Building low-modulus reconstruction method based on hierarchical feature compensation

Abstract

The invention discloses a building low-modulus reconstruction method based on hierarchical feature compensation, and the method comprises the following steps: initializing a proxy grid: generating a visual housing through an original building high modulus, taking the visual housing as an initial proxy grid, determining a reconstruction process starting point, and preparing an initial feature unit set for iterative optimization; iterative optimization of the proxy grid: selecting feature units from the feature unit set to perform Boolean operation with the proxy grid, optimizing the geometric structure of the proxy grid, and updating the feature unit set; and grid post-processing: carrying out fine processing on the optimized proxy grid, eliminating artifacts, and generating a high-quality building low model. Through the method, a live-action building model with any topological structure can be processed without being influenced by the number of patches and surface noise, and a building low-modulus grid with few patches and high visual similarity is generated; and important geometric features and structural consistency of the building model can be effectively reserved, and meanwhile, technical support is provided for building data processing of large-scale three-dimensional scene reconstruction.

Description

Technical Field

[0001] The present invention belongs to the fields of computer graphics, three-dimensional reconstruction and remote sensing science and technology, and specifically relates to a low-poly model reconstruction method of buildings based on hierarchical feature compensation. Background Art

[0002] With the rapid development and widespread adoption of 3D scene reconstruction technology, precise 3D building models acquired through remote sensing are being widely used in disaster assessment, tourism, public services, mapping, and visualization, facilitating social production and daily life. However, these high-precision building models typically contain a large amount of vertex and mesh data. The sheer size of these models places high demands on storage and computing resources, limiting their widespread adoption in mobile applications and real-time scene interactions. This situation has led to widespread demand for low-polygon building model generation technologies that deliver high information content and compression ratios.

[0003] Currently, architectural low-poly model generation technologies fall into two main categories: mesh simplification algorithms based on local operations and low-poly model reconstruction methods based on remeshing. Local-operation-based mesh simplification algorithms, represented by QEM (Quadric Error Metric) and SLAM (Structure-Aware Mesh Decimation), reduce the number of vertices and facets in the model through operations such as edge collapse. However, these methods rely on local error metrics and struggle to reflect changes in the global features of the architectural model, often resulting in over-smoothing of edge structures. While SLAM methods introduce planar feature constraints, due to the frequent presence of noise on the surfaces of real-world architectural models, traditional feature detection and extraction methods (such as RANSAC, region growing, and model fitting) are limited in their effectiveness in accurately identifying planar features. Furthermore, poor topology further complicates local optimization, potentially leading to detached faces or topological errors after simplification.

[0004] Low-poly reconstruction methods based on remeshing generate lightweight models by geometrically reconstructing the original point cloud or mesh data. The PolyFit (Polygonal Surface Reconstruction from Point Clouds) method simplifies the model by extracting geometric primitives, but its estimation of primitive parameters is susceptible to surface noise, resulting in a loss of geometric feature fidelity. Other methods rely on accurate contour generation to restore the building's characteristic structure. When the input model has poor topology or complex geometry, these methods cannot generate correct contours, potentially causing low-poly reconstruction failure.

[0005] In summary, the main problems of existing methods are: the interference of model surface noise on feature extraction is not effectively handled, resulting in difficulty in accurately recovering important visual features of the building model; the impact of poor topological structure on algorithm performance is not fully addressed, and topological errors can lead to inaccurate error estimation and even model reconstruction failure; existing methods focus on the optimization of geometric features and ignore the visual consistency of low-poly and high-poly models under multiple perspectives, which affects user experience and model usability. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method for reconstructing low-poly models of buildings based on hierarchical feature compensation. This method effectively reduces the visual difference between the low-poly model of a building and the original high-poly model of a building at a high simplification rate through multi-level feature extraction and optimization, avoids the interference of poor topological structures in the original high-poly model of a building, ensures that the generated low-poly model of a building has good mesh quality, and can effectively retain the important geometric features and overall structural consistency of the original high-poly model of a building. At the same time, it provides technical support for architectural data processing for large-scale three-dimensional scene reconstruction.

[0007] In order to solve the above technical problems, the present invention is implemented in the following ways: A low-poly model reconstruction method for buildings based on hierarchical feature compensation. The specific process steps are as follows: S1. Initialize the proxy mesh: Generate a visual shell for the original building high-poly model and use the visual shell as the initial proxy mesh to determine the starting point of the reconstruction process. At the same time, prepare the initial feature unit set for iterative optimization. S2, iterative optimization of proxy mesh: From the feature cell set, select feature cells and proxy mesh to perform Boolean operations, optimize the proxy mesh geometry, and update the feature cell set at the same time; S3. Mesh post-processing: Perform fine processing on the proxy mesh after iterative optimization to eliminate artifacts and generate high-quality building low-poly models.

[0008] Furthermore, the specific steps of step S1 are as follows: S11. Constructing a visual shell: Extract the projection outline of the building model through multi-view projection, simplify the two-dimensional projection outline using the Ramer-Douglas-Peucker algorithm, and then map it into three-dimensional space to generate the visual shell of the building model. The visual shell can tightly wrap the main geometric structure of the original high-poly model of the building. S12. Initialization of feature unit set: Utilize the volume difference between the original building high-poly model and the generated visual shell to extract multiple initial feature units. Set the volume difference as redundant information, mark the Boolean operation symbol of each initial feature unit, and then form an initial feature unit set after screening.

[0009] Furthermore, the specific steps of step S11 are as follows: S111, Viewpoint Selection: Use the region growing algorithm to fit the surface of the original building high-polygon into multiple plane regions. The normal of the plane region is used as the viewpoint direction, and the area of the plane region is used as the sorting weight. The top 50 viewpoints are selected for the projection outline. S112, Contour Extraction and Simplification: Render the 3D architectural model into a 2D image using parallel projection. Use image extraction to extract the contours of the projected image. Simplify the 2D contours using the Ramer-Douglas-Peucker algorithm. Perform a second round of contour smoothing before converting them to 3D space as geometric contours. S113. Generate visual shell: Stretch the geometric outline along the positive and negative directions of the normal line of the plane to generate a corresponding three-dimensional geometric body. Continuously perform Boolean intersections between the geometric bodies to generate a corresponding visual shell.

[0010] Furthermore, the specific steps of step S12 are as follows: S121, Feature Unit Extraction: By voxelizing the original building high-poly model and the visual shell generated in S113, redundant feature structures are extracted from the visual shell using the volume difference. These feature structures are further processed through voxel-based morphological opening operations to eliminate small feature structures, separate feature structures at weak connections, and smooth the boundaries of larger feature structures without significantly changing their size, thereby obtaining feature units without assigning operation symbols; S122, feature unit symbol confirmation: The feature units extracted in S121 are assigned Boolean operation symbols for subsequent feature compensation. The obtained feature unit operation symbols are determined according to the hierarchical structure. For the initial feature unit, it is the volume difference between the original building high-poly model and the proxy mesh, which is a concave feature. The expression of the feature unit operation symbol is as follows: in, O ( p (l) )express l Hierarchical characteristic unit operation symbol, O ( p (l-1) )express l -1 level feature unit operation symbol, + represents convex features, - represents concave features; S123, feature unit screening: The feature units assigned with symbols in S122 are eliminated with small volumes, and the intersection volume of the feature units and the original building high-poly model is used to determine whether they are noise parts: The ratio of the intersection volume of the concave feature and the original building high polynomial to the volume of the original building high polynomial should be close to 0. If the ratio is too large, it is regarded as noise. The ratio of the intersection volume of the convex feature and the original building high polynomial to the volume of the original building high polynomial should be close to 1. If the ratio is too small, it is regarded as noise. This is used to generate the initial candidate feature unit set to ensure the validity and accuracy of the candidate set.

[0011] Furthermore, the specific steps of step S2 are as follows: S21, feature unit set search: First, for the initial candidate feature unit set, generate a visual shell for each feature unit in the set according to step S11, and evaluate the contribution of each visual shell to the proxy mesh accuracy. The feature unit corresponding to the visual shell with the largest contribution value is selected as the result of this search; S22, feature fusion and update: Perform a Boolean operation (such as difference or sum) on the visual shell of the selected feature unit and the proxy mesh to generate a new proxy mesh, completing a feature compensation. At the same time, the current feature unit is removed from the candidate feature unit set, and the removed feature unit and its visual shell are combined to generate the next-level feature unit of the feature unit according to step S12, and the next-level feature unit is added to the feature unit set. S23, multi-level feature compensation: repeat steps S21 and S22, compensating the feature structure of the proxy mesh layer by layer until the set of candidate feature units is empty or the improvement value of the visual similarity difference between the proxy mesh and the original model is lower than the set threshold.

[0012] Furthermore, the specific method for selecting the characteristic unit in step 21 is as follows: S211, Node Search: Construct a global search tree on the set of candidate feature units. Use multi-step prediction to evaluate the short-term and long-term contribution of each feature unit. Each node in the global search tree represents the state of the proxy grid after a Boolean operation, and branches correspond to the selection of different candidate feature units. Through multiple paths in the search tree, calculate the comprehensive improvement value of each feature unit in terms of visual similarity difference, and select the feature unit with the largest contribution value as the current optimization direction. The node search process includes three steps: node selection, node expansion, and node simulation. The unexpanded nodes are selected for node expansion. If all nodes are expanded, a node is selected according to the formula, which is as follows: in, U(n) Indicates the selection weight of node n. The larger the value, the higher the priority. N(n) Indicates the number of times node n is visited; N(n') Representation node n' Number of visits, Q(n) Indicates the evaluation value of the current node; After selecting a node, it is expanded to generate the nodes of the next layer. Then, for the newly expanded nodes, feature compensation simulation is performed using the feature cells in the node and the proxy grid. The simulation stops when it reaches a given depth or no nodes can be expanded. The depth is set to 5. S212, Node Evaluation: During the search process, each node is evaluated. Specifically, the proxy mesh is subjected to a Boolean operation with the visual shell of the node's feature unit to compensate for the feature, resulting in a new proxy mesh for the node. The visual similarity difference between the new proxy mesh and the original building high-polygon is calculated. The visual similarity difference between the proxy mesh and the original building high-polygon is reflected by the depth map pixel information. A visual difference function is constructed based on the multi-view depth map. Its calculation formula is as follows: Where T represents the proxy grid M proxy With the original building high model M o The visual similarity difference between D Indicates the projection distance, n j Indicates the number of depth maps, n x Indicates the horizontal number of pixels in the depth map, n y Indicates the horizontal number of pixels in the depth map, n x* n y Indicates the number of pixels in the depth map under the viewing angle. v j Indicates the v j Depth map under different viewing angles, x, y Represents horizontal x , vertical y The pixel position, d proxy (x, y) Representing a proxy grid M proxy In the v j Depth map ( x, y ) position, d o (x, y) Indicates the original building high poly model M o In the v j Depth map ( x, y ) position’s depth value; S213, Backward propagation: After the node is evaluated, the node evaluation value is updated along the reverse path. The expression is as follows: Q ( n ) = max ( Q o ( n ), q ) in, Q ( n ) represents the current node n The updated evaluation value of Q o ( n ) represents the current node n The evaluation value before the update, q Representation node n The evaluation value of the child node backpropagated up, and the maximum value of Q0(n) and q is selected as the final result.

[0013] Furthermore, the specific steps of step S3 are as follows: S31, vertex adjustment: The proxy mesh that has been iteratively optimized in step S2 can be used as the initial building low-poly model. The vertex distance between the initial building low-poly model and the original building high-poly model is used to adjust the vertex of each vertex in the initial building low-poly model. v Find the nearest vertex on the original building high poly model v' , and v Move to v' , this operation reduces the tiny protrusions in the initial building low-poly model and smoothes the surface of the initial building low-poly model; S32. Surface optimization: Using the visual similarity difference function in step S212 as the error metric for face simplification, the QEM algorithm is used to further simplify the surface. The fragmented faces and small facet areas generated by Boolean operations in the initial building low-poly model are optimized and merged to ensure the overall manifold and watertightness of the mesh.

[0014] Compared with the prior art, the present invention has the following beneficial effects: The present invention effectively avoids the unstable geometry and topology of the original building model by introducing a visual shell to construct a multi-level feature structure, effectively extracting the important appearance structure and spatial feature structure of the building model, thereby improving visual similarity. By introducing a global tree search strategy to select feature units, the short-sightedness defects caused by local optimality are avoided. The hierarchical structure of geometric features maximizes the restoration of the spatial features of the model. By introducing a visual similarity index based on depth and contour, feature compensation is guided to optimize in the most effective direction to maximize visual similarity. The post-processing mechanism avoids the surface "fragmentation" caused by a large number of geometric operations, as well as the deviation errors caused by discrete operations such as rasterization and voxelization. This method can process real-life building models with arbitrary topological structures, unaffected by the number of facets and surface noise, and generate low-poly meshes of buildings with very few facets and high visual similarity. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flowchart of the present invention; Figure 2 This is a schematic diagram of the low-poly model of the building generated by the present invention; Figure 3 This is a schematic diagram of the intermediate results of the low-poly model generation process of the building according to the present invention. DETAILED DESCRIPTION

[0016] The specific implementation of the present invention will be further described in detail below with reference to the accompanying drawings and specific examples.

[0017] like Figures 1 - 2 As shown in the figure, a low-poly model reconstruction method of buildings based on hierarchical feature compensation is shown in the figure. The specific process steps are as follows: S1. Initialize the proxy mesh: Generate a visual shell from the original building high-poly model and use this visual shell as the initial proxy mesh to determine the starting point of the reconstruction process. At the same time, prepare the initial feature unit set for iterative optimization. The specific steps are as follows: S11. Constructing a visual shell: Extract the projection outline of the building model through multi-view projection, simplify the two-dimensional projection outline using the Ramer-Douglas-Peucker algorithm, and then map it into three-dimensional space to generate the visual shell of the building model. The visual shell can tightly wrap the main geometric structure of the original building high-poly model. The specific steps are as follows: S111, Viewpoint Selection: Use the region growing algorithm to fit the surface of the original building high-polygon into multiple plane regions. The normal of the plane region is used as the viewpoint direction, and the area of the plane region is used as the sorting weight. The top 50 viewpoints are selected for the projection outline. S112. Contour extraction and simplification: Render the architectural model in three-dimensional space into a two-dimensional image through parallel projection. Use image extraction to extract the contour of the projected image, and use the Ramer-Douglas-Peucker algorithm to simplify the two-dimensional contour. Perform a second round of contour smoothing and then convert it to three-dimensional space as a geometric contour. The Ramer-Douglas-Peucker algorithm defines the weight as the distance from the vertex to the straight line between the two adjacent points. Each time, the point with the smallest distance is selected and removed. The algorithm uses the largest vertex distance in the current result as the simplification target. When the distance of all points is greater than this distance, the iteration stops.

[0018] S113. Generate visual shells: Stretch the geometric outline along the positive and negative directions of the plane normal to generate corresponding three-dimensional geometric bodies. Continuously perform Boolean intersections between the geometric bodies to generate corresponding visual shells. During the intersection process, in order to reduce unnecessary calculations, the visual similarity index is used as a guide. When the similarity change is less than a given threshold, stop selecting geometric bodies.

[0019] S12. Initialization of feature unit set: Utilize the volume difference between the original building high-poly model and the generated visual shell to extract multiple initial feature units. The volume difference is set as redundant information. The Boolean operation symbol of each initial feature unit is marked. After screening, an initial feature unit set is formed. The specific steps are as follows: S121, Feature Unit Extraction: By voxelizing the original building high-poly model and the visual shell generated in S113, redundant feature structures are extracted from the visual shell using the volume difference. These feature structures are further processed through voxel-based morphological opening operations to eliminate small feature structures, separate feature structures at weak connections, and smooth the boundaries of larger feature structures without significantly changing their size, thereby obtaining feature units without assigning operation symbols; S122, feature unit symbol confirmation: The feature units extracted in S121 are assigned Boolean operation symbols for subsequent feature compensation. The obtained feature unit operation symbols are determined according to the hierarchical structure. For the initial feature unit, it is the volume difference between the original building high-poly model and the proxy mesh, which is a concave feature. The expression of the feature unit operation symbol is as follows: in, O ( p (l) )express l Hierarchical characteristic unit operation symbol, O ( p (l-1) )express l -1 level feature unit operation symbol, + represents convex features, - represents concave features; S123, feature unit screening: The feature units assigned with symbols in S122 are eliminated with small volumes, and the intersection volume of the feature units and the original building high-poly model is used to determine whether they are noise parts: The ratio of the intersection volume of the concave feature and the original building high polynomial to the volume of the original building high polynomial should be close to 0. If the ratio is too large, it is regarded as noise. The ratio of the intersection volume of the convex feature and the original building high polynomial to the volume of the original building high polynomial should be close to 1. If the ratio is too small, it is regarded as noise. This is used to generate the initial candidate feature unit set to ensure the validity and accuracy of the candidate set.

[0020] S2. Iterative Optimization of the Proxy Mesh: Select feature cells from the feature cell set and perform Boolean operations on the proxy mesh to optimize the proxy mesh geometry and update the feature cell set. The specific steps are as follows: S21. Feature unit set search: First, for the initial candidate feature unit set, generate a visual shell for each feature unit in the set according to step S11, and evaluate the contribution of each visual shell to the accuracy of the proxy mesh. The contribution value is determined by the improvement value of the visual similarity difference between the proxy mesh after the feature unit is applied and the original building high-poly model. The visual similarity difference improvement value can be calculated by the difference between the two visual similarity difference values. The feature unit corresponding to the visual shell with the largest contribution value is selected as the result of this search. The specific method of feature unit selection is as follows: S211, Node Search: Construct a global search tree on the set of candidate feature units. Use multi-step prediction to evaluate the short-term and long-term contribution of each feature unit. Each node in the global search tree represents the state of the proxy grid after a Boolean operation, and branches correspond to the selection of different candidate feature units. Through multiple paths in the search tree, calculate the comprehensive improvement value of each feature unit in terms of visual similarity difference, and select the feature unit with the largest contribution value as the current optimization direction. The node search process includes three steps: node selection, node expansion, and node simulation. The unexpanded nodes are selected for node expansion. If all nodes are expanded, a node is selected according to the formula, which is as follows: in, U(n) Indicates the selection weight of node n. The larger the value, the higher the priority. N(n) Indicates the number of times node n is visited; N(n') Representation node n' Number of visits, Q(n) Indicates the evaluation value of the current node; After selecting a node, it is expanded to generate the nodes of the next layer. Then, for the newly expanded nodes, feature compensation simulation is performed using the feature cells in the node and the proxy grid. The simulation stops when it reaches a given depth or no nodes can be expanded. The depth is set to 5. S212, Node Evaluation: During the search process, each node is evaluated. Specifically, the proxy mesh is subjected to a Boolean operation with the visual shell of the node's feature unit to compensate for the feature, resulting in a new proxy mesh for the node. The visual similarity difference between the new proxy mesh and the original building high-polygon is calculated. The visual similarity difference between the proxy mesh and the original building high-polygon is reflected by the depth map pixel information. A visual difference function is constructed based on the multi-view depth map. Its calculation formula is as follows: Where T represents the proxy grid M proxy With the original building high model M o The visual similarity difference between D Indicates the projection distance, n j Indicates the number of depth maps, n x Indicates the horizontal number of pixels in the depth map, n y Indicates the horizontal number of pixels in the depth map, n x* n y Indicates the number of pixels in the depth map under the viewing angle. v j Indicates the v j Depth map under different viewing angles, x, y Represents horizontal x , vertical y The pixel position, d proxy (x, y) Representing a proxy grid M proxy In the v j Depth map ( x, y ) position, d o (x, y) Indicates the original building high poly model M o In the v j Depth map ( x, y ) position depth value; in order to maximize feature optimization, the single-view difference value reduction is used as the node evaluation value in step S213 when calculating the visual difference.

[0021] S213, Backward propagation: After the node is evaluated, the node evaluation value is updated along the reverse path. The expression is as follows: Q ( n ) =max ( Q o ( n ), q ) in, Q ( n ) represents the current node n The updated evaluation value of Q o ( n ) represents the current node n The evaluation value before the update, q Representation node n The evaluation value of the child node backpropagated up, and the maximum value of Q0(n) and q is selected as the final result.

[0022] S22, feature fusion and update: Perform a Boolean operation (such as difference or sum) on the visual shell of the selected feature unit and the proxy mesh to generate a new proxy mesh, completing a feature compensation. At the same time, the current feature unit is removed from the candidate feature unit set, and the removed feature unit and its visual shell are combined to generate the next-level feature unit of the feature unit according to step S12, and the next-level feature unit is added to the feature unit set. S23, multi-level feature compensation: repeat steps S21 and S22, compensating the feature structure of the proxy mesh layer by layer until the set of candidate feature units is empty or the improvement value of the visual similarity difference between the proxy mesh and the original model is lower than the set threshold.

[0023] S3. Mesh post-processing: Perform fine processing on the proxy mesh after iterative optimization to eliminate artifacts such as jagged edges and irregular triangle distribution, and generate high-quality building low-poly models. The specific steps are as follows: S31, vertex adjustment: The proxy mesh that has been iteratively optimized in step S2 can be used as the initial building low-poly model. The vertex distance between the initial building low-poly model and the original building high-poly model is used to adjust the vertex of each vertex in the initial building low-poly model. v Find the nearest vertex on the original building high poly model v' , and v Move to v' , this operation reduces the tiny protrusions in the initial building low-poly model and smoothes the surface of the initial building low-poly model; S32. Surface optimization: Using the visual similarity difference function in step S212 as the error metric for face simplification, the QEM algorithm is used to further simplify the surface. The fragmented faces and small facet areas generated by Boolean operations in the initial building low-poly model are optimized and merged to ensure the overall manifold and watertightness of the mesh. Example

[0024] First, the UrbanScene3D dataset was selected for this case study. This dataset covers 16 scenes, including large-scale real urban areas and synthetic cities, totaling 136 square kilometers and containing a large number of real-world building models. To verify the effectiveness and practicality of this method, the scenes were instance-segmented and 100 noisy individual building models were extracted as raw input data. These building models were reconstructed from high-precision LiDAR imagery and not only have extremely high numbers of triangles (ranging from 10,000 to 2.5 million), but also exhibit complex topological structures, including non-watertightness, non-manifold, and self-intersection issues.

[0025] The evaluation of architectural low-poly mesh generation results is generally based on three aspects: simplification ratio, model quality (watertightness and manifoldness), and visual similarity difference (LightFieldDescriptor (LFD)). The simplification ratio is the ratio of the number of low-poly meshes to the number of high-poly meshes; a smaller value indicates better simplification. Watertightness refers to the mesh model having closed surfaces, no holes, and clearly defined interiors and exteriors. Building mesh models with good watertightness are better suited for 3D scene rendering and other applications. Manifoldness generally refers to the mesh model's lack of self-intersecting surfaces and its localized planar properties. Visual similarity (LFD) projects the 3D model onto a 2D plane and calculates the spatial differences and contour information of the models. For two models, the smaller the LFD, the smaller the difference and the greater the similarity. Furthermore, the success rate is used to verify the robustness and practicality of each method, indicating whether the method can successfully output results without crashing during execution.

[0026] To ensure a fair comparison with other low-poly model generation methods on the UrbanScene3D dataset, the parameters for each comparison method are set as follows: The simplification rates of the QEM, Blender, and SAMD methods are consistent with the simplification rates of the experimental results in this paper, but due to algorithmic flaws, the set simplification rates may not be achieved. For Polyfit, the mesh model is first sampled into a point cloud containing one million points using Monte Carlo, and then the low-poly model is generated using Polyfit. The simplification rate cannot be set. For the LPBM, BGM, MCC, and BCMR methods, the simplification rate cannot be set either, so the low-poly model is reconstructed according to the official parameters. On this basis, each method generates low-poly models of 100 noisy single-building models and compares their performance.

[0027]

[0028] Table 1 shows that the proposed method demonstrates strong robustness in terms of the success rate of low-poly model reconstruction. Although the proposed algorithm cannot achieve the lowest simplification rate, it achieves the highest similarity index with its LFD counterpart, demonstrating that the proposed method significantly improves the visual similarity of the model while maintaining a lower simplification rate. The proposed method effectively ensures the watertightness and manifold properties of the generated low-poly model, ensuring mesh quality, achieving a balance and improvement in simplification rate, visual similarity, and mesh quality.

[0029] like Figure 3 As shown in the figure, for a real-life building model with a large number of facets and surface noise, the proxy mesh is generated as the initial low-poly model through step S1. Its good geometric quality plays a key role in subsequent iterations. After iterative optimization in S2, the features of the proxy mesh are hierarchically compensated and optimized to restore the spatial structural features of the building model. Finally, after post-processing, a low-poly mesh with an appearance structure similar to the original building model is obtained, which proves its strong advantages in practical applications.

[0030] The above description is merely an embodiment of the present invention. It is stated again that, for a person skilled in the art, several improvements can be made to the present invention without departing from the principles of the present invention, and these improvements are also included in the scope of protection of the claims of the present invention.

Claims

1. A method for low-poly model reconstruction of buildings based on hierarchical feature compensation, characterized by: The specific process steps are as follows: S1. Initialize the proxy mesh: Generate a visual shell from the original building high-poly model, and use this visual shell as the initial proxy mesh to determine the starting point of the reconstruction process, while preparing the initial feature unit set for iterative optimization; S2, iterative optimization of proxy mesh: From the feature cell set, select feature cells and proxy mesh to perform Boolean operations, optimize the proxy mesh geometry, and update the feature cell set at the same time; S3. Mesh post-processing: Finely process the proxy mesh after iterative optimization to eliminate artifacts and generate high-quality building low-poly models.

2. The method for low-poly model reconstruction of buildings based on hierarchical feature compensation according to claim 1, characterized in that: The specific steps of step S1 are as follows: S11. Constructing a visual shell: Extract the projection outline of the building model through multi-view projection, simplify the two-dimensional projection outline using the Ramer-Douglas-Peucker algorithm, and then map it into three-dimensional space to generate the visual shell of the building model. The visual shell can tightly wrap the main geometric structure of the original high-poly model of the building. S12. Initialization of feature unit set: Utilize the volume difference between the original building high-poly model and the generated visual shell to extract multiple initial feature units. Set the volume difference as redundant information, mark the Boolean operation symbol of each initial feature unit, and then form an initial feature unit set after screening.

3. The method for low-poly model reconstruction of buildings based on hierarchical feature compensation according to claim 2, characterized in that: The specific steps of step S2 are as follows: S21, feature unit set search: First, for the initial candidate feature unit set, generate a visual shell for each feature unit in the set according to step S11, and evaluate the contribution of each visual shell to the proxy mesh accuracy. The feature unit corresponding to the visual shell with the largest contribution value is selected as the result of this search; S22, feature fusion and update: Perform a Boolean operation on the visual shell of the selected feature unit and the proxy mesh to generate a new proxy mesh, completing a feature compensation. At the same time, the current feature unit is removed from the candidate feature unit set, and the removed feature unit and its visual shell are combined to generate the next-level feature unit of the feature unit according to step S12, and the next-level feature unit is added to the feature unit set. S23, multi-level feature compensation: repeat steps S21 and S22, compensating the feature structure of the proxy mesh layer by layer until the set of candidate feature units is empty or the improvement value of the visual similarity difference between the proxy mesh and the original model is lower than the set threshold.

4. The method for low-poly model reconstruction of buildings based on hierarchical feature compensation according to claim 2, characterized in that: The specific steps of step S11 are as follows: S111, Viewpoint Selection: Use the region growing algorithm to fit the surface of the original building high-polygon into multiple plane regions. The normal of the plane region is used as the viewpoint direction, and the area of the plane region is used as the sorting weight. The top 50 viewpoints are selected for the projection outline. S112, Contour Extraction and Simplification: Render the 3D architectural model into a 2D image using parallel projection. Use image extraction to extract the contours of the projected image. Simplify the 2D contours using the Ramer-Douglas-Peucker algorithm. Perform a second round of contour smoothing before converting them to 3D space as geometric contours. S113. Generate visual shell: Stretch the geometric outline along the positive and negative directions of the normal line of the plane to generate a corresponding three-dimensional geometric body. Continuously perform Boolean intersections between the geometric bodies to generate a corresponding visual shell.

5. The method for low-poly model reconstruction of buildings based on hierarchical feature compensation according to claim 4, characterized in that: The specific steps of step S12 are as follows: S121, Feature Unit Extraction: By voxelizing the original building high-poly model and the visual shell generated in S113, redundant feature structures are extracted from the visual shell using the volume difference. These feature structures are further processed through voxel-based morphological opening operations to eliminate small feature structures, separate feature structures at weak connections, and smooth the boundaries of larger feature structures without significantly changing their size, thereby obtaining feature units without assigning operation symbols; S122, feature unit symbol confirmation: The feature units extracted in S121 are assigned Boolean operation symbols. The obtained feature unit operation symbols are determined according to the hierarchical structure. For the initial feature unit, it is the volume difference between the original building high-poly model and the proxy mesh. The expression of the feature unit operation symbol is as follows: in, O ( p (l) )express l Hierarchical characteristic unit operation symbol, O ( p (l-1) )express l -1 level feature unit operation symbol, + represents convex features, - represents concave features; S123, feature unit screening: The feature units assigned with symbols in S122 are eliminated with small volumes, and the intersection volume of the feature units and the original building high-poly model is used to determine whether they are noise parts: The ratio of the intersection volume of the concave feature and the original building high polynomial to the volume of the original building high polynomial should be close to 0. If the ratio is too large, it is regarded as noise; the ratio of the intersection volume of the convex feature and the original building high polynomial to the volume of the original building high polynomial should be close to 1. If the ratio is too small, it is regarded as noise. This is used to generate the initial candidate feature unit set.

6. The method for low-poly model reconstruction of buildings based on hierarchical feature compensation according to claim 3, characterized in that: The specific method of selecting the characteristic unit is as follows: S211, Node Search: Construct a global search tree on the set of candidate feature units. Use multi-step prediction to evaluate the short-term and long-term contribution of each feature unit. Each node in the global search tree represents the state of the proxy grid after a Boolean operation, and branches correspond to the selection of different candidate feature units. Through multiple paths in the search tree, calculate the comprehensive improvement value of each feature unit in terms of visual similarity difference, and select the feature unit with the largest contribution value as the current optimization direction. The node search process includes three steps: node selection, node expansion, and node simulation. The unexpanded nodes are selected for node expansion. If all nodes are expanded, a node is selected according to the formula, which is as follows: in, U(n) Indicates the selection weight of node n. The larger the value, the higher the priority. N(n) Indicates the number of times node n is visited; N(n') Representation node n' Number of visits, Q(n) Indicates the evaluation value of the current node; After selecting a node, it is expanded to generate the nodes of the next layer. Then, for the newly expanded nodes, feature compensation simulation is performed using the feature cells in the node and the proxy grid. The simulation stops when it reaches a given depth or no nodes can be expanded. S212, Node Evaluation: During the search process, each node is evaluated. Specifically, the proxy mesh is subjected to a Boolean operation with the visual shell of the node's feature unit to compensate for the feature, resulting in a new proxy mesh for the node. The visual similarity difference between the new proxy mesh and the original building high-polygon is calculated. The visual similarity difference between the proxy mesh and the original building high-polygon is reflected by the depth map pixel information. A visual difference function is constructed based on the multi-view depth map. Its calculation formula is as follows: Where T represents the proxy grid M proxy With the original building high model M o The visual similarity difference between D Indicates the projection distance, n j Indicates the number of depth maps, n x Indicates the horizontal number of pixels in the depth map, n y Indicates the horizontal number of pixels in the depth map, n x* n y Indicates the number of pixels in the depth map under viewing angle. v j Indicates the v j Depth map under different viewing angles, x, y Represents horizontal x , vertical y The pixel position, d proxy (x,y) Representing a proxy grid M proxy In the v j Depth map ( x,y ) position, d o (x,y) Indicates the original building high poly model M o In the v j Depth map ( x,y ) position’s depth value; S213, Backward propagation: After the node is evaluated, the node evaluation value is updated along the reverse path. The expression is as follows: Q ( n ) = max ( Q o ( n ), q ) in, Q ( n ) represents the current node n The updated evaluation value of Q o ( n ) represents the current node n The evaluation value before the update, q Representation node n The evaluation value back-propagated from the child nodes.

7. The method for low-poly model reconstruction of buildings based on hierarchical feature compensation according to claim 6, characterized in that: The specific steps of step S3 are as follows: S31, vertex adjustment: The proxy mesh that has been iteratively optimized in step S2 is used as the initial building low-poly model. The vertex distance between the initial building low-poly model and the original building high-poly model is used to adjust the vertex of each vertex in the initial building low-poly model. v Find the nearest vertex on the original building high poly model v' , and v Move to v' , reducing the tiny bumps in the initial building low-poly model and smoothing the surface of the initial building low-poly model; S32. Surface optimization: Using the visual similarity difference function in step S212 as the error metric for face simplification, the QEM algorithm is used to further simplify the surface. The fragmented faces and small facet areas generated by Boolean operations in the initial building low-poly model are optimized and merged to ensure the overall manifold and watertightness of the mesh.

Images