A method for simplifying a region of a three-dimensional building model based on deep learning
By extracting semantic features using a deep learning-based graph convolutional neural network and incorporating them into a surface collapse algorithm, the problem of low structural similarity and decreased visual fidelity after simplification of 3D building models in existing technologies is solved. This achieves efficient and accurate region simplification, improving the visual fidelity and structural integrity of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUANGXI ZHUANG AUTONOMOUS REGION NATURAL RESOURCES REMOTE SENSING INST
- Filing Date
- 2025-09-29
- Publication Date
- 2026-08-04
AI Technical Summary
Existing methods for simplifying 3D building models cannot accurately identify key areas when processing multi-source data, resulting in low structural similarity between the simplified model and the original model and a severe decrease in visual fidelity.
A graph convolutional neural network based on deep learning is used to extract semantic features of the 3D building model, generate regional importance scores, and integrate them into the error measurement calculation of the surface collapse algorithm. A dynamic collapse priority queue is constructed, and the process is optimized and simplified through iterative collapse operation to maintain the structural similarity and visual fidelity of key areas.
Effectively distinguish between critical and non-critical areas of the building, avoid oversimplification of critical areas, improve the visual fidelity and structural integrity of the simplified model, ensure the efficiency and relevance of the simplification process, and maintain the geometric topology of the model.
Smart Images
Figure CN121330217B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer graphics processing technology, specifically to a method for region simplification of 3D architectural models based on deep learning. Background Technology
[0002] Existing methods for simplifying 3D building models are limited by the mechanism of traditional surface collapse algorithms, which rely solely on geometric information to drive simplification priorities, when processing multi-source data such as textured triangular mesh models generated by oblique photogrammetry and LiDAR point cloud reconstruction models to generate high-detail building models. This results in the inability to accurately identify key areas of the building, such as doors, windows, and roof edges, as well as non-key areas. Consequently, key areas are oversimplified, ultimately leading to low structural similarity between the simplified model and the original model and a severe decrease in visual fidelity. Summary of the Invention
[0003] (a) Technical problems to be solved
[0004] To address the shortcomings of existing technologies, this invention provides a region simplification method for architectural 3D models based on deep learning, which solves the problems of low structural similarity and severe decrease in visual fidelity between the simplified model and the original model.
[0005] (II) Technical Solution
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] A deep learning-based method for region simplification of 3D building models includes the following steps:
[0008] Step 1: Extract semantic features from the input 3D building mesh model using a deep learning network to generate a regional importance score for each vertex, which represents the degree of simplification of the region where the vertex is located;
[0009] Step 2: The importance score of the region is used as a weight and incorporated into the error metric calculation of the surface collapse algorithm to construct a dynamic collapse priority queue, in which the lower the importance score of a vertex, the higher its collapse priority in the queue.
[0010] Step 3: Perform iterative collapse operation based on the dynamic collapse priority queue. The iterative collapse operation includes: selecting the vertex with the highest collapse priority in the priority queue to perform collapse, updating the error metric value of the adjacent vertices affected by the collapse, and readjusting their priority in the queue.
[0011] Step 4: When the preset iteration termination conditions are met, stop the collapse operation and output the simplified 3D building model; the iteration termination conditions include: the number of simplified model patches reaches the target number threshold, and the structural similarity index between the key areas of the model and the corresponding areas of the original model reaches the preset fidelity threshold.
[0012] Preferably, the deep learning network is a graph convolutional neural network; the input of the graph convolutional neural network includes the following features of each vertex: three-dimensional spatial coordinates, normal vector direction, Gaussian curvature scalar value representing the curvature of the surface, and the number of adjacent vertices directly connected to that vertex;
[0013] The output of the graph convolutional neural network is an importance score for each vertex, which is a value between 0 and 1; where an importance score of 0 means that the vertex can be completely collapsed, and an importance score of 1 means that the vertex cannot be collapsed.
[0014] Preferably, the error metric function for incorporating the regional importance score is defined as:
[0015] Part 1: Calculate the basic error value of the vertex in the traditional face collapse algorithm. This basic error value is calculated based on the sum of the quadratic forms of the geometric errors of all the triangular faces adjacent to the vertex.
[0016] Part Two: The regional importance score is multiplied by the base error value as a penalty weight, so that the error value of vertices with high importance scores is significantly increased, thus reducing their priority of being collapsed;
[0017] Part 3: Add a penalty term based on the Euclidean norm of the change in the normal vector before and after vertex collapse. This penalty term is multiplied by a preset weight coefficient, which ranges from 0.2 to 0.5.
[0018] The final error metric is the result of multiplying the first part by the weight of the regional importance score, plus the result of the normal change penalty term in the third part. The base error value is calculated using the traditional quadratic error metric (QEM), and the specific calculation formula is as follows:
[0019]
[0020] Where, faces(υ) represents the set of all triangles adjacent to vertex υ;
[0021] It is the quadratic error matrix of the patch f (where q f =[a,b,c,d] T (It is the coefficient vector of the plane equation ax + by + cz + d = 0 for the surface f);
[0022] P is the homogeneous coordinate vector [x, y, z, 1] of vertex υ. T ;
[0023] n f It is the normal vector of the facet f.
[0024] Preferably, the management mechanism of the dynamic collapse priority queue includes: assigning vertices to different processing categories based on their importance scores: when the importance score of a vertex is lower than a first preset threshold, it is placed in a high-priority collapse queue for priority processing; when the importance score of a vertex is higher than a second preset threshold, it is marked as a permanently retained vertex, removed from the collapse queue, and no longer considered for collapse.
[0025] After each vertex collapse operation, the updated error metric values of all adjacent vertices within two hops of the collapsed vertex are recalculated, and the positions of these adjacent vertices in the dynamic collapse priority queue are readjusted based on the updated values.
[0026] Preferably, the iteration termination condition is a dual threshold constraint: a first quantity threshold, which is satisfied when the total number of simplified model patches is less than or equal to the preset target number of patches; and a second visual fidelity threshold, which is satisfied when the average structural similarity index between the key region set of the model and the corresponding region set of the original model is greater than or equal to the preset fidelity threshold; the key regions include door and window regions and roof edge regions; and the preset fidelity threshold is 0.95.
[0027] Preferably, before generating the region importance score through the deep learning network, data preprocessing is performed on the input original 3D building mesh model, including: topological reconstruction, repairing non-manifold edges and holes in the model; calculating the Gaussian curvature scalar value of each vertex in the model; and constructing adjacency relation matrices between vertices at different scales, including first-order neighborhood, second-order neighborhood and third-order neighborhood.
[0028] Preferably, the graph convolutional neural network includes an edge pooling layer; the operations performed by the edge pooling layer include: for an edge connecting two vertices, calculating the average of the feature vectors of the two vertices at both ends of the edge; performing linear transformation and nonlinear activation function processing on the average value; and selecting the feature with the largest activation value in the local region connected by the edge as the aggregated output feature to achieve the aggregation of local structural features of the building.
[0029] Preferably, the input 3D building mesh model originates from one or more of the following multi-source data: a textured triangular mesh model generated by oblique photogrammetry; a triangular mesh model reconstructed from LiDAR point cloud data; a parametric model exported from building information modeling software; and the method maintains the geometric topology of the output simplified model as a manifold structure during the collapse operation.
[0030] (III) Beneficial Effects
[0031] This invention provides a region simplification method for 3D building models based on deep learning. It has the following beneficial effects:
[0032] 1. This method uses graph convolutional neural networks to extract vertex semantic features and generate region importance scores, which can distinguish key areas of a building, doors, windows, roof edges and non-key areas, avoid the oversimplification of key areas, and effectively improve the visual fidelity and structural integrity of the simplified model.
[0033] 2. This method incorporates the regional importance score into the error metric calculation of the surface collapse algorithm and combines it with the Euclidean norm penalty term for the change of normal vector. This allows for dynamic adjustment of vertex collapse priority, ensuring that the collapse probability of highly important vertices is reduced and geometric distortion is minimized.
[0034] 3. This method improves the efficiency and specificity of the simplification process and avoids ineffective iterations by constructing a dynamic collapse priority queue, classifying vertex processing categories according to importance scores, and updating the errors of adjacent vertices and queue positions after each collapse.
[0035] 4. This method uses a dual threshold constraint of a face number threshold and a visual fidelity threshold to terminate the iteration, which satisfies the need for lightweight model and ensures the consistency between key regions and the original model through structural similarity index, thus balancing "lightweight" and "detail preservation".
[0036] 5. This method, through preprocessing such as repairing non-manifold edges / holes, calculating Gaussian curvature, and constructing multi-scale adjacency matrices, can be adapted to multi-source input models, such as oblique photogrammetry, LiDAR, and BIM export, thus improving model compatibility.
[0037] 6. This method aggregates local structural features of buildings through edge pooling layers of graph convolutional neural networks, enhancing the effectiveness of semantic feature extraction and making importance scoring more consistent with the logic of building structures. During the collapse operation, the geometric topology of the output model is maintained as a manifold structure to avoid topological anomalies in the simplified model, ensuring the model's usability in subsequent applications. Attached Figure Description
[0038] Figure 1 This is a system flowchart of the present invention. Detailed Implementation
[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0040] Please see Figure 1 The present invention provides a technical solution:
[0041] A deep learning-based method for region simplification of 3D building models includes the following steps:
[0042] Step 1: Extract semantic features from the input 3D building mesh model using a deep learning network to generate a regional importance score for each vertex, which represents the degree of simplification of the region where the vertex is located;
[0043] Step 2: Incorporate the regional importance score as a weight into the error metric calculation of the surface collapse algorithm to construct a dynamic collapse priority queue, in which the lower the importance score of a vertex, the higher its collapse priority in the queue.
[0044] Step 3: Perform iterative collapse operation based on the dynamic collapse priority queue. The iterative collapse operation includes: selecting the vertex with the highest collapse priority in the priority queue to perform collapse, updating the error metric value of the adjacent vertices affected by the collapse, and readjusting their priority in the queue.
[0045] Step 4: When the preset iteration termination conditions are met, stop the collapse operation and output the simplified 3D building model; the iteration termination conditions include: the number of simplified model patches reaches the target number threshold, and the structural similarity index between the key areas of the model and the corresponding areas of the original model reaches the preset fidelity threshold.
[0046] Deep learning networks are graph convolutional neural networks; the input of a graph convolutional neural network includes the following features for each vertex: three-dimensional spatial coordinates, normal vector direction, Gaussian curvature scalar value representing the curvature of the surface, and the number of adjacent vertices directly connected to that vertex;
[0047] The output of the graph convolutional neural network is an importance score for each vertex, which is a value between 0 and 1; where an importance score of 0 means that the vertex can be completely collapsed, and an importance score of 1 means that the vertex cannot be collapsed.
[0048] The error metric function for incorporating regional importance scoring is defined as:
[0049] Part 1: Calculate the basic error value of the vertex in the traditional face collapse algorithm. This basic error value is calculated based on the sum of the quadratic forms of the geometric errors of all the triangular faces adjacent to the vertex.
[0050] Part Two: The regional importance score is multiplied by the base error value as a penalty weight, which significantly increases the error value of vertices with high importance scores, thus reducing their priority for collapse.
[0051] Part 3: Add a penalty term based on the Euclidean norm of the change in the normal vector before and after vertex collapse. This penalty term is multiplied by a preset weight coefficient, which ranges from 0.2 to 0.5.
[0052] The final error metric is the result of multiplying the first part by the weight of the region importance score, plus the result of the normal change penalty term in the third part.
[0053] The above error metric function can be calculated using the following formula:
[0054]
[0055] Where S(υ) is the vertex importance score;
[0056] These are the basic geometric error values calculated using the traditional QEM method;
[0057] λ is the weighting coefficient, with a value ranging from 0.2 to 0.5;
[0058] ||nn′||2 represents the Euclidean norm between the normal vector n before and after the vertex collapse, used to measure the degree of change in the normal vector.
[0059] The management mechanism of the dynamic collapse priority queue includes: assigning vertices to different processing categories based on their importance scores: when a vertex's importance score is lower than the first preset threshold, it is placed in the high-priority collapse queue for priority processing; when a vertex's importance score is higher than the second preset threshold, it is marked as a permanently retained vertex, removed from the collapse queue, and no longer considered for collapse.
[0060] After each vertex collapse operation, the updated error metric values of all adjacent vertices within two hops of the collapsed vertex are recalculated, and the positions of these adjacent vertices in the dynamic collapse priority queue are readjusted based on the updated values.
[0061] The iteration termination condition is a dual threshold constraint: the first quantity threshold is satisfied when the total number of simplified model patches is less than or equal to the preset target number of patches; the second visual fidelity threshold is satisfied when the average structural similarity index between the key region set of the model and the corresponding region set of the original model is greater than or equal to the preset fidelity threshold; the key regions include the door and window regions and the roof edge regions; the preset fidelity threshold is 0.95.
[0062] Before generating region importance scores through a deep learning network, data preprocessing is performed on the input original 3D building mesh model, including: topological reconstruction, repairing non-manifold edges and holes in the model; calculating the Gaussian curvature scalar value of each vertex in the model; and constructing adjacency relation matrices between vertices at different scales, including first-order, second-order, and third-order neighborhoods.
[0063] Graph convolutional neural networks include edge pooling layers. The operations performed by the edge pooling layers include: for an edge connecting two vertices, calculating the average of the feature vectors of the two vertices at both ends of the edge; performing linear transformation and non-linear activation function processing on the average value; and selecting the feature with the largest activation value in the local region connected by the edge as the aggregated output feature to achieve the aggregation of local structural features of the building.
[0064] The input 3D building mesh model is derived from one or more of the following multi-source data: a textured triangular mesh model generated by oblique photogrammetry; a triangular mesh model reconstructed from LiDAR point cloud data; a parametric model exported from building information modeling software; and the method maintains the geometric topology of the output simplified model as a manifold structure during the collapse operation.
[0065] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0066] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for region simplification of a 3D building model based on deep learning, characterized in that, Includes the following steps: Step 1: Extract semantic features from the input 3D building mesh model using a deep learning network to generate a regional importance score for each vertex; Step 2: The regional importance score is used as a weight and incorporated into the error metric calculation of the surface collapse algorithm to construct a dynamic collapse priority queue; Step 3: Perform iterative collapse operations based on the dynamic collapse priority queue; Step 4: When the preset iteration termination condition is met, stop the collapse operation and output the simplified 3D building model; The deep learning network is a graph convolutional neural network; the input of the graph convolutional neural network includes the following features of each vertex: three-dimensional spatial coordinates, normal vector direction, Gaussian curvature scalar value representing the curvature of the surface, and the number of adjacent vertices directly connected to that vertex; The output of the graph convolutional neural network is an importance score for each vertex, which is a value between 0 and 1; where an importance score of 0 means that the vertex can be completely collapsed, and an importance score of 1 means that the vertex cannot be collapsed. The error metric function for the integration of regional importance scores is defined as follows: Part 1: Calculate the basic error value of vertices in the traditional face collapse algorithm; Part Two: Multiply the regional importance score as a penalty weight by the base error value; Part 3: Add a penalty term based on the Euclidean norm of the change in the normal vector before and after vertex collapse. This penalty term is multiplied by a preset weight coefficient, which ranges from 0.2 to 0.
5. The final error metric is the result of multiplying the first part by the weight of the regional importance score, plus the result of the normal change penalty term in the third part.
2. The method for region simplification of a 3D building model based on deep learning according to claim 1, characterized in that: The management mechanism of the dynamic collapse priority queue includes: assigning vertices to different processing categories based on their importance scores: when the importance score of a vertex is lower than a first preset threshold, it is placed in the high-priority collapse queue for priority processing; when the importance score of a vertex is higher than a second preset threshold, it is marked as a permanently retained vertex, removed from the collapse queue, and no longer considered for collapse.
3. The method for region simplification of a 3D building model based on deep learning according to claim 1, characterized in that: The iteration termination condition is a dual threshold constraint: the first quantity threshold is satisfied when the total number of simplified model patches is less than or equal to the preset target patch number; the second visual fidelity threshold is satisfied when the average structural similarity index between the key region set of the model and the corresponding region set of the original model is greater than or equal to the preset fidelity threshold.
4. The method for region simplification of a 3D building model based on deep learning according to claim 1, characterized in that: Before generating region importance scores through the deep learning network, data preprocessing is performed on the input original 3D building mesh model.
5. The method for region simplification of a 3D building model based on deep learning according to claim 1, characterized in that: The graph convolutional neural network includes an edge pooling layer; the edge pooling layer performs the following operations: for an edge connecting two vertices, it calculates the average of the feature vectors of the two vertices at both ends of the edge; it performs linear transformation and nonlinear activation function processing on the average value; and it selects the feature with the largest activation value in the local region connected by the edge as the aggregated output feature to achieve the aggregation of local structural features of the building.
6. The method for region simplification of a three-dimensional building model based on deep learning according to claim 1, characterized in that: The input 3D building mesh model is derived from one or more of the following multi-source data: a textured triangular mesh model generated by oblique photogrammetry; a triangular mesh model reconstructed from LiDAR point cloud data; A parametric model exported from building information modeling software; the method maintains the geometric topology of the output simplified model as a manifold structure during the collapse operation.