Automatic contact surface fusion method for three-dimensional model based on feature distance field regression network

Abstract

The application discloses a kind of based on feature distance field regression network's three-dimensional model contact surface automatic fusion method, belong to three-dimensional model field, comprising: obtaining multiple three-dimensional model components to be spliced and organizing as unified polygonal mesh input, after extracting the open boundary information of each component, contact surface suture is executed with the connectivity of implicit surface patching and keeping, and initial fusion grid is generated.After cleaning and boundary alignment by manifold, the sharp feature near contact surface is estimated using three-dimensional sharp feature distance field regression network model, and feature guide constraint containing feature points, feature lines and extension direction is generated.According to the constraint, adaptive reconstruction optimization is performed on the surface patching area, and a transition grid is obtained.Based on the L1 energy of the surface Hessian matrix, surface sharpening optimization is performed, and finally local topology repair is performed to output a complete three-dimensional model.The application can realize automatic fusion of high-quality three-dimensional model while maintaining smooth transition of contact surface.

Description

Technical Field

[0001] This invention belongs to the field of three-dimensional models, and in particular relates to an automatic fusion method for contact surfaces of three-dimensional models based on a feature distance field regression network. Background Technology

[0002] In scenarios such as the digital restoration of fragmented cultural relics and the virtual assembly of complex components, the acquired 3D models are often characterized by multiple independent components with open boundaries due to limitations such as scanning equipment occlusion, missing data, or physical damage. After initial spatial alignment of these components, the core issue affecting the quality of subsequent display, simulation analysis, and re-editing is how to automatically merge the gaps at the interface caused by geometric mismatches or missing data. Current technologies often employ Boolean operations, hole filling, implicit surface reconstruction, mesh stitching, or surface smoothing to process the contact surfaces of 3D models. These methods primarily rely on geometric rules such as the positional relationships of model boundaries, local distances, normal variations, or curvature thresholds to fill and connect the contact areas.

[0003] However, most of the existing methods mentioned above are rule-driven or locally geometrically heuristic methods, lacking the ability to intelligently perceive potential sharp features, feature line extension directions, and cross-gap structural relationships in complex contact areas. In practical applications, when the boundaries are irregular, weak features are obvious, noise interference is strong, or the contact area spans a large distance, it is difficult to reliably identify potential feature lines near the contact surface by simply relying on manually set curvature thresholds or local geometric rules. This can easily lead to the weakening of sharp structures, loss of details, surface shrinkage and deformation, and topological defects such as holes, overlaps, and T-shaped connections. Furthermore, existing technologies struggle to unify deep learning feature perception, feature-guided reconstruction, geometric surface optimization, and topology repair into a single processing flow. This results in low stability, feature fidelity, and topological reliability of 3D model contact surface fusion in complex scenarios, making it impossible to effectively restore and continue the sharp corners and fine geometric features of the original model while maintaining a smooth transition of the contact surface. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides an automatic fusion method for contact surfaces of 3D models based on a feature distance field regression network, comprising: Obtain multiple 3D model components to be stitched together, and organize these components into a unified polygonal mesh input. Based on the polygonal mesh input, extract the open boundary information of each model component; Based on the open boundary information, implicit surface patching is performed to generate an initial surface patching mesh that matches the boundary of the original model. Based on the initial patch mesh, perform contact surface stitching that maintains connectivity to generate an initial fused mesh that is connected to the original model; The initial fused mesh is manifold cleaned and its boundaries aligned to obtain the initial mesh for surface patching; Based on the initial mesh of the patch surface, the sharp features near the contact surface are learned and estimated using a three-dimensional sharp feature distance field regression network model to generate feature guidance constraints that include feature points, feature lines and their extension directions. Based on the feature-guided constraints, adaptive reconstruction optimization is performed on the patching region to obtain a transition mesh; based on the transition mesh, L1 energy surface sharpening optimization based on the surface Hessian matrix is ​​performed to generate a sharpened contact surface fusion surface. Local topology repair is performed on the sharpened contact surface fusion surface to output a complete 3D model.

[0005] Optionally, the implicit surface patching process further includes: The polygonal mesh input surface is sampled to obtain a discrete set of sampling points; Based on the discrete sampling point set, an implicit surface matching the boundary of the original model and its corresponding octree structure are generated using the Poisson surface reconstruction method.

[0006] Optionally, the step of performing contact surface stitching to maintain connectivity further includes: Based on the initial patch mesh, a connectivity-preserving double contour stitching method facing the mesh boundary is used to generate new triangular patches in the boundary gap region; By extending the boundary and pruning redundant elements, the newly added triangular facets are connected to the boundary of the original model to form the initial fused mesh.

[0007] Optionally, the manifold cleanup and boundary alignment of the initial fused mesh further includes: The initial fused mesh is subjected to non-manifold cell culling and local degenerate patch cleaning to obtain a cleaned mesh; The edge vertices of the patched area in the cleaned mesh are coupled positionally with the original model open boundary corresponding to the open boundary information to obtain a patched initial mesh that satisfies the manifold condition and is connected to the original model.

[0008] Optionally, generating the feature-guided constraints further includes: Based on the initial mesh of the patched surface, local geometric fragments within the original model boundary and the neighborhood of the patched surface are extracted; the local geometric fragments are input into the three-dimensional sharp feature distance field regression network model; The network model outputs the distance field prediction value, feature direction prediction information, and confidence level from each sampling position to the potential sharp feature through the local geometric feature encoding module, distance field regression module, and feature direction prediction module. Based on the distance field prediction value, feature direction prediction information, and confidence level, candidate feature points and candidate feature lines are generated, and the extension trend is determined to form the feature guidance constraint.

[0009] Optionally, the adaptive reconstruction optimization of the patched area further includes: Based on the feature points, feature lines and their extension directions in the feature guidance constraints, determine the guidance target position and vertex constraint weights; A variational shape reconstruction method based on a quadratic error metric is adopted, and the transition mesh is generated by adaptively adjusting the vertex arrangement and mesh connection relationship of the surface patching region according to the guiding target position and vertex constraint weight.

[0010] Optionally, the execution of L1 energy surface sharpening optimization based on the surface Hessian matrix further includes: Obtain the vertex coordinates of the patch region in the transition mesh as optimization variables; construct a joint energy objective function; The joint energy objective function includes: an L1 sparsification energy term based on the surface Hessian matrix, a boundary constraint term for fixing the position of the boundary vertices of the original model, and a feature constraint term for maintaining the continuity and orientation consistency of the feature lines. By minimizing the joint energy objective function, the vertex coordinates of the patching region are updated to generate the sharpened contact surface fusion surface.

[0011] Optionally, performing local topology repair further includes: Detect local abnormal regions in the sharpened contact surface fusion surface, including holes, mesh gaps, local overlaps, T-shaped connections, degenerate triangles, or non-manifold units; A feature-preserving mesh repair method based on restricted power graphs is used to reconstruct and repair the detected local abnormal regions.

[0012] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0013] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects: First, by introducing a 3D sharp feature distance field regression network, the identification of discrete feature points and feature lines is transformed into a continuous distance field regression problem. This effectively overcomes the shortcomings of traditional methods, such as reliance on manual curvature thresholds and difficulty in perceiving weak features and irregular boundaries, significantly improving the robustness of sharp feature recognition near the contact surface. Second, during the fusion process, predicted feature points, feature lines, and extension directions are used as guiding constraints to adaptively reconstruct the patching region and perform sharpening optimization based on the L1 norm of the surface Hessian matrix. This achieves a smooth transition of the contact surface while perfectly preserving the sharp edges and corner details of the original model, avoiding feature blunting and shrinkage deformation that occur in traditional methods. Finally, post-processing topology repair effectively eliminates defects such as holes, overlaps, and non-manifold units. The output 3D model is geometrically continuous and topologically reliable, suitable for high-fidelity virtual display, simulation analysis, and re-editing applications. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a model splicing scene with different cross-sectional shapes according to an embodiment of the present invention; Figure 3 This is a schematic diagram comparing the fidelity technology effects of an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0018] Example 1 This embodiment provides an automatic fusion method for contact surfaces of 3D models based on a feature distance field regression network, including: Using a unified polygonal mesh composed of multiple 3D models to be stitched as input, without altering the connectivity and geometry of the original models, the process employs implicit surface patching, connectivity-preserving contact surface stitching, manifold cleanup and boundary alignment, sharp feature detection and guided feature construction based on a feature distance field regression network, feature-guided adaptive reconstruction, feature-preserving surface optimization, and post-processing topology repair to obtain a fusion result with continuous contact surfaces, natural transitions, clear sharp features, and reliable topology. The overall workflow framework is as follows: Figure 1 As shown, the specific steps include: Step 1: Construct the input data for the 3D model to be stitched. Obtain multiple 3D model components to be stitched and organize them into a unified polygonal mesh input; extract the open boundary, boundary vertices, boundary edges and their adjacent face information of each model component to determine the area to be filled and the contact area, providing basic geometric data for subsequent automatic fusion.

[0019] Step 2: Perform sampling and implicit surface patching. Sample the surface of the input model and generate implicit surfaces that match the original boundaries and their corresponding octree structures based on the sampled points using the filtered Poisson surface reconstruction method. Optionally, guide sampling points are added to control the connection or separation trends between model components, obtaining the initial surface patching results for the contact area.

[0020] Step 3: Perform contact surface stitching that maintains connectivity. For implicit surfaces, a connectivity-preserving double contour stitching method oriented towards the mesh boundary is adopted to mesh the boundary gap region; through boundary extension, redundant element clipping, and generation of new triangular patches, the patched mesh is reliably connected to the original model boundary, forming an initial fused mesh without destroying the connectivity of the original model.

[0021] Step 4: Manifold cleanup and boundary alignment based on topological constraints. The extracted initial fused mesh undergoes non-manifold cell removal and localized degenerate patch cleanup. The edge vertices of the patched mesh are then positionally coupled to the open boundary of the original model, resulting in a patched initial mesh that satisfies the manifold condition and is connected to the original model.

[0022] Step 5: Feature-guided construction based on the 3D sharp feature distance field regression network model. For the original model boundary and the neighborhood of the patch surface, a 3D sharp feature distance field regression network model based on local geometric fragments is used to perform learned estimation of sharp features near the contact surface, obtaining the distance values ​​from each sampling position to the nearest sharp feature, the corresponding feature direction information, and the extension trends of potential feature points and feature lines. The feature points, feature lines, and their direction information are then mapped onto the initial fusion mesh to construct feature-guided constraints for adaptive reconstruction of the patch surface region.

[0023] Step 6: Perform feature-guided adaptive reconstruction optimization on the patched area. A variational shape reconstruction method based on a quadratic error metric is used to adaptively arrange vertices and reconstruct the mesh for the newly added patched area. During the reconstruction process, the feature points, feature lines, and their extension directions obtained in Step 5 are used as guiding constraints to enable the mesh wiring and vertex distribution of the patched area to actively conform to the feature orientation. This improves the triangle quality, local detail density, and feature alignment capability while ensuring the continuity of the contact surface area.

[0024] Step 7: L1 Energy Surface Sharpening Optimization Based on the Surface Hessian Matrix. Based on the transition mesh output in Step 6, and using the feature points, feature lines, and original model boundaries obtained in Step 5 as constraints, a joint energy objective function is constructed, including data fidelity terms, boundary constraint terms, feature constraint terms, and L1 norm regularization terms based on the surface Hessian matrix. By minimizing this objective function and updating the vertex coordinates of the patching region, the contact surface region forms a piecewise smooth surface and sharp polygonal lines while maintaining the original boundary positions and feature continuity, achieving explicit feature fidelity.

[0025] Step 8: Local Topology Repair and Output Based on Restricted Power Graph. To eliminate local mesh flipping, self-intersection, holes, mesh gaps, local overlaps, T-connections, degenerate triangles, non-manifold elements, and other topological defects that may accumulate during surface reconstruction, feature propagation, and surface optimization, a restricted power graph-based method is used to reconstruct and repair local abnormal regions. The final output is a 3D model stitching result that maintains the original connectivity and is free of topological defects.

[0026] This invention proposes an automatic fusion method for the contact surfaces of 3D models based on a feature distance field regression network. Taking a polygonal mesh composed of multiple 3D models to be stitched as input, the method achieves automatic contact surface fusion, smooth transition, feature continuation, and defect repair by implicit surface patching, contact surface stitching that maintains connectivity, manifold cleaning and boundary alignment, sharp feature detection and feature-guided construction, adaptive reconstruction of the surface patching region guided by features, feature-preserving surface optimization, and post-processing topology repair. Finally, it outputs a complete, continuous, and topologically reliable 3D model.

[0027] Example 2 This embodiment provides an automatic fusion method for contact surfaces of 3D models based on a feature distance field regression network, including: Obtain multiple 3D model components to be stitched together, and organize these components into a unified polygonal mesh input. Based on the polygonal mesh input, extract the open boundary information of each model component; Based on the open boundary information, implicit surface patching is performed to generate an initial surface patching mesh that matches the boundary of the original model. Based on the initial patch mesh, perform contact surface stitching that maintains connectivity to generate an initial fused mesh that is connected to the original model; The initial fused mesh is manifold cleaned and its boundaries aligned to obtain the initial mesh for surface patching; Based on the initial mesh of the patch surface, the sharp features near the contact surface are learned and estimated using a three-dimensional sharp feature distance field regression network model to generate feature guidance constraints that include feature points, feature lines and their extension directions. Based on the feature-guided constraints, adaptive reconstruction optimization is performed on the patching region to obtain a transition mesh; based on the transition mesh, L1 energy surface sharpening optimization based on the surface Hessian matrix is ​​performed to generate a sharpened contact surface fusion surface. Local topology repair is performed on the sharpened contact surface fusion surface to output a complete 3D model.

[0028] Specifically, it includes: (1) Input Model Construction and Boundary Extraction: First, multiple 3D model components to be stitched are acquired and organized into a unified polygonal mesh input. For each model component, the open boundary, boundary vertices, boundary edges, and adjacent face information are extracted to determine the contact surface area and the area to be filled between components. The input model is denoted as: (1) in, For the input model, Indicates the first A 3D model component to be assembled. This indicates the number of model components. By performing a unified analysis of the boundaries of each model component, basic geometric constraints are provided for subsequent implicit surface patching and contact surface stitching.

[0029] (2) Sampling and Implicit Surface Completion: The surface of the input 3D model is sampled to obtain a discrete set of sampling points and corresponding normal information. An implicit surface is then generated using a Poisson surface reconstruction method. Let the set of sampling points be: (2) in, Indicates the first One sampling point, Indicates the total number of sampling points. This is the sampling point set. An implicit function field is constructed based on the sampling point set, and implicit surface patching results and an octree structure matching the input model boundary are obtained. If necessary, guiding sampling points can be added to control the connection or separation trends between components, thereby improving the rationality of contact surface patching.

[0030] (3) Connectivity-preserving contact surface stitching: For the implicit surface obtained in step (2), a connectivity-preserving double-contour stitching method is used to generate new triangular patches in the boundary gap region. During the stitching process, the original model boundary is extended to the local nodes of the octree, and connected triangular patches consistent with the original boundary are generated near the boundary; at the same time, redundant elements are trimmed to ensure that the newly added patch only covers the area to be merged. Through this process, the initial merging mesh is formed in the contact surface region without changing the connectivity and geometry of the original model. Let the newly added patch mesh be denoted as . Then the initial fusion model can be expressed as: (3) in, For the initial fusion model, To add a new mesh.

[0031] For model splicing scenarios with different cross-sectional shapes, such as Figure 2 As shown, the heterogeneous model to be spliced ​​is input ( Figure 2 a) When using the traditional smooth suturing method ( Figure 2 (b) After processing, obvious feature breaks and inward shrinkage will appear at the connection point; while the method of the present invention can achieve a smooth and feature-aligned splicing effect. Figure 2 c).

[0032] (4) Manifold Cleaning and Mesh Correction: The initial fusion model M0 is cleaned by removing non-manifold edges, non-manifold vertices, and locally degenerate triangles. For local octree nodes in the boundary neighborhood that do not meet the connectivity requirements, adjustments are made from the bottom up to ensure that the newly added patch region meets the manifold condition. After this step, a structurally continuous patch mesh that meets the requirements of subsequent optimization is obtained, providing stable input for the next stage of adaptive reconstruction and feature propagation. At the same time, the edge vertices of the patch mesh are positionally coupled or aligned with the open boundary of the original model to ensure the continuity of the connection between the patch region and the boundary of the original model.

[0033] (5) Sharp Feature Detection and Feature-Guided Construction Based on Feature Distance Field Regression Network: Sharp feature analysis is performed on the boundary and neighborhood of the original model. A three-dimensional sharp feature distance field regression network model is used to identify feature points, feature lines, and their extension trends near the contact surface. Specifically, the open boundary neighborhood and neighborhood of the original model are first used as feature perception regions. Local geometric fragments are extracted within these regions as network input data. The local geometric fragments include local point clouds or mesh vertex sets, vertex normal information, local curvature information, relative position information, and neighborhood connectivity relationships, which are used to characterize local shape changes and potential sharp structures near the contact surface.

[0034] For any sampling point in the neighborhood of the patch face Define it to the set of potential sharp feature lines. The distance field function is: (4) in, Indicates sampling point The distance to the nearest sharp feature line. By transforming the problem of identifying discrete feature points or feature lines into a continuous distance field regression problem, the model's feature localization accuracy for weak features, noisy boundaries, and irregular contact areas can be improved.

[0035] The 3D sharp feature distance field regression network model includes a local geometric feature encoding module, a distance field regression module, and a feature direction prediction module. The local geometric feature encoding module extracts multi-scale geometric features from the input local geometric segment; the distance field regression module predicts the distance from the sampling point to the nearest sharp feature; and the feature direction prediction module predicts the extension direction of the potential feature line at that location. The network output can be represented as: (5) in, Indicates the first A local geometric segment centered on a sampling point The parameter is A three-dimensional sharp feature distance field regression network model, This represents the distance field value predicted by the network. Indicates the predicted feature direction, This indicates the confidence level that the sampling point belongs to the neighborhood of sharp features.

[0036] During network training, a weighted mean squared error loss function is used to constrain the distance field regression results, and its expression is: (6) in, Indicates sampling point The corresponding true distance field label, This represents the distance-aware weights. Sampling points closer to sharp feature lines are assigned higher weights to improve the network's prediction accuracy in the feature line neighborhood.

[0037] After obtaining the predicted distance field, local minima, high-confidence sampling points, or sampling points that meet the distance threshold conditions are selected as candidate feature points. These candidate feature points are then connected to form candidate feature lines, taking into account the predicted feature direction, spatial proximity, and directional consistency. Subsequently, the feature points, feature lines, and their directional information are mapped onto the fused mesh cleaned in step (4) using nearest-neighbor projection or neighborhood interpolation, thereby forming feature guidance constraints for adaptive reconstruction and surface optimization of subsequent surface patching regions.

[0038] (6) Feature-guided adaptive reconstruction of the surface patching region: For the newly added surface patching region, a variational shape reconstruction method based on a quadratic error metric is adopted to optimize the vertex positions in the surface patching mesh and adjust the local mesh connection relationship.

[0039] Let the set of vertices of the patch region be... Then the feature-guided adaptive reconstruction objective can be expressed as: (7) in, As vertex The local quadratic error measure term, The distance field value, feature direction information and confidence level obtained from step (5) are used to control the strength of the sharp feature constraint on the vertex. The guide target position is determined by feature points, feature lines, and their extension directions. This refers to the weighting coefficient of the guiding term.

[0040] Furthermore, the local quadratic error metric term can be expressed as: (8) in, Representing a vertex The local neighborhood plane set, and They represent the first The normal vector and bias term of a local fitting plane. Equations (7) and (8) are used together to constrain the patch area to conform to the feature orientation near the contact surface while maintaining overall continuity.

[0041] This step uses the feature points, feature lines and their extension directions obtained in step (5) as guiding constraints, so that the vertex distribution, local mesh wiring and triangle division of the patching area can actively fit the feature direction near the contact surface. Under the premise of ensuring the continuity of the contact surface area, the vertex distribution density of the sharp area and the detail area is adaptively increased, thereby improving the triangle quality and local feature alignment ability, and reducing the dilution of the sharp structure during the patching process.

[0042] The feature points, feature lines, and their orientation information serve as guiding constraints in vertex placement and mesh routing during the adaptive reconstruction of the surface patching region, and further serve as feature preservation constraints in the construction of the objective function during the subsequent surface optimization process.

[0043] (7) Feature-preserving smooth fusion of the contact surface: Using the feature points, feature lines, and original model boundaries obtained in steps (5) and (6) as constraints, the surface of the patching region is optimized using an L1 sharpening and sparsification optimization method based on the surface Hessian matrix. Let the optimized vertex coordinate matrix of the patching surface be: (9) in Indicates the first in the patch region The three-dimensional coordinates of each vertex. Indicates the total number of vertices in the patch region; the Used to uniformly represent the positional state of all vertices to be optimized in the contact surface region. To optimize the variables, a surface optimization objective function is constructed. ,in This represents the L1 sparsification energy term based on the surface Hessian matrix, used to suppress disordered curvature changes at non-feature locations in the patch region; This represents the boundary constraint term, used to fix the position of the boundary vertices of the original model; These represent characteristic constraint terms, used to maintain the continuity, directional consistency, and local polyline relationships of characteristic lines; The weight coefficients corresponding to the above energy constraints are used to adjust the balance between the degree of sparsity, the strength of boundary fixation, and the feature alignment weights. Furthermore, for the boundary vertices of the original model, a position fixation constraint is applied to keep them stationary during the optimization process; for vertices located on feature lines, a direction deviation penalty and a neighborhood continuity constraint are applied to ensure that the optimized surface achieves a smooth transition at the contact surface while still maintaining sharp edges, polygonal transitions, and local boundary details, thus balancing fusion smoothness and geometric feature fidelity.

[0044] The high-fidelity technical effect produced by this step is as follows: Figure 3 As shown: Comparison with the original damage model ( Figure 3 a) Traditional methods ( Figure 3 b) The patched area exhibits a rounded, blunted state, losing its sharp features; while the method of this invention ( Figure 3 c) While achieving a smooth surface transition, the sharp corners of the original model were perfectly restored and preserved.

[0045] (8) Post-processing topology repair and result output: To eliminate holes, mesh gaps, local overlaps, T-connections, degenerate triangles, non-manifold elements, and other topological defects that may accumulate during surface reconstruction, feature propagation, and surface optimization, post-processing repair is performed on the stitched model after surface optimization. A feature-preserving mesh repair method based on restricted power diagrams is adopted to detect and correct local abnormal regions in the contact surface area and its neighborhood; at the same time, the propagated feature lines and sharp boundaries are preserved. Let the final output model be denoted as... ,in This represents the complete 3D model after automatic surface patching, feature propagation, surface optimization, and topology repair. The final output is a geometrically continuous, structurally complete, and topologically reliable 3D model.

[0046] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0047] On the other hand, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0048] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An automatic fusion method for contact surfaces of 3D models based on a feature distance field regression network, characterized in that, include: Obtain multiple 3D model components to be stitched together, and organize these components into a unified polygonal mesh input. Based on the polygonal mesh input, extract the open boundary information of each model component; Based on the open boundary information, implicit surface patching is performed to generate an initial surface patching mesh that matches the boundary of the original model. Based on the initial patch mesh, perform contact surface stitching that maintains connectivity to generate an initial fused mesh that is connected to the original model; The initial fused mesh is manifold cleaned and its boundaries aligned to obtain the initial mesh for surface patching; Based on the initial mesh of the patch surface, the sharp features near the contact surface are learned and estimated using a three-dimensional sharp feature distance field regression network model to generate feature guidance constraints that include feature points, feature lines and their extension directions. Based on the feature-guided constraints, adaptive reconstruction optimization is performed on the patching region to obtain a transition mesh; based on the transition mesh, L1 energy surface sharpening optimization based on the surface Hessian matrix is ​​performed to generate a sharpened contact surface fusion surface. Local topology repair is performed on the sharpened contact surface fusion surface to output a complete 3D model.

2. The method according to claim 1, characterized in that, The implicit surface patching process further includes: The polygonal mesh input surface is sampled to obtain a discrete set of sampling points; Based on the discrete sampling point set, an implicit surface matching the boundary of the original model and its corresponding octree structure are generated using the Poisson surface reconstruction method.

3. The method according to claim 1, characterized in that, The step of performing contact surface stitching to maintain connectivity further includes: Based on the initial patch mesh, a connectivity-preserving double contour stitching method facing the mesh boundary is used to generate new triangular patches in the boundary gap region; By extending the boundary and pruning redundant elements, the newly added triangular facets are connected to the boundary of the original model to form the initial fused mesh.

4. The method according to claim 1, characterized in that, The manifold cleaning and boundary alignment of the initial fused mesh further includes: The initial fused mesh is subjected to non-manifold cell culling and local degenerate patch cleaning to obtain a cleaned mesh; The edge vertices of the patched area in the cleaned mesh are coupled positionally with the original model open boundary corresponding to the open boundary information to obtain a patched initial mesh that satisfies the manifold condition and is connected to the original model.

5. The method according to claim 1, characterized in that, Generating the feature-guided constraints further includes: Based on the initial mesh of the patched surface, local geometric fragments within the original model boundary and the neighborhood of the patched surface are extracted; the local geometric fragments are input into the three-dimensional sharp feature distance field regression network model; The network model outputs the distance field prediction value, feature direction prediction information, and confidence level from each sampling position to the potential sharp feature through the local geometric feature encoding module, distance field regression module, and feature direction prediction module. Based on the distance field prediction value, feature direction prediction information, and confidence level, candidate feature points and candidate feature lines are generated, and the extension trend is determined to form the feature guidance constraint.

6. The method according to claim 5, characterized in that, The adaptive reconstruction optimization of the patching region further includes: Based on the feature points, feature lines and their extension directions in the feature guidance constraints, determine the guidance target position and vertex constraint weights; A variational shape reconstruction method based on a quadratic error metric is adopted, and the transition mesh is generated by adaptively adjusting the vertex arrangement and mesh connection relationship of the surface patching region according to the guiding target position and vertex constraint weight.

7. The method according to claim 1, characterized in that, The execution of L1 energy surface sharpening optimization based on the surface Hessian matrix further includes: Obtain the vertex coordinates of the patch region in the transition mesh as optimization variables; construct a joint energy objective function; The joint energy objective function includes: an L1 sparsification energy term based on the surface Hessian matrix, a boundary constraint term for fixing the position of the boundary vertices of the original model, and a feature constraint term for maintaining the continuity and orientation consistency of the feature lines. By minimizing the joint energy objective function, the vertex coordinates of the patching region are updated to generate the sharpened contact surface fusion surface.

8. The method according to claim 1, characterized in that, The process of performing local topology repair further includes: Detect local abnormal regions in the sharpened contact surface fusion surface, including holes, mesh gaps, local overlaps, T-shaped connections, degenerate triangles, or non-manifold units; A feature-preserving mesh repair method based on restricted power graphs is used to reconstruct and repair the detected local abnormal regions.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computing program, it implements the method of any one of claims 1-8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-8.

Images