Lightweight design method and system of building three-dimensional model based on artificial intelligence

Abstract

The present invention relates to the technical field of architectural model design, and more specifically, to an artificial intelligence-based lightweight design method and system for architectural three-dimensional models. The present invention extracts geometric-semantic features of architectural models through multi-scale curvature filtering and semantic segmentation networks, constructs a fused feature vector matrix, utilizes an adaptive neural simplification network integrating a graph convolutional network and a dual-branch attention mechanism, performs differentiable resampling based on vertex importance weights, generates a simplified intermediate model, restores surface microstructures with the help of a generative adversarial network, corrects mesh holes with Delaunay triangulation, reconstructs key decorative components with non-uniform rational B-spline surfaces, and outputs a lightweight model. The method solves the problems of insufficient geometric feature retention and semantic information fragmentation in traditional technologies, realizes intelligent, efficient and lightweight design of architectural three-dimensional models, ensures visual fidelity while compressing data, and meets the needs of digital management of buildings throughout their life cycle.

Description

Technical Field

[0001] The present invention relates to the technical field of building model design, and in particular to a lightweight design method and system for building three-dimensional models based on artificial intelligence. Background Art

[0002] Architectural model design technology is an important technology. In modern construction projects and digital twin scenarios, architectural three-dimensional model lightweight technology is the core means to achieve efficient cross-platform model transmission, real-time rendering and collaborative design by reducing the amount of model data and retaining key geometric features.

[0003] With the increasing complexity of buildings and the advancement of smart city construction, building an intelligent lightweight model that takes into account both detail preservation and data compression is of key significance to reducing computing resource consumption and improving engineering collaboration efficiency. The traditional lightweight method based on fixed thresholds can no longer meet the optimization needs of high-precision models. However, the traditional lightweight technology of building three-dimensional models has the core problems of insufficient geometric feature preservation and semantic information fragmentation. The existing solution uses a fixed-ratio simplified grid, and then does not establish an intelligent evaluation model for curvature distribution and edge density. When processing models containing complex decorative components, the detail loss rate in the curvature mutation area increases, resulting in distortion of key structural features. In addition, the single geometric simplification method does not integrate the semantic information of components, and in curtain wall grids, carved components and other parts, the details are lost. Because the importance level of components cannot be identified, functional and decorative components are simplified equally, and the semantic integrity of the model is reduced. In addition, traditional technologies lack a detail recovery mechanism, which makes it impossible to effectively reconstruct the surface microstructure of the simplified model, resulting in a mismatch between the rendering effect and the original design deviation. This leads to visual discontinuity in the virtual reality scene, affecting the efficiency of design review. This geometric-semantic simplification and the lack of detail recovery capabilities ultimately make it difficult for traditional solutions to balance the data compression ratio and visual fidelity after the lightweighting of complex building models, and cannot meet the digital management needs of the entire building life cycle. To solve this technical problem, we provide an artificial intelligence-based lightweight design method and system for building 3D models. Summary of the Invention

[0004] The purpose of the present invention is to provide a lightweight design method and system for building three-dimensional models based on artificial intelligence to solve the problems raised in the above background technology.

[0005] 1. Because traditional technologies do not adequately preserve geometric features and result in a high loss of detail in complex components, this case uses multi-scale curvature filtering and edge density detection to extract a geometric-semantic feature matrix, which can preserve details in high-curvature areas and reduce distortion of key structural features.

[0006] 2. Since traditional methods fragment semantic information and fail to identify component importance, this case uses a semantic segmentation network to output component labels and integrates them into the feature matrix. This can simplify the components according to their importance and improve the semantic integrity of the model.

[0007] To achieve the above objectives, one of the objectives of the present invention is to provide a lightweight design method for a three-dimensional building model based on artificial intelligence, comprising the following steps:

[0008] S1. The building geometry feature parsing module processes the original building grid model, extracts the curvature distribution map through a multi-scale curvature filtering algorithm, generates an edge density heat map in combination with an edge density detector, and uses a semantic segmentation network to output component semantic labels to construct a feature vector matrix that integrates geometry and semantics.

[0009] S2. Input the feature vector matrix into an adaptive neural simplification network, which integrates a graph convolutional network and a dual-branch attention mechanism. The graph convolutional network generates an adjacency feature map based on the vertex K-nearest neighbor topology. The attention mechanism integrates the normal vector directionality constraint and the curvature gradient field, and generates a simplified intermediate model with a vertex retention rate greater than 85% through a differentiable resampling module.

[0010] S3, the residual detail compensation module receives the simplified intermediate model and the original normal map, uses a generative adversarial network to recover high-frequency features, the generator performs residual learning with an encoder-decoder structure, introduces a texture-geometry alignment loss function to reconstruct the surface microstructure, and the discriminator verifies visual consistency based on multi-scale structural similarity indicators and curvature continuity constraints;

[0011] S4. Use Delaunay triangulation to correct mesh holes, reconstruct key decorative components through non-uniform rational B-spline surfaces, and output a lightweight model.

[0012] As a further improvement of this technical solution, the workflow of the semantic segmentation network in S1 includes:

[0013] Convert the original mesh model into voxelized point cloud data, and index the voxelized point cloud data through the octree space;

[0014] Adopting the adaptive radius feature extraction module to calculate the local geometric descriptor, including normal deviation, curvature change rate and point density;

[0015] A graph neural network based on a bidirectional feature propagation mechanism fuses local descriptors with global context information to output component semantic labels.

[0016] As a further improvement of this technical solution, when constructing the eigenvector matrix in S1:

[0017] Perform Gaussian pyramid layered sampling on the curvature distribution map to extract multi-resolution curvature histogram features;

[0018] Generate edge trend distribution matrix through edge density heat map, and calculate main direction angle variance as characteristic component;

[0019] The semantic label is mapped into a one-hot encoded vector and concatenated with the geometric feature component to form a high-dimensional feature matrix.

[0020] As a further improvement of this technical solution, the execution logic of the dual-branch attention mechanism in S2 includes:

[0021] Topological attention branch: calculates the message passing weight between vertices based on the adjacency feature graph and aggregates k-hop neighbor features;

[0022] Geometric attention branch: It integrates the normal vector directionality constraint to establish a normal consistency score and combines it with the curvature gradient field to generate a geometric saliency map;

[0023] The topological and geometric attention outputs are dynamically weighted through the gated fusion module to generate vertex importance weight distribution.

[0024] As a further improvement of the present technical solution, the operation of the differentiable resampling module in S2 includes:

[0025] A probability sampling mask is generated according to the vertex importance weight distribution, edge collapse operation is performed on low-weight vertex clusters, the adjacency matrix is ​​updated synchronously, and a retention strategy based on surface differential coordinates is implemented for vertices in high curvature areas.

[0026] As a further improvement of this technical solution, the process of constructing the texture-geometry alignment loss function in S3 is as follows:

[0027] The tangent space normal vector field of the original normal map is extracted, and a normal vector reprojection module is established in the generator decoder. The cosine distance of the angle between the simplified model surface and the original normal field is calculated, and multi-scale structural similarity indicators are integrated to establish spatial correlation constraints.

[0028] As a further improvement of this technical solution, the discriminator verification of S3 includes:

[0029] Perform Laplace mesh smoothing on the input model;

[0030] Calculate the differential coordinate difference between the simplified model and the original model in the direction of the main curvature of the surface;

[0031] Detect surface G1 continuity breakpoints using curvature continuity constraints.

[0032] As a further improvement of this technical solution, the optimization logic of the Delaunay triangulation in S4 is as follows:

[0033] Identify grid hole boundary vertex chains and construct minimum spanning tree connection paths;

[0034] A local curvature-driven vertex insertion strategy is used to optimize the triangulation results;

[0035] Performs a conformal parametric unfolding of the newly added triangles to match the original texture coordinates.

[0036] As a further improvement of this technical solution, the reconstruction of the key decorative components of S4 includes:

[0037] The boundary lines of decorative components are identified based on semantic labels, and the isosurface control points are extracted using the marching cubes algorithm. The control mesh of the non-uniform rational B-spline surface is constructed and basis function fitting is performed.

[0038] A second object of the present invention is to provide a system for implementing any one of the above-mentioned artificial intelligence-based lightweight design methods for building three-dimensional models, comprising:

[0039] The multimodal feature parsing unit receives the building mesh model and texture data, implements multi-scale curvature filtering and edge density detection through a parallel computing architecture, generates a curvature distribution map and edge heat map, and simultaneously runs a semantic segmentation network to encode component semantic labels into feature vectors, which are then matrix-fused with geometric features to construct a multidimensional feature space.

[0040] The AI ​​lightweight decision-making unit integrates the core processing modules of the graph convolutional network and the dual-branch attention mechanism. The graph convolutional network analyzes the topological relationship of the K nearest neighbors of the vertex to construct an adjacency feature graph. The dual-branch attention mechanism simultaneously calculates the normal direction constraint weight and the curvature gradient field, generates a vertex importance distribution map through the gated fusion module, drives the differentiable resampling module to perform edge folding operations and high curvature area retention strategies, and outputs a simplified intermediate model that meets the vertex retention rate threshold.

[0041] The detail reconstruction and verification unit includes a residual detail generative adversarial network and an industrial-grade topology optimization module. The generator learns the residual features of the original normal map based on the codec architecture and restores the surface microstructure through the texture-geometry alignment mechanism. The discriminator verification subsystem detects the G1 continuity breakpoints of the surface and triggers iterative optimization. Finally, curvature-driven Delaunay triangulation is used to repair mesh holes, and non-uniform rational B-spline surfaces are combined to reconstruct key decorative components to generate lightweight output that meets the building information modeling standard.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] The present invention uses multi-scale curvature filtering and edge density detection, combined with a semantic segmentation network to construct a fusion geometry-semantics feature vector matrix. This can accurately identify the key features of the building model and the importance level of the components, avoid the loss of details of key parts such as complex decorative components, and improve the model's geometric feature retention rate and semantic integrity. It uses an adaptive neural simplification network that integrates a graph convolutional network and a dual-branch attention mechanism to perform differentiable resampling based on the vertex importance weight distribution. While ensuring that the vertex retention rate exceeds a preset threshold, it achieves efficient compression of the model data volume and solves the feature distortion problem caused by traditional fixed-ratio simplification. With the help of the residual detail compensation module of the generative adversarial network, the texture-geometry alignment loss function and multi-scale structural similarity verification are introduced, which can effectively restore surface microstructures such as stone texture, ensure the visual consistency of the simplified model with the original design, and reduce rendering deviation. Through Delaunay triangulation and non-uniform rational B-spline surface reconstruction, mesh holes can be corrected and key decorative components can be reconstructed, so that the lightweight model achieves a balance between data compression and visual fidelity, meeting the needs of digital management of the entire building life cycle. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is the overall workflow diagram of the present invention;

[0045] Figure 2 It is a schematic diagram of the overall structure of the present invention;

[0046] The meaning of each number in the figure is:

[0047] 1. Multimodal feature analysis unit; 2. AI lightweight decision-making unit; 3. Detail reconstruction and verification unit. DETAILED DESCRIPTION

[0048] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0049] See also Figure 1 As shown, one of the purposes of this embodiment is to provide a lightweight design method for a building three-dimensional model based on artificial intelligence, comprising the following steps:

[0050] S1. The building geometry feature parsing module processes the original building grid model, extracts the curvature distribution map through a multi-scale curvature filtering algorithm, generates an edge density heat map in combination with an edge density detector, and uses a semantic segmentation network to output component semantic labels to construct a feature vector matrix that integrates geometry and semantics.

[0051] S2. Input the feature vector matrix into an adaptive neural simplification network, which integrates a graph convolutional network and a dual-branch attention mechanism. The graph convolutional network generates an adjacency feature map based on the vertex K-nearest neighbor topology. The attention mechanism integrates the normal vector directionality constraint and the curvature gradient field, and generates a simplified intermediate model with a vertex retention rate greater than 85% through a differentiable resampling module.

[0052] S3, the residual detail compensation module receives the simplified intermediate model and the original normal map, uses a generative adversarial network to recover high-frequency features, the generator performs residual learning with an encoder-decoder structure, introduces a texture-geometry alignment loss function to reconstruct the surface microstructure, and the discriminator verifies visual consistency based on multi-scale structural similarity indicators and curvature continuity constraints;

[0053] S4. Use Delaunay triangulation to correct mesh holes, reconstruct key decorative components through non-uniform rational B-spline surfaces, and output a lightweight model.

[0054] Since the curvature distribution and edge density of the building model are key indicators reflecting the geometric features, a multi-scale curvature filtering algorithm is first performed on the original building mesh model. That is, Gaussian kernels of different sizes are used to calculate the Gaussian curvature and average curvature of the model vertices to generate a multi-scale curvature distribution map. The map displays the curvature value of each vertex in pseudo-color coding. At the same time, all edges of the model are traversed through the edge density detector, and the normal angle between the adjacent faces of each edge is calculated. When the angle exceeds the preset threshold, it is determined to be an edge. The number of edges per unit area is counted to generate an edge density heat map. The darker the color in the heat map, the higher the edge density. Multi-scale curvature filtering can capture geometric features of different degrees of precision. Edge density detection can locate the model structure boundary and improve the integrity of geometric feature extraction. In order to obtain the semantic information of the model, the semantic segmentation network is used to process the original mesh model. The workflow of the semantic segmentation network in S1 includes:

[0055] The original mesh model is converted into voxelized point cloud data, and the voxelized point cloud data is indexed through the octree space. The octree recursively divides the three-dimensional space into 8 sub-cubes, and each node stores the point cloud data in the corresponding space. The indexing mechanism can accelerate neighborhood queries, improve the efficiency of point cloud neighbor search, and lay the foundation for subsequent feature calculations. The adaptive radius feature extraction module is used to calculate the local geometric descriptor. For each point cloud point, the search radius is dynamically adjusted according to its neighborhood point density. The calculation content includes normal deviation, curvature change rate and point density, where the normal deviation is the angle between the current point normal and the neighborhood average normal, the curvature change rate is the difference between the current point curvature and the neighborhood average curvature, and the point density is the number of point clouds per unit volume. These descriptors can effectively characterize the local geometric characteristics of the point cloud and provide underlying features for semantic segmentation. The graph neural network with feature propagation mechanism fuses local descriptors with global context information. The graph neural network uses point clouds as nodes and adjacency relationships as edges to build a graph structure. The initial feature of each node is a local geometric descriptor. The network aggregates k-hop neighbor features through forward propagation to obtain global context information. During backpropagation, the global features are backpropagated to each node, the local feature representation is corrected, and finally the component semantic label of each point is output through the softmax classifier. The label category is preset according to the construction industry standard, providing a semantic basis for the fusion of geometric and semantic features. In order to construct a high-dimensional feature matrix that integrates geometric features and semantic information, after completing the extraction of curvature distribution map and edge density heat map and the generation of semantic labels, it is necessary to perform structured integration of multi-source features. The specific implementation method is as follows: when constructing the feature vector matrix in S1:

[0056] Considering that the curvature characteristics of the building model at different scales have hierarchical characteristics, the curvature distribution map is sampled by Gaussian pyramid layer to capture multi-scale geometric details. First, the original curvature distribution map is used as the bottom layer of the pyramid, that is, the 0th layer. The bottom layer image is convolutionally blurred using a 5×5 Gaussian kernel and downsampled by 2×2 to generate the first layer. Similarly, a 4-layer pyramid is constructed with resolutions of 1 / 1, 1 / 2, 1 / 4, and 1 / 8, respectively. For each layer of curvature map, the curvature value is normalized to the interval [0,1] and then divided into 32 bins to calculate the histogram. The mean, variance, skewness, and kurtosis of each layer's histogram are extracted as statistical features. Each layer has 4 dimensions, and finally a 4-layer × 4-dimensional = 16-dimensional multi-resolution curvature histogram feature vector is formed, which effectively distinguishes the curvature distribution differences between carved components and flat walls.

[0057] Since edge direction can reflect the structural regularity of the model surface, an edge direction distribution matrix is ​​generated through the edge density heat map to quantify this feature. First, the model surface is divided into M×N equal-area grids. The edge edges in each grid are linearly fitted by Hough transform to obtain the main direction angle of the edge, which ranges from [0 to π). The edge is determined by the normal angle being greater than 60 degrees. Then, an M×N edge direction distribution matrix is ​​constructed, with the matrix elements being the main direction angles of the corresponding grids. The variance of all elements in the matrix is ​​calculated as the feature component. The smaller the variance, the more regular the edge direction. This feature can effectively characterize the structural regularity of the model surface and improve the accuracy of the structured description of edge features.

[0058] In order to integrate semantic information into the feature matrix, one-hot encoding is used to convert discrete semantic labels into vector representations. According to the category of building components, each semantic label is mapped to a 20-dimensional one-hot vector. For example, the 5th bit in the vector corresponding to "decorative component" is 1, and the rest are 0. The vector is spliced ​​with the geometric feature component according to the dimension. The geometric feature component is 16-dimensional curvature feature + 1-dimensional edge direction variance + other geometric features to form a high-dimensional feature matrix. Each row of the matrix corresponds to a model vertex, and the columns are semantic coding, multi-resolution curvature feature, and edge direction feature. For example, in the feature vector of a vertex, the semantic coding part shows "decorative component", the curvature feature shows a high variance value, and the edge direction variance shows a high value, indicating that the vertex belongs to the high curvature area of ​​complex decorative components. The tensor splicing operation makes the semantic information and geometric features form a complementary representation, providing a multi-dimensional basis for vertex importance evaluation in subsequent lightweight processing.

[0059] The eigenvector matrix is ​​input into the adaptive neural simplification network, which integrates a graph convolutional network and a dual-branch attention mechanism. The execution logic of the dual-branch attention mechanism in S2 includes: in order to accurately evaluate the importance of vertices in the process of lightweighting the building 3D model, the dual-branch attention mechanism in S2 performs weighted processing on the eigenvector matrix based on the topological and geometric dimensions. The specific implementation is as follows:

[0060] Topological attention branch: Calculates the message passing weight between vertices based on the adjacency feature graph and aggregates k-hop neighbor features. Since the structural importance of the model vertex is closely related to its neighborhood topological relationship, the topological attention branch calculates the message passing weight between vertices based on the adjacency feature graph. First, the graph convolutional network is used to parse the K-nearest neighbor topological relationship of the vertex to construct the adjacency feature graph. The adjacency feature graph of each vertex contains the feature vectors of its K nearest neighbor vertices. Then, the message passing weight between vertex i and adjacent vertex j is calculated using the attention weight formula. This weight reflects the similarity of the features between the vertices. Then, the k-hop neighbor features are aggregated according to the weight, so that the features of each vertex contain local topological structure information, effectively identifying topological key parts such as load-bearing columns.

[0061] Geometric attention branch: The normal vector directional constraint is integrated to establish the normal consistency score, and the curvature gradient field is combined to generate a geometric saliency map. Given the key influence of geometric features on the visual fidelity of the model, the geometric attention branch integrates the normal vector directional constraint and the curvature gradient field to generate a geometric saliency map. First, the normal consistency score of the vertex is calculated. For each vertex i, the cosine value of the angle between its normal direction and the normal vector of the adjacent vertex is calculated, and the average is taken as the normal consistency score. The lower the score, the more drastic the change in the normal direction. Then, the geometric saliency map is generated by combining the curvature gradient field. The curvature gradient field is calculated by the curvature difference of adjacent vertices. For each vertex, its curvature value and curvature gradient are normalized and weighted summed to obtain the geometric saliency value, ensuring the preservation of geometric features such as curtain wall grid lines.

[0062] The gated fusion module dynamically weights the topological and geometric attention outputs to generate vertex importance weight distribution. In order to achieve the organic fusion of topological and geometric features, the gated fusion module dynamically weights the outputs of the two attentions. The gated fusion module uses the sigmoid activation function to calculate the fusion weight. ;in, and are the output vectors of topological attention and geometric attention respectively, and is a learnable parameter, and the vertex importance weight distribution after fusion is ; For example, when processing decorative components, the geometric attention weight Automatically increase to highlight the importance of geometric features; when processing structural columns, topological attention weight It emphasizes the importance of topological structure, provides accurate weight basis for subsequent differentiable resampling, and generates a simplified intermediate model with a vertex retention rate greater than 85% through the differentiable resampling module.

[0063] Among them, the differentiable resampling module operations include:

[0064] To achieve intelligent simplification of the model based on vertex importance weight distribution, the differentiable resampling module needs to perform a series of operations including probabilistic sampling mask generation, simplification of low-weight vertex clusters, and preservation of high curvature areas. The specific implementation is as follows:

[0065] A probability sampling mask is generated based on the vertex importance weight distribution, and edge folding operations are performed on low-weight vertex clusters. The adjacency matrix is ​​updated synchronously, and a retention strategy based on surface differential coordinates is implemented for vertices in high curvature areas. Since the vertex importance weight distribution has characterized the contribution of each vertex to the model structure and vision, the weight value is first normalized to the interval [0,1] to generate a probability sampling mask. For each vertex, its sampling probability is equal to the normalized weight value. This mask converts the abstract importance weight into an executable sampling probability, making the model simplification process random and interpretable. The vertex retention strategy generated by the mask is more in line with the subtle differences in weight distribution, improving the rationality of sampling. In order to efficiently reduce the number of model vertices, edge folding operations are performed on low-weight vertex clusters whose sampling probability is lower than the preset threshold. First, the low-weight vertices are divided into several clusters using the DBSCAN clustering algorithm. For each cluster, the vertex with the lowest weight in the cluster is selected as the folding target vertex, and other vertices in the cluster are folded to this vertex. The error cost of edge folding is calculated during the folding process. To ensure that the simplification error is within an acceptable range, after edge collapse is completed, the adjacency matrix is ​​updated synchronously, that is, the adjacency relationship of the collapsed vertex is deleted, its adjacent vertices are pointed to the target vertex, and the K-nearest neighbor topological relationship of the target vertex is recalculated. This operation effectively reduces the number of model vertices while maintaining the basic shape of the model through error control. Considering the importance of high curvature areas, such as building corners and decorative carvings, to model details, a retention strategy based on surface differential coordinates is implemented for vertices in high curvature areas with a sampling probability higher than a preset threshold. First, the surface differential coordinates of each vertex are calculated, that is, the tangent vector and normal vector of the vertex are obtained through local surface fitting to construct a local coordinate system. The differential coordinates can represent the positional relationship of the vertex on the local surface. During the resampling process, the differential coordinate information of high curvature vertices is forced to be retained. Even if they are marked as low weight in the edge collapse operation, they are prevented from being collapsed through coordinate constraints. This strategy ensures the preservation of details in high curvature areas and avoids the problem of detail loss caused by oversimplification of high curvature areas in traditional simplification methods.

[0066] The residual detail compensation module receives the simplified intermediate model and the original normal map, and uses a generative adversarial network to restore high-frequency features. The generator performs residual learning with an encoder-decoder structure and introduces a texture-geometry alignment loss function to reconstruct the surface microstructure. Since the simplified intermediate model will lose high-frequency details such as stone texture and decorative carvings after lightweighting, the residual detail compensation module uses a generative adversarial network to learn the residual features between the original model and the simplified model. The generator adopts an encoder-decoder structure. The encoder takes the mesh vertex coordinates of the simplified intermediate model and the simplified normal map as input, where the mesh vertex coordinates are obtained through 3D modeling software. The normal map is a texture image that records the surface normal direction and can be exported from the original model. Features are extracted through a multi-layer convolutional neural network. The decoder generates residual details based on these features and fuses them with the geometric information of the simplified model to reconstruct the surface microstructure. To ensure that the generated surface microstructure is consistent with the original model in texture and geometry, a texture-geometry alignment loss function needs to be constructed. The construction process of the texture-geometry alignment loss function in S3 is as follows:

[0067] Extract the tangent space normal vector field of the original normal map, establish a normal vector reprojection module in the generator decoder, calculate the cosine distance of the angle between the simplified model surface and the original normal field, and integrate multi-scale structural similarity indicators to establish spatial correlation constraints. Each pixel in the original normal map stores the normal vector of the corresponding point on the model surface. In order to convert the normal vector from the three-dimensional space of the model to the local texture space, it is necessary to extract the tangent space normal vector field. First, for each triangular face of the original model, calculate its tangent vector and sub-tangent vector. The tangent vector and sub-tangent vector are calculated by the partial derivatives of the vertex texture coordinates to construct the tangent space coordinate system. Then, the normal vector in the normal map is converted from the model coordinate system to the tangent space. The tangent space normal vector field is obtained from the spatial coordinate system, which can accurately reflect the change of normal direction in the texture space. In the decoder of the generator, a normal vector reprojection module is established. This module reprojects the normal vector corresponding to the generated residual details from the tangent space to the three-dimensional space of the model and fuses it with the normal vector of the simplified model. Specifically, for the generated residual height map, its corresponding normal vector is obtained by differential calculation, that is, the gradient direction of the height map reflects the normal change. These normal vectors are then converted back from the tangent space to the model coordinate system and weightedly fused with the normal vector of the simplified model so that the generated details dominate the normal direction. The cosine distance of the angle between the simplified model surface and the original normal field is calculated. The formula is ;in are the normal vectors of the simplified model and the original model respectively. The smaller the distance, the closer the normal direction. At the same time, a multi-scale structural similarity index is introduced to evaluate the similarity between the generated surface microstructure and the original model from different resolutions. Finally, the angle cosine distance and the multi-scale structural similarity index are weighted and fused to establish a spatial correlation constraint to form a texture-geometry alignment loss function. ;in is the weight of each scale, is the cosine distance of the included angle, The loss function ensures that the generated surface microstructure is highly consistent with the original model in terms of normal direction and overall structure. The discriminator verifies visual consistency based on the multi-scale structural similarity index and curvature continuity constraint. The discriminator in S3 needs to perform multi-dimensional verification on the input model. The verification of the S3 discriminator includes:

[0068] Perform Laplace mesh smoothing on the input model. Since the surface of the generated simplified model may contain high-frequency noise introduced by residual learning, it needs to be smoothed by Laplace mesh first. The specific operation is:

[0069] For each vertex in the model, the average position of its adjacent vertices is calculated and used as the new position of the vertex. This process can effectively eliminate small fluctuations on the model surface by iterating 3 to 5 times, making the mesh smoother and providing clean geometric input for subsequent curvature calculations. In order to quantify the geometric differences between the simplified model and the original model, it is necessary to calculate the differential coordinate difference between the two in the direction of the principal curvature of the surface. First, the principal curvature and the corresponding principal direction vector of each vertex are calculated for the original model and the simplified model respectively. The principal curvature is solved by fitting a quadratic equation to the local surface. The principal direction is the eigenvector corresponding to the curvature. Then, the three-dimensional coordinates of the vertex are converted into differential coordinates based on the principal curvature direction. The vertex is taken as the origin, and the corresponding principal direction vector and the normal vector constitute a local coordinate system. The coordinates of the vertex neighborhood points in this coordinate system are the differential coordinates. The L2 norm of the differential coordinate differences of all corresponding vertices is calculated, and the mean is taken as the geometric difference index. This index can sensitively reflect the change of the surface shape. In order to ensure a smooth transition of the model surface, the curvature continuity constraint is required to detect the G1 continuity breakpoint of the surface, where the G1 continuity of the surface is tangent continuity. First, the model is calculated. The principal curvature direction vectors of the vertices at both ends of each edge are taken, and the direction corresponding to the maximum principal curvature is taken, and the angle between the two vectors is calculated. When the angle exceeds the preset threshold, the edge is determined to be a G1 continuity breakpoint. At the same time, combined with the curvature value continuity constraint, if the principal curvature value difference between the vertices at both ends of the edge exceeds the threshold, it is also determined to be a breakpoint. The two types of breakpoints are superimposed to generate a G1 continuity breakpoint distribution map. The highlighted breakpoints in the map are the areas that need to be repaired. This detection method can accurately identify discontinuous areas on the model surface, provide a clear target for subsequent mesh optimization, and ensure that there are no visual faults when the model is rendered. Step S4 uses Delaunay triangulation to correct the mesh holes in the area that needs to be repaired, and reconstructs key decorative components through non-uniform rational B-spline surfaces to output a lightweight model.

[0070] To correct mesh holes in the simplified model and ensure geometric accuracy and texture continuity in the repaired area, S4 uses an optimized Delaunay triangulation method. The optimization logic of the Delaunay triangulation in S4 is as follows:

[0071] Identify the boundary vertex chain of mesh holes and construct the minimum spanning tree connection path. Since the boundary vertex chain of mesh holes is the basis of triangulation, first traverse all edges of the model to identify the edges that belong only to a single triangle, namely the boundary edges, and connect these boundary edges in sequence into a closed vertex chain, namely the hole boundary. Then, build a minimum spanning tree based on the boundary vertex chain, using the vertices in the vertex chain as nodes and the Euclidean distance between vertices as edge weights. Use the Prim algorithm to generate a minimum spanning tree connecting all boundary vertices to ensure that the total length of the edges in the tree is the shortest. This step provides the optimal connection path for hole filling, making the initial mesh structure of subsequent triangulation more reasonable. In order to make the triangulation result conform to the geometric characteristics of the model surface, a local curvature-driven vertex insertion strategy is adopted. First, calculate the curvature value of the hole boundary vertex chain and the surrounding area, use the quadratic surface fitting method to calculate the Gaussian curvature of each point, determine the area with drastic curvature changes, such as the area with a curvature value greater than 0.5, and then , insert new vertices in the area with high curvature. For each boundary edge, if the average curvature of its two end vertices is greater than the preset threshold, insert a new vertex at the midpoint of the edge and calculate the curvature value of the new vertex to ensure that the position of the inserted vertex conforms to the local surface morphology. After inserting the new vertex, use the Delaunay triangulation algorithm to generate a triangular mesh for the hole area. The Delaunay triangulation algorithm ensures that the circumscribed circle of any triangle does not contain other vertices. It optimizes the triangle quality by maximizing the minimum angle criterion to avoid narrow triangles and effectively avoids the surface distortion problem caused by traditional uniform vertex insertion. In order to ensure the texture continuity of the repaired area, the newly added triangular facets are subjected to shape-preserving parameterized expansion. First, obtain the texture coordinates of the original model in the area around the hole and determine the texture mapping relationship. Then, for the newly added triangular facets, a shape-preserving parameterization method based on the Laplace equation is used to map the triangular facets in the three-dimensional space to the two-dimensional texture space while keeping the angle unchanged. The specific steps are as follows:

[0072] Establish the position constraint equation of the triangle vertex ;in The texture coordinates are obtained by solving the Poisson equation to obtain the texture coordinates of the newly added vertices, so that the texture coordinates of the newly added facets can smoothly transition with the surrounding area. Finally, the texture coordinates are optimized and the energy function minimization method is used to adjust the texture coordinates to make the texture mapping more conformal, avoid texture faults or distortions in the repaired area, and ensure the consistency of the overall visual effect of the model, laying a solid foundation for the subsequent reconstruction of key decorative components and the output of lightweight models. The key decorative components are reconstructed through non-uniform rational B-spline surfaces and the lightweight model is output.

[0073] The reconstruction of key decorative components of S4 includes:

[0074] The boundary lines of decorative components are identified based on semantic labels, and the isosurface control points are extracted using the marching cubes algorithm. The control mesh of the non-uniform rational B-spline surface is constructed and basis function fitting is performed.

[0075] To achieve high-precision reconstruction of key decorative components and improve the visual fidelity and semantic integrity of the lightweight model, S4 performs specialized reconstruction of decorative components based on semantic information and geometric features. The specific implementation is as follows:

[0076] Since the semantic labels of the decorative components have been generated in S1 through the semantic segmentation network, first, all the vertices of the target component are screened according to the semantic labels. For example, all the vertices with the semantic label "carving" are extracted from the model to form a point cloud set of the decorative component. Then, the boundary line of the decorative component is identified by the boundary tracing algorithm. For each point cloud point, the normal vector angle of its adjacent vertices is calculated. When the angle exceeds the preset threshold, it is determined to be a boundary point. The boundary points are connected in sequence to form a closed boundary line. For example, the point cloud boundary points of a carved component are obtained by Delaunay triangulation. Corner subdivision is used to construct topological relationships, and then continuous boundary line loops are formed by tracing along the triangle edges, which avoids the problem of misidentification of complex decorative components by traditional geometric threshold methods. In order to convert the point cloud data of decorative components into a parametric surface model, the marching cube algorithm is used to extract isosurface control points. First, the point cloud data of the decorative components is voxelized to construct a three-dimensional voxel grid. Then, an isosurface threshold is defined. The isosurface threshold takes the average distance from the point cloud to the fitting surface. Each voxel cube is traversed, and according to the intersection of the cube vertex and the isosurface, a triangular facet approximation to the isosurface is generated. For example, in the voxel grid of a certain carved component, when the voxel vertex is on one side of the isosurface, it is marked as 1, and the other side is marked as 0. By searching the preset 14 cube vertex state combination table, the corresponding triangle patch vertices, namely the isosurface control points, are generated. The isosurface control points generated by this algorithm can accurately describe the surface morphology of the decorative component and provide a high-quality control point set for subsequent surface fitting. In order to achieve parametric representation and smooth reconstruction of the decorative component, a non-uniform rational B-spline surface is constructed based on the extracted isosurface control points. First, the isosurface control points are topologically mapped. The relationship is organized into a grid-like control vertex array. The density of the control grid is adjusted according to the curvature distribution of the component, that is, control vertices are added in the area with high curvature and control vertices are reduced in the area with low curvature to form a non-uniform control grid. For example, the edge curvature of the petals of a certain relief decoration is high, the control grid vertex spacing is 0.01 meters, and the plane area spacing is 0.05 meters. Then, the order of the non-uniform rational B-spline surface is determined, and the order is 3, that is, cubic B-spline, and a weight is assigned to each control vertex. Finally, the surface equation is calculated by fitting the basis function: for any point on the surface , whose coordinates are determined by the control vertex , weight and B-spline basis functions 、 Determine the formula ;in, are parameter coordinates, The order is set. By adjusting the control vertex positions and weights, the fitting error between the rational B-spline surface and the isosurface control points is made less than 0.001 meters, achieving a balance between detail retention and lightweight. The reconstructed rational B-spline surface model can not only accurately restore complex shapes such as carvings and reliefs, but also realize flexible editing of the model through parametric control. Compared with the traditional mesh reconstruction method, it has significantly improved visual fidelity and data efficiency, providing a high-precision lightweight model foundation for the digital management of the entire life cycle of buildings.

[0077] A second object of the present invention is to provide a system for implementing a lightweight design method for a three-dimensional building model based on artificial intelligence, comprising:

[0078] The multimodal feature parsing unit 1 receives the building grid model and texture data, implements multi-scale curvature filtering and edge density detection through a parallel computing architecture, generates a curvature distribution map and an edge heat map, and simultaneously runs a semantic segmentation network to encode component semantic labels into feature vectors, which are then matrix-fused with geometric features to construct a multidimensional feature space.

[0079] The AI ​​lightweight decision unit 2 integrates the core processing modules of the graph convolutional network and the dual-branch attention mechanism. The graph convolutional network analyzes the topological relationship of the K nearest neighbors of the vertex to construct an adjacency feature graph. The dual-branch attention mechanism simultaneously calculates the normal direction constraint weight and the curvature gradient field, generates a vertex importance distribution map through the gated fusion module, drives the differentiable resampling module to perform edge folding operations and high curvature area retention strategies, and outputs a simplified intermediate model that meets the vertex retention rate threshold.

[0080] The detail reconstruction and verification unit 3 includes a residual detail generative adversarial network and an industrial-grade topology optimization module. The generator learns the residual features of the original normal map based on the codec architecture and restores the surface microstructure through the texture-geometry alignment mechanism. The discriminator verification subsystem detects the continuity breakpoints of the surface G1 and triggers iterative optimization. Finally, curvature-driven Delaunay triangulation is used to repair mesh holes, and non-uniform rational B-spline surfaces are combined to reconstruct key decorative components to generate lightweight output that meets the building information modeling standard.

[0081] The present invention extracts the geometric-semantic features of the building model through multi-scale curvature filtering and semantic segmentation network, constructs a fused feature vector matrix, utilizes an adaptive neural simplification network integrating graph convolutional network and dual-branch attention mechanism, performs differentiable resampling based on vertex importance weights, generates a simplified intermediate model, restores the surface microstructure with the help of generative adversarial network, combines Delaunay triangulation to correct mesh holes, uses non-uniform rational B-spline surfaces to reconstruct key decorative components, and outputs a lightweight model. This solves the problems of insufficient geometric feature retention and semantic information fragmentation in traditional technologies, realizes intelligent, efficient and lightweight three-dimensional building models, ensures visual fidelity while compressing data, and meets the needs of digital management of buildings throughout their life cycle.

[0082] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A lightweight design method for building three-dimensional models based on artificial intelligence, characterized by: include: S1. The building geometry feature parsing module processes the original building grid model, extracts the curvature distribution map through a multi-scale curvature filtering algorithm, generates an edge density heat map in combination with an edge density detector, and uses a semantic segmentation network to output component semantic labels to construct a feature vector matrix that integrates geometry and semantics. S2. Input the feature vector matrix into an adaptive neural simplification network, which integrates a graph convolutional network and a dual-branch attention mechanism. The graph convolutional network generates an adjacency feature map based on the vertex K-nearest neighbor topology. The attention mechanism integrates the normal vector directionality constraint and the curvature gradient field, and generates a simplified intermediate model with a vertex retention rate greater than 85% through a differentiable resampling module. S3, the residual detail compensation module receives the simplified intermediate model and the original normal map, uses a generative adversarial network to recover high-frequency features, the generator performs residual learning with an encoder-decoder structure, introduces a texture-geometry alignment loss function to reconstruct the surface microstructure, and the discriminator verifies visual consistency based on multi-scale structural similarity indicators and curvature continuity constraints; S4. Use Delaunay triangulation to correct mesh holes, reconstruct key decorative components through non-uniform rational B-spline surfaces, and output a lightweight model.

2. The method for lightweight design of a three-dimensional building model based on artificial intelligence according to claim 1, characterized in that: The workflow of the semantic segmentation network in S1 includes: Convert the original mesh model into voxelized point cloud data, and index the voxelized point cloud data through the octree space; Adopting the adaptive radius feature extraction module to calculate the local geometric descriptor, including normal deviation, curvature change rate and point density; A graph neural network based on a bidirectional feature propagation mechanism fuses local descriptors with global context information to output component semantic labels.

3. The lightweight design method of building three-dimensional model based on artificial intelligence according to claim 1 is characterized in that: When constructing the eigenvector matrix in S1: Perform Gaussian pyramid layered sampling on the curvature distribution map to extract multi-resolution curvature histogram features; Generate edge trend distribution matrix through edge density heat map, and calculate main direction angle variance as characteristic component; The semantic label is mapped into a one-hot encoded vector and concatenated with the geometric feature component to form a high-dimensional feature matrix.

4. The method for lightweight design of a three-dimensional building model based on artificial intelligence according to claim 1, characterized in that: The execution logic of the dual-branch attention mechanism in S2 includes: Topological attention branch: calculates the message passing weight between vertices based on the adjacency feature graph and aggregates k-hop neighbor features; Geometric attention branch: It integrates the normal vector directionality constraint to establish a normal consistency score and combines it with the curvature gradient field to generate a geometric saliency map; The topological and geometric attention outputs are dynamically weighted through the gated fusion module to generate vertex importance weight distribution.

5. The lightweight design method of building three-dimensional model based on artificial intelligence according to claim 1 is characterized in that: The differentiable resampling module operation in S2 includes: A probability sampling mask is generated according to the vertex importance weight distribution, edge collapse operation is performed on low-weight vertex clusters, the adjacency matrix is ​​updated synchronously, and a retention strategy based on surface differential coordinates is implemented for vertices in high curvature areas.

6. The lightweight design method of building three-dimensional model based on artificial intelligence according to claim 1 is characterized in that: The construction process of the texture-geometry alignment loss function in S3 is as follows: The tangent space normal vector field of the original normal map is extracted, and a normal vector reprojection module is established in the generator decoder. The cosine distance of the angle between the simplified model surface and the original normal field is calculated, and multi-scale structural similarity indicators are integrated to establish spatial correlation constraints.

7. The lightweight design method of building three-dimensional model based on artificial intelligence according to claim 1 is characterized in that: The S3 discriminator verification includes: Perform Laplace mesh smoothing on the input model; Calculate the differential coordinate difference between the simplified model and the original model in the direction of the main curvature of the surface; Detect surface G1 continuity breakpoints using curvature continuity constraints.

8. The method for lightweight design of building three-dimensional models based on artificial intelligence according to claim 1, characterized in that: The optimization logic of Delaunay triangulation in S4 is as follows: Identify grid hole boundary vertex chains and construct minimum spanning tree connection paths; A local curvature-driven vertex insertion strategy is used to optimize the triangulation results; Performs a conformal parametric unfolding of the newly added triangles to match the original texture coordinates.

9. The method for lightweight design of building three-dimensional models based on artificial intelligence according to claim 1, characterized in that: The reconstruction of the key decorative components of the S4 includes: The boundary lines of decorative components are identified based on semantic labels, and the isosurface control points are extracted using the marching cubes algorithm. The control mesh of the non-uniform rational B-spline surface is constructed and basis function fitting is performed.

10. A system for implementing the artificial intelligence-based lightweight design method for a three-dimensional building model according to any one of claims 1 to 9, characterized in that: include: The multimodal feature parsing unit (1) receives the building grid model and texture data, implements multi-scale curvature filtering and edge density detection through a parallel computing architecture, generates a curvature distribution map and an edge heat map, and simultaneously runs a semantic segmentation network to encode component semantic labels into feature vectors, and performs matrix fusion with geometric features to construct a multidimensional feature space; The AI ​​lightweight decision-making unit (2) integrates the core processing module of the graph convolutional network and the dual-branch attention mechanism, in which the graph convolutional network analyzes the topological relationship of the K nearest neighbors of the vertex to construct the adjacency feature graph, and the dual-branch attention mechanism simultaneously calculates the normal direction constraint weight and the curvature gradient field, generates the vertex importance distribution map through the gated fusion module, drives the differentiable resampling module to perform the edge folding operation and the high curvature area retention strategy, and outputs a simplified intermediate model that meets the vertex retention rate threshold; The detail reconstruction and verification unit (3) includes a residual detail generative adversarial network and an industrial-grade topology optimization module. The generator learns the residual features of the original normal map based on the codec architecture and restores the surface microstructure through the texture-geometry alignment mechanism. The discriminator verification subsystem detects the G1 continuity breakpoints of the surface and triggers iterative optimization. Finally, the curvature-driven Delaunay triangulation is used to repair the mesh holes, and the non-uniform rational B-spline surface is combined to reconstruct the key decorative components to generate a lightweight output that meets the building information modeling standard.

Images