment operation performs interpolation smoothing within the feature point aggregation neighborhood, using cubic B-spline interpolation. After updating the coordinates of each feature point, its local surface normal vector and principal curvature in the model are recalculated. The output of all updated feature point sets is the corrected model feature point set, which maintains the same number of feature points as the original set. The positional error of each feature point is controlled within 0.001 units, completing the disparity correction process.

[0058] Preferably, the reconstruction of the simulation 3D model based on the feature points of the correction model in step S2 includes:

[0059] Multi-scale spatial interpolation is performed on the feature points of the correction model to obtain multi-scale interpolated point cloud data;

[0060] Adaptive density resampling is performed on multi-scale interpolated point cloud data to obtain uniform density point cloud data;

[0061] Reconstructing a continuous surface based on uniform density point cloud data, thereby generating an initial surface mesh;

[0062] Global connectivity optimization is performed on the initial surface mesh to obtain a connected and optimized mesh;

[0063] Surface manifold reconstruction is performed based on connected optimization meshes to generate manifold optimization reconstruction data;

[0064] The simulation 3D model is reconstructed based on the manifold optimization data.

[0065] In this embodiment, multi-scale spatial interpolation is performed on the feature points of the corrected model. An interpolation method based on hierarchical wavelet kernel functions is used to construct multiple scale levels in the original 3D space. Each level corresponds to a wavelet scale factor, which is set to 0.005, 0.01, 0.02, and 0.04 units. At each scale, locally weighted interpolation is performed on the feature points of the corrected model within the radius of a neighborhood sphere. The interpolation uses locally supported Mexican interpolation. The Hat wavelet kernel function is used, whose distance sensitivity is controlled by the parameter σ, which is equal to the current scale factor. During interpolation, a sphere with a radius of 2σ is established around each feature point to search for a neighborhood. Interpolation points with a fixed distribution are generated in this neighborhood. The coordinates of each interpolation point are generated by weighting the points in the neighborhood, where the weight is the wavelet function value at that point. Finally, a set of interpolation points is generated at each scale. The set of interpolation points at all scale levels is unified into multi-scale interpolated point cloud data. The total number of this point cloud data is automatically expanded according to the input feature point density and scale level, and the final number is maintained between 20 and 50 times the number of the original corrected feature points. Spatial density estimation is performed on the interpolated point cloud. The local density value of each point is calculated using the kernel density estimation method. The kernel function is a three-dimensional Gaussian kernel, and the bandwidth parameter h is set to 0.For a unit length of 0.1, points with a density below the 25th percentile in the density estimation results are marked as low-density regions, and points with a density above the 75th percentile are marked as high-density regions. The point cloud is then subjected to preservation or dilution operations based on the local density values ​​of the points. New points are inserted in low-density regions by interpolating the midpoints between points, preserving the consistency of the normal vector direction during interpolation. In high-density regions, Farthest Point Sampling is performed to remove locally redundant points. The removal operation prioritizes removing points closest to the center point based on Euclidean distance. After adjustment, uniform density point cloud data is obtained, with the number of points per unit cube fluctuating within ±3. A continuous surface reconstruction operation is then performed based on the uniform density point cloud data, using Poisson surface reconstruction. The process employs a reconstruction method. First, the normal vector of each point in the point cloud data is estimated. A local neighborhood of each point is constructed using the k-nearest neighbor method (k=20), and the normal direction is extracted using Principal Component Analysis (PCA). Consistent redirection is performed on all normal vector directions. A minimum spanning tree method is used to establish paths between points and propagate the normal direction. After obtaining the normals, an octree voxel structure is constructed with a depth of 10 layers. The spatial range of the root node is the maximum dimension of the bounding box of the point cloud. The normal direction is accumulated in each voxel node, and the Poisson equation is solved. The implicit function value is obtained by solving the Poisson equation using the multigrid method. The Marching Cubes algorithm is used to extract the isosurface from the implicit function. Finally, the initial surface mesh is output. Each facet in this mesh has a triangular topology, and the vertex position accuracy is controlled within ±0.Within a unit length range of 002, an adjacency graph model is constructed based on the mesh topology. Each node in the graph represents a triangular facet, and each edge represents the shared edge relationship between adjacent faces. First, all non-manifold edges and boundary holes are detected. The boundary is closed using an edge chain tracing method. Non-triangular faces appearing during the closure operation are subdivided using the Loop subdivision algorithm, decomposing each polygon into the smallest triangular facet group. Simultaneously, normal consistency is checked for all faces using bidirectional Breadth-First Search (BFS) to traverse the entire mesh, ensuring that the mesh normal vectors are consistent among adjacent faces. Based on this, a minimum cut graph model is constructed, connecting or pruning disconnected segments using the shortest path. The minimum energy connection strategy is selected based on the target topology integrity requirements. Finally, a connected optimized mesh is output, and surface manifold reconstruction is performed using a Laplacian-Beltrami surface adjustment method. The algorithm (Smoothing) first constructs a Laplacian adjacency weight matrix for each vertex, using inverse distance weighting coefficients for weight calculation. Then, it treats vertex coordinates as solution variables and performs minimum energy deformation on the entire mesh. The objective function is global energy minimization, i.e., minimizing the sum of the Laplacian residuals of all vertices, while maintaining edge length errors within ±5%. Boundary preservation constraints are introduced during the solution process, with Dirichlet boundary conditions set to remain unchanged for boundary vertices. After solving, it generates manifold optimization reconstruction data with continuous surfaces, consistent normals, and closed boundaries. A simulation 3D model is then generated based on this manifold optimization reconstruction data, employing a model generation method based on bilateral normal filtering and local geometric consistency matching. In the manifold data, the local curvature tensor of each vertex is first calculated. The curvature tensor consists of the principal curvature direction and magnitude. Bilateral filtering is applied to all curvature tensors to preserve the geometric sharpness of edge features. After filtering, a subdivision enhancement operation is performed using Modified... The Butterfly subdivision algorithm increases vertex density to more than four times that of the original point cloud. Then, normal reestimation and edge polyline enhancement are performed. Boundary segments with a principal curvature greater than 0.1 are marked as explicit boundaries in all vertices, and model patch bounding box information is constructed. Finally, a simulated 3D model is output, with vertex coordinates, normal directions, patch topology, and boundary information all output to a specified data channel in standard OBJ file format. The number of model faces is controlled to be between three and five times that of the original initial mesh, and the number of vertices exceeds the initial number of feature points by more than ten times to meet simulation accuracy requirements.

[0066] Preferably, step S3 includes the following steps:

[0067] Step S31: Collect basic model usage data; deconstruct the functional requirements of the basic model usage data to generate functional deconstruction data, wherein the number of functional requirement classification dimensions is limited to 12-20 structural functional dimensions;

[0068] Step S32: Perform a requirement weight distribution analysis based on the functional deconstruction data to obtain requirement weight data, wherein the maximum weight of a single functional module shall not exceed 25% of the total global weight;

[0069] Step S33: Perform ideal state mapping based on demand weight data to generate ideal demand data, wherein the ideal state mapping matrix is ​​fixed to a 6×6 state-response correspondence matrix;

[0070] Step S34: Perform spatial deconstruction mapping on the ideal requirement data to obtain spatial modeling requirement data, wherein the minimum granularity of spatial deconstruction is limited to 0.5 cubic voxels;

[0071] Step S35: Perform target structure abstract simulation based on spatial modeling requirement data to generate target abstract simulation structure; determine the ideal three-dimensional model structure through target abstract simulation structure.

[0072] In this embodiment, the structural purpose registration interface module is invoked to collect design purpose data of the 3D model under different scenarios through an industry-standard application interface API. The collected dimensions include 12 categories of purpose features: load domain, connection method, operating frequency, environmental adaptability level, dynamic response mode, thermal stress level, structural dynamic feedback parameters, assembly interface constraints, functional domain reconstruction expectations, task response time steps, motion degree of freedom parameters, and rigid area identifiers. All data is input into the purpose data container in JSON format. Subsequently, a structural function decomposition engine based on basis function decomposition is used to deconstruct the functional requirements of the above purpose data. In this process, each purpose feature dimension is assigned an independent functional category vector. A dimensional mapping matrix is ​​constructed and feature matching is performed, ultimately outputting functional deconstruction data. The deconstruction process limits the number of structural functional dimensions to between 12 and 20. Each dimension undergoes positive independence verification using a purpose-structure orthogonal classifier to ensure no redundancy or overlap between functional dimensions. All classification labels employ non-cross-coding rules, and each functional dimension must include a specific physical interpretation label for subsequent weight indexing. An independent sequence of requirement items is constructed for each functional dimension, and each functional dimension is projected onto a unified functional intensity space using a proportional density projection method. Subsequently, a weight normalization calculation engine based on a multi-layer attention mechanism is applied to perform weighted calculations on all requirement items, ensuring that the maximum weight of each functional module within the overall functional framework does not exceed the global weight. The total weight is 25%. If the input consists of 18 functional dimensions, the maximum weight of each dimension will not exceed 1 / 4. All weight calculations are based on iterative distribution optimization using structural similarity matching indicators. The initial distribution is uniform, and adjustments are made according to actual functional relevance indicators during the iteration process. Normalization is required after each adjustment. The final output is a demand weight data vector set, which is stored in a weight data table indexed by structural labels. This table is a two-dimensional array, with columns representing functional dimensions and rows representing the distribution intensity of functional points. An ideal state response matrix is ​​established based on the demand weight data, with a dimension of 6 rows and 6 columns. The rows represent six basic state parameters of the structure, including static stability, dynamic response frequency, load extension, and structural stability. The matrix is ​​constructed using residual energy, relative inertia distribution, and regional tension field distribution. Columns represent six response dimensions: maximum allowable deformation, ultimate load strength, mean dynamic response time, minimum fatigue displacement threshold, stable operating time window, and optimal structural adjustment margin. During construction, the matrix is ​​initialized using a rule-based filling method. Each element value is calculated using the function Fij = Wi × Rj, where Wi represents the functional dimension weight, and Rj represents the standardized influence coefficient corresponding to the response dimension. This coefficient is extracted from an empirical coefficient table derived from a typical simulation sample database. Each mapping must ensure that the matrix satisfies the state-response monotonically increasing constraint. Finally, the ideal requirement data matrix is ​​output and stored in a structural format for input to the subsequent spatial mapping module.The aforementioned ideal requirement data is spatially mapped and expanded according to the six-dimensional response dimensions. First, the six response dimensions are mapped to the three-dimensional coordinate axes and their gradient directions, forming a six-dimensional tensor reduced to three-dimensional space. Then, through three-dimensional vector field expansion, the logical response values ​​are projected into the voxels of the spatial structure unit. The voxel unit side length is set to 0.5 cubic units, with each voxel having a side length of 1 mm. The spatial mapping process employs a voxel-level minimum response interpolation mechanism, meaning all response value allocation must be based on the smallest voxel as the deconstruction granularity, and data must not be written across granularities during interpolation. After the spatial structure mapping is completed, Poisson filling is used to compensate and fill voxel gaps, ensuring the existence of a breakpoint-free voxel structure. Finally, the spatial modeling requirement data tensor is output. The tensor dimension is determined by the voxel space range, and the tensor value type is a multi-dimensional array, containing three types of content: response value intensity, structure number identifier, and function mapping path index. The structure is called based on the spatial modeling requirement data tensor. The topology generation engine uses a simulated annealing-based topology generation path optimizer to perform initial global structural fitting, constructing a target abstract simulation structure. This structure is generated based on minimum response constraints, with the initial structure generated from polygonal surface basic units. After constructing the initial mesh, a geometric deformation rule library is invoked to perform morphological fitting deformation based on the voxel position corresponding to each response vector. During structural simulation, each deformation operation is limited to execution within a single voxel space, restricting the maximum structural extension to no more than 15% of the maximum response extension axis length within the tensor. All simulated structures undergo global connectivity determination through a morphological consistency matching module. The generated model is output in standard STEP format as the target abstract simulation structure model. This model will serve as the basic input for the ideal 3D model structure in the next stage. All structural nodes are tagged with a unique structural number and deformation path record for subsequent structural reconstruction path backtracking. After output, the model is archived and saved to the structural abstract data pool, completing the structure generation task.

[0073] Of particular importance, step S35 includes:

[0074] Multi-layer isomorphic abstraction simulation is performed on spatial modeling requirement data to generate preliminary structural abstract metadata;

[0075] The initial structural abstract metadata is subjected to heterogeneous interactive mapping to obtain a multidimensional interactive mapping structure;

[0076] Response constraint structure data is obtained by filtering the multidimensional interactive mapping structure data.

[0077] A step-by-step iterative fusion is performed based on response constraint structure data to obtain the target abstract simulation structure;

[0078] The target abstract simulation structure is integrated with spatial topology consistency to generate a spatially consistent structure.

[0079] Based on the spatial consistency structure, global structural coupling and reorganization are performed to obtain the ideal three-dimensional model structure.

[0080] In this embodiment, all input spatial components are type-identified based on the generated spatial modeling requirement data. Homogeneous clustering is then performed according to the material parameters, volume parameters, and connection boundary information in the component attribute labels. An improved hierarchical adaptive clustering (MHAC) algorithm is used to construct an attribute-structure nested tree. In each nested structure, nodes represent structural component units, and edges represent structural connections between components. During clustering, a component volume similarity threshold of 0.15 and consistent material category coding are used as constraints. The maximum number of clustered nodes in each layer is set to 12. Components are merged into isogeneous units when the similarity between them meets the preset threshold and there are at least two boundary connection surfaces between them. Finally, preliminary structural abstract metadata is output, represented in a nested hierarchical form using JSON (JavaScript Objects). The data is saved in Notation (JavaScript object notation) format. Each node includes a component code, geometric parameters, boundary connection vectors, and node depth information. When performing heterogeneous interactive mapping on the initial structural abstract metadata, a heterogeneous attribute cross matrix is ​​constructed. The matrix dimension is m×n, where m represents the total number of isomorphic abstract nodes and n represents the number of heterogeneous attribute dimensions. In this embodiment, the value of n is fixed at 8. The matrix includes component mechanical response, thermal response, structural fatigue cycle, structural functional strength label, deformation tolerance factor, boundary conflict degree, energy consumption correlation value, and dynamic stability score. A heterogeneous feature interaction network is constructed using a Cooperative Interactive Self-Attention Mechanism to map the initial structural abstract metadata into an interaction tensor data structure containing multidimensional response relationships. Each element in this interaction tensor contains a source node number, a target attribute number, an interaction weight value, and a coupling response relationship identifier. Finally, a multidimensional interaction mapping structure is generated, which is a multidimensional sparse tensor of extended type. When performing response constraint filtering on the multidimensional interaction mapping structure data, a set of constraint conditions is set, including a maximum deformation response threshold of 1.2% and an upper limit of thermal expansion coefficient of 3.6×10. -6 / K, the structural fatigue bearing cycle shall not be less than 2000 cycles, the boundary conflict degree shall not be higher than 0.75, and the dynamic stability score shall not be lower than 0.85. A logical-gated convolutional filter network is used to discriminate the responses of each dimension in the interaction tensor channel by channel. Structural units that meet all filtering conditions are marked with a retention flag, while nodes that do not meet the conditions and their attribute numbers are recorded, generating response constraint structural data. This data is stored in key-value pairs with structural unit numbers as keys and attribute retention flags as values. When performing step-by-step iterative fusion based on the response constraint structural data, a multi-branch structural iterative fusion network is used. The network initially uses structural units marked as 1 as a base and performs three rounds of recursive fusion layer by layer. In each round of fusion, the current layer structural unit and its adjacent units in three-dimensional space are geometrically reconstructed and their attributes are fused. The reconstruction process uses the Bézier three-dimensional surface fitting method for boundary fitting. The attribute fusion uses a bidirectional gated recurrent network (Bi-GRU) to perform time-series-preserving fusion of each attribute. The fusion convergence conditions are set as follows: structural deformation rate is less than 0.5%, energy consumption index change is less than 0.2%, and boundary conflict degree decrease is greater than 0.1. After fusion, a target abstract simulation structure is generated, represented in octree format. Each node contains the component spatial index, fusion level, comprehensive response value, and spatial boundary coordinate range. During spatial topological consistency integration of the target abstract simulation structure, all nodes and their spatial index data are extracted, a 3D adjacency graph structure is constructed, and a spatial topological consistency check algorithm is used. The CheckingAlgorithm performs topological rule comparisons on the connection edges between all nodes, setting the judgment rules as boundary continuity, angle matching, and voxel overlap ratio not exceeding 5%. For node connection pairs that do not meet the topological consistency requirements, a local reconstruction operation is performed. A local boundary reconstruction model is used to fit the boundary surfaces of adjacent nodes using the average least squares method. After topological integration, a spatially consistent structure is generated, and the structure output is a voxel-level mesh structure model. Each voxel records its absolute position in the 3D coordinate system, adjacent voxel numbers, boundary identifiers, and topological consistency flags. When performing global structural coupling and recombination based on the spatially consistent structure, all spatially consistent structural units are used as input to construct a global structural coupling graph model. Nodes are structural units, and edges are coupling paths. A method based on the Structural Function Coupling Graph is used, setting the coupling weights between nodes based on the functional dependency paths in the previously generated functional deconstruction data. The shortest structural energy consumption path search algorithm is used to search for the optimal combination path of structural functions. The searched structural combination paths are mapped to 3D space, and the structural units associated with the corresponding paths are spatially rearranged. During the rearrangement process, a Structural Voxel Reconstruction Engine is used. Voxel Rebuilder transforms the inter-element connection boundaries into continuous surfaces and performs a global mesh reconstruction operation. The final ideal 3D model structure is stored in standard STL (stereolithography) format, with the mesh density set to include a triangular facet with a minimum side length of 0.1 units per cubic voxel.

[0081] Preferably, step S4, which involves dividing the core region of the reconstructed simulation 3D model based on the model's basic usage data, includes:

[0082] Based on the model's basic usage data, the regions are prioritized to obtain the region priority data;

[0083] The reconstructed simulation 3D model is spatially segmented based on regional priority data to obtain initial model slices;

[0084] Interactive region mapping is performed on the region priority data and the initial model slices to generate mapped region slice data;

[0085] Hotspot region assessment is performed on the initial model slices based on the mapped region slice data, thereby obtaining key region model slices.

[0086] In this embodiment, by reading the labeled purpose classification dataset in the task configuration module, five sets of feature data are extracted for each purpose label: functional instruction parameters, target task accuracy threshold, interaction frequency level, operation density factor, and region dependency matrix. The above data are uniformly vector-encoded by constructing a five-dimensional purpose vector group. The length of the functional instruction parameter vector is set to 128 dimensions, the task accuracy threshold is mapped to the normalized distribution interval [0.1, 1.0], and the interaction frequency level is divided into 5 levels and assigned values ​​of 0.2, 0.4, 0.6, 0.8, and 1 respectively.0. The operation density factor is calculated by counting the number of operations per unit area in the simulation interaction record and then performing log compression. The dimension of the region dependency matrix is ​​set to n×n, where n represents the number of recognition regions in the model. Subsequently, the functional fit score between each group of purpose vectors and the standard function instruction set is calculated using cosine similarity. Then, all regions are grouped by purpose and sorted in descending order of fit score to form a region priority list. During the spatial block segmentation of the reconstructed simulation 3D model based on the region priority data, the region priority list is indexed and mapped to the spatial labels in the reconstructed simulation 3D model. The Octree octagonal space partitioning algorithm is used to perform 3D voxelization of the model's spatial structure. The initial voxel side length is 10 mm. Each spatial node is labeled with a priority level based on its ranking in the priority list. A multi-resolution segmentation factor is introduced during partitioning, setting the boundary voxel precision of high-priority regions to 5 mm, medium-priority regions to 10 mm, and low-priority regions to maintain the original model precision. Spatial boundary continuity is ensured through point cloud distribution density constraints, ultimately generating an initial model slice set. This set contains the spatial boundary voxel number, region identifier, spatial orientation matrix, and voxel count statistics for each slice. During the interactive region mapping process of the region priority data and the initial model slices, the 3D spatial annotation tool VTK (Visualization) is used. The Toolkit (a 3D visualization toolkit) loads the initial slices into a GPU-accelerated rendering environment. Relying on OpenGL, it renders the outlines of each slice and their positional distribution in 3D coordinate space in real time. It then reads the region usage labels matching each slice from the priority data. Engineers manually confirm the correlation between region labels and slices through mouse trajectory capture and point-and-click mapping. The system synchronously writes all mapping results into a mapping matrix. This mapping matrix uses sparse matrix storage to record the mapping relationship between each usage label and the slice's spatial index number. Simultaneously, it performs a secondary confirmation by combining the Euclidean distance between the centroid coordinates of each slice in 3D space and the center point of the usage region. Regions with a distance less than 5% of the model's principal scale are labeled as directly mapped regions, while those between 5% and 10% are labeled as staggered edge regions. Finally, a set of mapped region slice data is generated. Using this data as evaluation input, heatmap data under the corresponding usage labels from historical simulation records is retrieved. A 3D hotspot scoring model is constructed by extracting the frequency of unit voxel operations, the number of interactions per unit time, and the density of interaction behavior types from the heatmap. The hotspot scoring model uses voxels as the smallest evaluation unit, with a threshold set to 1 of the average interaction density.Five times the threshold is applied. When the number of voxels exceeding this threshold in a certain area slice exceeds 40% of the total number of voxels in the slice, the slice is designated as a hotspot region. Based on the hotspot level, it is further divided into three levels: high-intensity hotspot region, medium-intensity hotspot region, and weak hotspot region. All evaluation results are exported as an evaluation result set containing slice number, hotspot level label, thermal center point coordinates, and interaction density vector, and converted into .csv structured data. Finally, a set of key area model slices is generated and synchronously passed to the next processing module.

[0087] Preferably, step S4 involves comparing the highlighted areas based on the ideal 3D model structure and the key area model slices, and analyzing the direction of model differences based on the highlighted area comparison results, including:

[0088] Locate the corresponding structural slice area in the ideal 3D model structure by slicing key areas of the model;

[0089] Overlap projection comparison is performed on key area model slices and corresponding structural slice areas to obtain the comparison results of highlighted areas;

[0090] Locally differing areas can be located by comparing the results of highlighted areas;

[0091] Convert key area model slices into a model space coordinate system;

[0092] Based on the model space coordinate system, the coordinate orientation of the differences in the local difference region is determined;

[0093] The direction of the model difference can be inferred from the coordinates of the difference.

[0094] In this embodiment, key area model slices are loaded sequentially into the 3D slice reading module according to their numbers. This module calls a geometric boundary parsing algorithm based on voxel segmentation to extract the outer contour boundary of each slice model. The contour boundary is represented by a 3D bounding box formed by the outermost vertices. Each bounding box has six values ​​representing its minimum and maximum x, y, and z coordinates, which are then uniformly stored as structure data. Subsequently, the 3D index matching module is started to load the ideal 3D model structure. This model must be a closed voxelized triangular patch model. The model needs to be automatically voxelized during import, with a voxel segmentation accuracy set to 1 mm. Each voxel block consists of its center coordinates and voxel number. The system iterates through the coordinate range in the bounding box of each key area model slice. The search process involves enumerating each coordinate point within the bounding box, matching the voxel number of that point, and statistically analyzing the ideal model region numbers corresponding to all overlapping voxels. The region number with the highest frequency in the statistical results is selected as the corresponding structural region for the current key slice. Finally, all triangular faces of the corresponding structural region are exported as separate geometric model files. The key region slice and its corresponding ideal structural slice files are loaded separately through a 3D model comparison module. After loading each model, this module first performs triangular facet normal vector correction to ensure that the facet normal vectors uniformly face outwards from the model. Then, it performs uniform scale normalization on both models, mapping the coordinate values ​​of all points to a uniform unit space, with the smallest coordinate value as the origin. The maximum boundary value is the unit size. Then, the two models are geometrically aligned in standard coordinate space. The geometric center is calculated by extracting the average value of all vertices. After model alignment, the overlapping projection process begins. The projection algorithm is based on a point-to-point nearest neighbor search. For the center point of each facet in the key area model slice, the system finds its nearest neighbor facet in the ideal structure slice model and measures the geometric distance between them. If the distance is less than the set standard tolerance value of 1 mm, it is considered an overlapping area and recorded as a green label; otherwise, it is recorded as a red label as a difference area. After all facets are compared, the comparison label information is mapped back to the original model. In the rendering module, facets with different labels are assigned corresponding colors, through O The penGL rendering method outputs a highlighted image and generates a patch difference report. This report includes the patch IDs, location coordinates, difference distance information, and the corresponding structure file paths. The system reads the difference description file generated from the comparison results. This file records the IDs and spatial coordinates of all triangular patches marked as difference regions. The system aggregates the geometric centers of these patches, grouping patches less than 3 mm apart into the same difference cluster. The clustering process uses a spatial connectivity analysis method with a fixed threshold condition. It recursively searches for all other patches within 3 mm of the current patch's center point and adds them to the current cluster until no new connections are added to any patch within the cluster.The system then calculates the minimum bounding box for each differential cluster. This bounding box is defined by the minimum and maximum x, y, and z values ​​among all vertices of the cluster. The center point of the bounding box is the spatial location identifier of the local differential region. The system records each cluster number, number of contained patches, boundary coordinate range, spatial center coordinates, and cluster number in a structured data table. This data table is uniformly saved as a differential region list file in .csv format, with each row corresponding to a differential region. The system parses the original position information of the key slice model, which includes the origin offset and modeling orientation rotation information during slice modeling. The offset data is represented in three-dimensional coordinates, and the rotation information is represented by three Euler angles. The system first converts all the coordinates... The vertex data of the triangular facets is read into a 3D coordinate array, and its spatial coordinates are adjusted using a spatial coordinate transformation module. The transformation process consists of two stages: the first stage is rotation transformation, where the system converts Euler angles into a 3D rotation matrix and performs matrix operations on the vertex coordinates one by one to adjust their orientation; the second stage is translation, where the system adds an offset vector to each rotated vertex and translates it to its new position in the model's global coordinate system. After all point adjustments are completed, the system performs an integrity check on the model structure to ensure that the vertex connections are not misaligned. Then, the updated model data is exported again as a new standard triangular facet file, and the 3D coordinate values ​​of the center point of each difference region are extracted from the list file of difference regions in the global coordinate system. The system synchronously reads the geometric center coordinates of the entire ideal 3D model. This geometric center is obtained by reading all vertex data and averaging their x, y, and z coordinates. The system subtracts the x, y, and z values ​​of the center point of each difference region from the x, y, and z values ​​of the model's center point to obtain the spatial offset data of that difference region relative to the model's geometric center. Then, by comparing the absolute values ​​of the coordinate offsets, the system determines the axial direction of the maximum offset. If the x-axis offset is the largest, it is marked as forward / backward; if the y-axis offset is the largest, it is marked as left / right; and if the z-axis offset is the largest, it is marked as up / down. During analysis, the system also records the positive and negative values ​​of the offset for each direction to determine the specific orientation. For example, a positive x-axis offset indicates a forward deviation, and a negative x-axis offset indicates a negative x-axis offset. This is the backward bias. All discrepancy regions are assigned explicit coordinate orientation labels according to the above rules, and the original coordinate values, offsets, and orientation labels are uniformly recorded in a structured JSON file. Each discrepancy region has a complete description of its spatial coordinates and orientation. Based on the aforementioned discrepancy region information file with orientation labels, the distribution of all discrepancy regions in each main direction is statistically analyzed. The statistics include the number of discrepancy regions appearing in each direction, the total number of discrepancy patches, the percentage of patches in that direction out of the total number of discrepancy patches, and the total distance between the center point of all discrepancy regions in that direction and the geometric center of the model. The system categorizes these statistical data into six directions: front, back, left, right, up, and down.Subsequently, based on the set judgment threshold, the main direction of difference is determined. If the proportion of differing patches in a certain direction exceeds 45% of all differing patches, and the average offset distance is the maximum among all directions, then the system takes that direction as the main direction of difference in the current model, and outputs the judgment result and its corresponding statistical information as a .diffdir file. The file content includes the main direction label, offset percentage, average offset distance, and a list of associated differing region numbers.

[0095] Preferably, step S5 includes the following steps:

[0096] Step S51: Based on the comparison results of the highlighted areas, adjust the geometric amplitude calibration to obtain the adjustment geometric amplitude data; determine the structural adjustment amplitude range through the adjustment geometric amplitude data;

[0097] Step S52: Map the adjustment constraints according to the range of structural adjustment and the direction of model differences to obtain model adjustment constraint data;

[0098] Step S53: Fine-tune and correct the model adjustment constraint data based on the model adjustment constraint data to obtain the fine-tuned and corrected 3D model;

[0099] Step S54: Compare the model by fine-tuning and correcting the 3D model and adjusting the model constraint data. When the model comparison result shows that the model is inconsistent, repeat the fine-tuning and correction operation until the model slices in the key areas are consistent with the ideal 3D model structure, thereby generating an adjusted and optimized 3D model.

[0100] In this embodiment, based on the comparison results of the highlighted areas of the generated 3D model and the ideal 3D structural model, 3D coordinate deviation data corresponding to the highlighted areas are collected. The deviation data is generated based on the Euclidean distance difference between the actual position of each highlighted pixel in the 3D model and the corresponding position in the ideal structure. The deviations of all highlighted points are decomposed in the X, Y, and Z directions in the structural coordinate system to form a 3D vector matrix data set. A custom spatial vector accumulator is used to average these 3D deviation vector data, extract the maximum and minimum offset vector components, and construct the geometric offset amplitude range of the current model based on this. Then, the bounding box is used to... The Box calculation method generates the minimum envelope size required for the structure to be adjusted in space, thereby determining the minimum and maximum adjustment thresholds for the structural adjustment range in the X, Y, and Z axes. Each threshold is recorded in millimeters. Based on the obtained structural adjustment range and the distribution direction data of the highlighted area in the 3D model coordinate space, a preset model difference direction mapping function is called. The structural difference distribution direction is projected onto the 3D coordinate system axis of the adjustment range through vector projection. A directional adjustment weight matrix is ​​constructed on each axis. The value of each matrix is ​​normalized by dividing the projection length of the difference direction on that axis by the total projection length, generating a set of directional adjustment constraint parameters. Combined with the structural adjustment range, constraint mapping is performed on the 3D model adjustment vector space. Model adjustment constraint data is constructed in the tensor space built in TensorFlow in the form of sparse tensors. Point-to-point linear interpolation is used to generate intermediate transition constraints in the mapping. The adjustment force vector value is obtained through interpolation at the transition point. This data will be used to control the direction and range of subsequent fine-tuning. Based on the generated model adjustment constraint data, local structural regions with deviations in the model are extracted and fine-tuned. The correction process calls a control point correction algorithm based on 3D B-spline surfaces. The fine-tuning of the control point positions affects the topology of the mesh model. Specifically, a control point mapping relationship is established for all mesh nodes in the key structural regions. Each control point is assigned a set of 3D adjustment vectors from the adjustment constraint data. These vectors are automatically smoothed according to the curvature continuity of the local region. The adjustment range of the control points is limited to the previously obtained structural adjustment range. After each control point update, the local surface is regenerated and its topological continuity is checked. Regions that do not meet C2 continuity are recorded and enter the next fine-tuning cycle. All processes are completed by a custom 3D mesh control module built in C++, with the mesh size starting at 0.The model is divided into 2-millimeter units. After obtaining the fine-tuned 3D model, it is structurally compared with the ideal 3D model. The structural comparison uses a point cloud registration method, employing an improved Iterative Closest Point (ICP) algorithm for point cloud relocalization. The comparison process focuses on key areas, comparing the 3D coordinate position deviations and normal vector direction differences of all structural key points. The registration error is calculated by recording the deviation value for each key point and calculating the average and maximum deviations. When the average deviation exceeds the preset tolerance threshold of 0.15 mm or the maximum deviation exceeds 0.3 mm, the point numbers and offsets of all points that do not meet the error requirements are recorded and automatically... Following the aforementioned fine-tuning and correction process, the 3D offsets of the control points in the corresponding areas are readjusted, and a fine-tuning process is executed. All comparison results are exported in CSV format for subsequent review. This correction process is repeated until all structural slices in the key areas, after 3D reconstruction, are completely consistent with the key point coordinates and normal directions of the ideal 3D model structure and pass the structural continuity test. Finally, the optimized complete 3D model is output and encapsulated for export in both OBJ and STL formats for multi-platform compatibility. All processing is controlled by a self-built 3D modeling and adjustment system, which automatically records the entire process log for traceability and analysis.

[0101] Of particular importance, step S54 includes:

[0102] A global correspondence mapping is performed between the fine-tuned and corrected 3D model and the model adjustment constraint data to obtain global comparison mapping data;

[0103] Local consistency determination is performed based on global comparison mapping data to generate local consistency determination data;

[0104] Extracting the highlighted areas of residuals from data based on local consistency discrimination;

[0105] The residual highlighted areas are subjected to differential hierarchical aggregation to obtain hierarchical differential aggregation data;

[0106] Based on the hierarchical difference aggregation data and model adjustment constraint data, targeted cyclic fine-tuning and correction are performed to generate a cyclically corrected 3D model;

[0107] A global consistency verification process is performed on the iteratively corrected 3D model. If the verification passes, the adjusted and optimized 3D model data is generated; if it fails, the above steps are repeated to continue the iterative correction.

[0108] In this embodiment, all vertex coordinates and their topological structure information of the fine-tuned 3D model are extracted. The adjustment vectors and constraint ranges contained in the model adjustment constraint data are mapped to the corresponding vertices of the fine-tuned model. A corresponding mapping matrix is ​​formed by establishing a point-to-point mapping relationship. The rows of this matrix represent the vertices of the fine-tuned model, and the columns represent the adjustment parameter indices in the constraint data. Sparse matrix storage is used to reduce computational resource consumption. Subsequently, a GPU-accelerated parallel computing module is called to perform a product operation on the corresponding mapping matrix, calculating the resultant force of all adjustment vectors on each vertex of the fine-tuned model. This yields global alignment mapping data, which is output as a vertex coordinate offset vector field, with the offset unit measured in millimeters. The vertex shader of the OpenGL graphics rendering engine is used to accelerate computation during processing. A block partitioning algorithm is used to divide the fine-tuned 3D model into several voxel blocks, each voxel block having a size... The size is set to a 1 mm cube. The consistency index of the region is determined by calculating the mean and variance of the offsets of all vertices within the voxel block. The variance threshold is set to 0.02 mm². Regions below this threshold are considered locally consistent, while those above are marked as locally inconsistent. The generated local consistency data is stored in binary image format, where 1 represents consistency and 0 represents inconsistency. The storage format uses a three-dimensional array for easy subsequent indexing. The local consistency calculation is accelerated using CUDA-based parallel computing. The set of voxel blocks in the locally inconsistent regions is used as the basis for the residual regions. The vertices in the residual regions are mapped back to the fine-tuning and correction 3D model space coordinate system. The mesh surfaces formed by these vertices are extracted as residual highlight regions. The vertex shader program assigns a highlight red color to the surface of this region, with the specific RGB values ​​set to (255, 0, 0) and the opacity set to 0.8. OpenGL is used. The Framebuffer object is rendered and labeled. The boundaries of the residual region are extracted using the Marching Cubes algorithm, with a boundary smoothing parameter set to 0.5 to ensure natural transitions at the edges of highlighted areas. Based on the offset of the vertices in the residual region, the offsets are divided into levels at 0.1 mm intervals, with each level forming a separate aggregate subset. The K-means clustering algorithm is used to spatially cluster the vertices within each level, with the number of clusters automatically adjusted according to the vertex density. The density threshold is no less than 50 vertices per cubic millimeter. The arithmetic mean of the coordinates of all member vertices is calculated for the cluster centers. Each aggregate subset corresponds to a difference intensity value, which is the maximum value of the vertex offset within that level. The final generated hierarchical difference aggregation data is stored in a dictionary structure, with the level number as the key and the corresponding cluster center coordinates and difference intensity parameter as the value. The coordinates of each difference cluster center are used as the fine-tuning target point. Combined with the constraint vectors in the corresponding coordinate directions in the model adjustment constraint data, the gradient descent algorithm is used to adjust and correct the positions of the control points in the 3D model, with the gradient descent learning rate set to 0.005. The iteration count is set to a maximum of 100. During fine-tuning, the continuity of the model mesh and the normal direction are checked after each adjustment. If the check fails, the previous state is restored and readjustment is performed. Weighting coefficients are applied to the X, Y, and Z axes respectively, with weighting coefficients of 0.6, 0.3, and 0.1, ensuring that the adjustment process prioritizes the main directions of difference. The fine-tuning algorithm is implemented in C++ and combined with the Eigen matrix library for efficient matrix calculation. A multi-scale shape comparison algorithm is called, covering the entire model and key areas. The coordinate deviation and normal angle deviation of all vertices are calculated, with a coordinate deviation threshold of 0.1 mm and a normal angle deviation threshold of 5 degrees. The verification process is achieved by constructing shape histograms and curvature statistics. All data is stored in a database for historical comparison analysis. When all detection indicators meet the threshold requirements, the final adjusted and optimized 3D model data is generated. The model data is exported in a standard format supporting multiple CAD systems, including STEP and IGES formats. If the thresholds are not met, the directional cyclic fine-tuning correction step is automatically called again to continue iteration until all verification indicators are passed.

[0109] The present invention also provides a three-dimensional model adjustment system for performing the three-dimensional model adjustment method described above, the three-dimensional model adjustment system comprising:

[0110] The image acquisition module is used to acquire multi-angle images of the original model; perform multi-view normalization on the multi-angle images of the original model to obtain a normalized view image set; and extract feature points of the original model based on the normalized view image set.

[0111] The model reconstruction module is used to perform disparity correction on the feature points of the original model, thereby obtaining the corrected model feature points; and to reconstruct the simulation 3D model based on the corrected model feature points.

[0112] The requirement mapping module is used to collect basic model usage data; to perform ideal requirement mapping based on the basic model usage data to generate ideal requirement data; and to determine the ideal 3D model structure based on the ideal requirement data.

[0113] The difference comparison module is used to segment the core region of the reconstructed simulation 3D model based on the model's basic application data, thereby obtaining key region model slices; it compares the highlighted regions with the ideal 3D model structure and the key region model slices, and analyzes the direction of model differences based on the highlighted region comparison results;

[0114] The cyclic adjustment module is used to determine the range of structural adjustment based on the comparison results of the highlighted areas; according to the range of structural adjustment and the direction of model differences, the key area model slices are cyclically fine-tuned and corrected until the key area model slices are consistent with the ideal 3D model structure, thereby generating an adjusted and optimized 3D model.

[0115] This invention, through the application of an image acquisition module, enables the system to comprehensively collect multi-angle images of the original model, ensuring data diversity and integrity. The generation of a normalized viewpoint image set improves the accuracy of subsequent feature point extraction. The extracted feature points of the original model provide a solid foundation for subsequent model reconstruction. The implementation of parallax correction technology effectively eliminates errors in the images, ensuring the accuracy of feature points, thereby improving the quality and realism of the reconstructed 3D model. The reconstructed simulation 3D model can realistically reflect the geometric features of the original model. The requirement mapping module, by collecting basic model usage data, ensures the clarity of design objectives. The realization of ideal requirement mapping allows the design process to closely revolve around practical application requirements. The generated ideal requirement data provides a clear basis for determining the ideal 3D model structure. The segmentation of the core region makes the analysis of key region model slices more focused, improving the efficiency and accuracy of subsequent comparisons. The difference comparison module compares the ideal 3D model structure with the key region model slices in highlighted areas. Yes, it can quickly identify model differences and deficiencies, analyze the direction of model differences to clarify the specific areas for optimization, and the cyclic adjustment module determines the structural adjustment range based on the comparison results of highlighted areas, providing a scientific basis for fine-tuning. Combined with the cyclic fine-tuning of the model difference direction, it ensures the consistency of key area model slices with the ideal 3D model structure. The generated adjusted and optimized 3D model not only improves the overall performance of the model, but also provides greater adaptability for subsequent applications. The entire system effectively integrates image processing, data analysis, and model optimization technologies, significantly improving the efficiency and accuracy of 3D model adjustment, providing strong technical support for various application scenarios, promoting the further development of 3D modeling technology, enhancing the functionality and reliability of models in practical applications, and ultimately realizing the intelligent and automated design and adjustment of 3D models. This promotes technological progress and innovation in related industries, ensures the efficient application of models in complex and ever-changing environments, and enhances product competitiveness and market responsiveness.

[0116] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed, implements the three-dimensional model adjustment method as described in any of the above claims.

[0117] This invention provides a highly efficient 3D model adjustment capability through a computer program stored on a computer-readable storage medium, ensuring the accuracy and consistency of model data. It integrates image acquisition and processing technologies, improving the accuracy of model feature extraction, optimizing the reconstruction process, and enhancing the realism of the simulated 3D model. Based on the requirement mapping function, it clarifies the design objectives, ensuring that the model structure meets the actual application requirements. It uses difference comparison technology to quickly identify model defects, promoting targeted optimization. The cyclic adjustment function enables fine-grained correction of model details, enhancing the model's adaptability and performance. The digital and automated characteristics of the entire system promote the improvement of design efficiency and the rational use of resources.

[0118] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0119] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for adjusting a three-dimensional model, characterized in that, Includes the following steps: Step S1: Obtain multi-angle images of the original model; Multi-view normalization is performed on the multi-angle images of the original model to obtain a normalized view image set; Feature points of the original model are extracted based on a normalized viewpoint image set; Step S2: Perform disparity correction on the feature points of the original model to obtain the corrected model feature points; reconstruct the simulation 3D model based on the corrected model feature points; Step S3: Collect basic application data for the model; Ideal requirement data is generated by mapping the model's basic usage data to ideal requirements data; the ideal 3D model structure is then determined using the ideal requirement data, including the following steps: Step S31: Collect basic application data of the model. Specifically, collect the design application data of the 3D model in different scenarios through the industry standard application interface (API). The collected dimensions include 12 application feature dimensions: load domain, connection method, operation frequency, environmental adaptability level, dynamic response mode, thermal stress level, structural dynamic feedback parameters, assembly interface constraints, functional domain reconstruction expectation, task response time step, motion degree of freedom parameters, and rigid area identification. Deconstruct the functional requirements of the basic application data of the model to generate functional deconstruction data. The number of functional requirement classification dimensions is limited to 12-20 structural functional dimensions. Step S32: Perform a requirement weight distribution analysis based on the functional breakdown data to obtain requirement weight data, wherein the maximum weight of a single functional module shall not exceed 25% of the total global weight; Step S33: Perform ideal state mapping based on demand weight data to generate ideal demand data, wherein the ideal state mapping matrix is ​​fixed to a 6×6 state-response correspondence matrix; including: establishing an ideal state response matrix based on demand weight data, setting the mapping matrix to a dimension of 6 rows and 6 columns, where the rows represent the six basic state parameters of the structure, and the columns represent the six response dimensions; Step S34: Perform spatial deconstruction mapping on the ideal requirement data to obtain spatial modeling requirement data. The minimum granularity of spatial deconstruction is limited to 0.5 cubic voxels. Specifically, the ideal requirement data is spatially mapped and expanded according to the six-dimensional response dimensions. The six response dimensions are mapped to the three-dimensional spatial coordinate axes and their gradient directions, forming a six-dimensional tensor to a three-dimensional space. The logical response values ​​are projected into the spatial structure unit voxels through the expansion of the three-dimensional vector field. Step S35: Perform target structure abstract simulation based on spatial modeling requirement data to generate target abstract simulation structure; determine the ideal 3D model structure through the target abstract simulation structure; Step S4: Based on the basic application data of the model, the core region of the reconstructed simulation 3D model is segmented to obtain key region model slices; the highlighted regions are compared with the key region model slices based on the ideal 3D model structure, and the direction of model differences is analyzed based on the comparison results of the highlighted regions; Step S5: Determine the structural adjustment range based on the comparison results of the highlighted areas; perform interval cyclic fine-tuning and correction on the key area model slices according to the structural adjustment range and the direction of model differences, until the key area model slices are consistent with the ideal 3D model structure, thereby generating an adjusted and optimized 3D model.

2. The three-dimensional model adjustment method according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Obtain multi-angle images of the original model; perform spectral layering processing on the multi-angle images of the original model to obtain layered view images, wherein the number of spectral layering bands is limited to 7-11 principal component channels; Step S12: Perform color gamut normalization on the layered view images to generate a normalized view image set, wherein the maximum view offset angle between adjacent image frames does not exceed 15°. Step S13: Perform dynamic viewing angle registration on the color gamut normalized image to generate a registered viewing angle image; perform global illumination correction based on the registered viewing angle image to generate an illumination-corrected image. Step S14: Extract multidimensional response features from the illumination-corrected image, wherein the number of response channels is limited to no less than 128 feature channels; Step S15: Aggregate spatial feature points based on multidimensional response features to obtain the original model feature points, wherein the aggregation radius is limited to a Euclidean neighborhood within the range of 0.8 to 1.5 units in length.

3. The three-dimensional model adjustment method according to claim 1, characterized in that, Step S2, which involves disparity correction of the feature points of the original model, includes: Determine the angle direction of each image in the multi-angle image of the original model based on the feature points of the original model; Based on the angle direction of each image, the original model multi-angle images are mapped to the same space, thereby obtaining the model projection space data for each angle. Based on the spatial data projected from each angle model, the original model's multi-angle images are compared with corresponding three-dimensional images to obtain the difference data of the three-dimensional images at each angle. Based on the difference data of three-dimensional images from various angles, the projection space data of the model from various angles is mapped to viewpoint differences, thereby generating parallax data of the projection model; By projecting disparity data, the feature points of the original model are disparity corrected, thereby obtaining the corrected model feature points.

4. The three-dimensional model adjustment method according to claim 1, characterized in that, Step S2, which involves reconstructing the simulation 3D model based on the feature points of the correction model, includes: Multi-scale spatial interpolation is performed on the feature points of the correction model to obtain multi-scale interpolated point cloud data; Adaptive density resampling is performed on multi-scale interpolated point cloud data to obtain uniform density point cloud data; Reconstructing a continuous surface based on uniform density point cloud data, thereby generating an initial surface mesh; Global connectivity optimization is performed on the initial surface mesh to obtain a connected and optimized mesh; Surface manifold reconstruction is performed based on connected optimization meshes to generate manifold optimization reconstruction data; The simulation 3D model is reconstructed based on the manifold optimization data.

5. The three-dimensional model adjustment method according to claim 1, characterized in that, Step S4 involves dividing the core region of the reconstructed simulation 3D model based on the model's basic usage data, including: Based on the model's basic usage data, the regions are prioritized to obtain the region priority data; The reconstructed simulation 3D model is spatially segmented based on regional priority data to obtain initial model slices; Interactive region mapping is performed on the region priority data and the initial model slices to generate mapped region slice data; Hotspot region assessment is performed on the initial model slices based on the mapped region slice data, thereby obtaining key region model slices.

6. The three-dimensional model adjustment method according to claim 1, characterized in that, In step S4, a comparison of highlighted areas is performed between the ideal 3D model structure and the key area model slices. Based on the comparison results, the directions of model differences are analyzed, including: Locate the corresponding structural slice area in the ideal 3D model structure by slicing key areas of the model; Overlap projection comparison is performed on key area model slices and corresponding structural slice areas to obtain the comparison results of highlighted areas; Locally differing areas can be located by comparing the results of highlighted areas; Convert key area model slices into a model space coordinate system; Based on the model space coordinate system, the coordinate orientation of the differences in the local difference region is determined; The direction of the model difference can be inferred from the coordinates of the difference.

7. The three-dimensional model adjustment method according to claim 1, characterized in that, Step S5 includes the following steps: Step S51: Based on the comparison results of the highlighted areas, adjust the geometric amplitude calibration to obtain the adjustment geometric amplitude data; determine the structural adjustment amplitude range through the adjustment geometric amplitude data; Step S52: Map the adjustment constraints according to the range of structural adjustment and the direction of model differences to obtain model adjustment constraint data; Step S53: Fine-tune and correct the model adjustment constraint data based on the model adjustment constraint data to obtain the fine-tuned and corrected 3D model; Step S54: Compare the model by fine-tuning and correcting the 3D model and adjusting the model constraint data. When the model comparison result shows that the model is inconsistent, repeat the fine-tuning and correction operation until the model slices in the key areas are consistent with the ideal 3D model structure, thereby generating an adjusted and optimized 3D model.

8. A three-dimensional model adjustment system, characterized in that, For performing the three-dimensional model adjustment method as described in claim 1, the three-dimensional model adjustment system includes: The image acquisition module is used to acquire multi-angle images of the original model; perform multi-view normalization on the multi-angle images of the original model to obtain a normalized view image set; and extract feature points of the original model based on the normalized view image set. The model reconstruction module is used to perform disparity correction on the feature points of the original model, thereby obtaining the corrected model feature points; and to reconstruct the simulation 3D model based on the corrected model feature points. The requirement mapping module is used to collect basic model usage data; to perform ideal requirement mapping based on the basic model usage data to generate ideal requirement data; and to determine the ideal 3D model structure based on the ideal requirement data. This includes the following steps: Collect basic usage data of the model; deconstruct the functional requirements of the basic usage data of the model to generate functional deconstruction data, wherein the number of functional requirement classification dimensions is limited to 12-20 structural functional dimensions; Based on the functional breakdown data, a requirement weight distribution analysis is performed to obtain requirement weight data, wherein the maximum weight of a single functional module shall not exceed 25% of the total global weight; Ideal state mapping is performed based on demand weight data to generate ideal demand data, wherein the ideal state mapping matrix is ​​fixed to a 6×6 state-response correspondence matrix. Spatial deconstruction mapping is performed on the ideal requirement data to obtain spatial modeling requirement data, wherein the minimum granularity of spatial deconstruction is limited to 0.5 cubic voxels; Based on spatial modeling requirements data, target structure abstraction simulation is performed to generate target abstract simulation structure; the ideal 3D model structure is determined through the target abstract simulation structure. The difference comparison module is used to segment the core region of the reconstructed simulation 3D model based on the model's basic application data, thereby obtaining key region model slices; it compares the highlighted regions based on the ideal 3D model structure and the key region model slices, and analyzes the direction of model differences based on the highlighted region comparison results; The cyclic adjustment module is used to determine the range of structural adjustment based on the comparison results of the highlighted areas; according to the range of structural adjustment and the direction of model differences, the key area model slices are cyclically fine-tuned and corrected until the key area model slices are consistent with the ideal 3D model structure, thereby generating an adjusted and optimized 3D model.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed, it implements the three-dimensional model adjustment method as described in any one of claims 1 to 7.

Images