Feature-driven point cloud completion method and system for cad parametric model

Abstract

The application discloses a feature-driven point cloud completion method and system for a CAD parameterized model, relates to the technical field of three-dimensional geometry processing and computer vision, and comprises the following steps: constructing a multi-source feature line collection and standardization system, obtaining feature line candidates through multiple paths such as analysis of CAD information, mesh point cloud geometric clues and interactive editing rules, and outputting a unified format feature representation after consistency evaluation, cleaning and repairing; designing a local feature extraction method based on adaptive weights of multiple relationships, dynamically adjusting weights according to the relationship between target points and neighbor points in three-dimensional space and high-dimensional feature space in a DGCNN framework, and quantifying the contribution of each neighbor point to the local feature of the target point; and proposing a feature line deep information injection technology, fusing feature line geometric information into a point cloud completion deep learning model through an encoder, using a feature line weighted chamfer distance loss function to strengthen the fidelity of boundary and acute angle regions, and effectively improving the accuracy and geometric detail retention capability of point cloud completion.

Description

Technical Field

[0001] This invention relates to the fields of three-dimensional geometry processing and computer vision technology, specifically to a feature-driven point cloud completion method and system for CAD parametric models. Background Technology

[0002] Point cloud completion technology is one of the core tasks in the field of 3D vision, aiming to recover complete 3D geometry from incomplete point cloud data. With the widespread use of 3D scanning equipment and depth sensors, point cloud data is widely used in industrial scenarios such as reverse engineering, quality inspection, and digital twins. However, due to factors such as the viewing angle limitations of scanning equipment, object self-occlusion, and surface reflection characteristics, the acquired point cloud data often has large areas of missing data, which poses a serious challenge to subsequent geometric reconstruction, dimensional measurement, and model analysis.

[0003] Existing point cloud completion methods are mainly divided into two categories: traditional geometry-based methods and deep learning-based methods. Traditional methods, such as radial basis function-based interpolation and Poisson reconstruction-based surface restoration, can achieve good results in certain scenarios, but they often rely on manually designed geometric priors and are difficult to handle complex topologies and cases with missing heights. Deep learning methods, such as PointNet, PointNet++, PCN, and PoinTr, automatically extract point cloud features and generate completion results through end-to-end learning, achieving significant progress in general point cloud completion tasks.

[0004] However, the following technical problems exist when directly applying existing point cloud completion methods to CAD parametric models: First, CAD models possess unique geometric priors. CAD models are composed of a limited set of primitives, such as planes, spheres, and cylinders, and possess distinct feature lines and boundary information. Existing general point cloud completion methods do not fully utilize these unique geometric constraints of CAD, resulting in insufficient structural fidelity in the completed models. For example, while the PoinTr method proposed by Yu et al. uses a Transformer architecture to improve feature representation, its local feature extraction still relies on simple K-nearest neighbor search, making it difficult to accurately capture local features in complex geometric regions such as sharp edges and high curvature areas, leading to a smoothing effect on the edges of the completed model.

[0005] Secondly, feature lines, as key geometric information in CAD models, are crucial for the accuracy and detail preservation of point cloud completion. Feature lines typically represent the boundaries and curvature changes of a model, acting as a skeleton in reverse engineering and parametric reconstruction. While some existing techniques attempt to incorporate feature line information into point cloud processing—such as CurveNet enhancing feature representation through curvature multi-task learning and Points2Surf learning implicit surface representations—these methods lack a deep fusion mechanism between feature lines and point cloud completion tasks. Specifically, existing methods suffer from three shortcomings: at the feature line acquisition level, there is a lack of a unified multi-source feature line standardization system; feature lines from different sources differ in sampling density, noise level, and topological consistency, directly affecting subsequent completion quality. At the feature extraction level, local feature extractors do not consider the guiding role of feature lines in contributing to neighboring points, resulting in inaccurate feature representation in high-curvature regions. At the information injection level, feature line information is often integrated into the network through simple splicing, lacking a deep cross-modal interaction mechanism and failing to fully leverage the geometric constraints of feature lines.

[0006] Third, the loss function design is not optimized for the geometric characteristics of CAD models. Existing methods generally use standard chamfer distance as the loss function, giving equal weight to all points. However, in CAD models, geometric details near feature lines have a greater impact on the model's engineering usability. For example, the fidelity of sharp edges and corners directly determines the assembly accuracy of parts. Without special processing of feature line areas, the completion result is prone to bulges, collapses, and other phenomena, reducing the model's engineering application value.

[0007] To address the aforementioned issues, this invention proposes a feature-driven point cloud completion method for CAD parametric models. By constructing a multi-source feature line acquisition and standardization system, designing a local feature extraction method with multi-relationship adaptive weights, proposing a deep information injection technology for feature lines, and employing a feature line-weighted chamfer distance loss function, the accuracy of point cloud completion and the ability to preserve geometric details are effectively improved. Summary of the Invention

[0008] Purpose of the invention: To provide a feature-driven point cloud completion method and system for CAD parametric models, and to solve the problems mentioned above.

[0009] Technical solution: A feature-driven point cloud completion method and system for CAD parametric models, including the following steps: Step S1: Construct a multi-source feature line acquisition and standardization system to obtain feature line data from the CAD model. The sources of the feature line data include CAD information parsing and extraction, automatic detection based on grids or point clouds, and interactive editing and rule generation. Step S2 involves performing consistency evaluation, cleaning, repair, and standardization on the feature line data obtained in step S1, and outputting a feature representation in a unified format, which includes a fixed-length point sequence, topological relationships, and metadata. Step S3: Based on the local feature extraction method of multi-relationship adaptive weight, local features are extracted from the incomplete point cloud. Under the DGCNN framework, the contribution weight of each neighboring point to the local features of the target point is dynamically adjusted according to the geometric relationship measurement between the target point and its neighboring points in the three-dimensional space and the similarity relationship measurement in the high-dimensional feature space. Step S4: The feature line information obtained in step S2 is fused into the point cloud completion deep learning model through the feature line deep information injection technology. The feature lines are encoded by the SplineLineEncoder to obtain line-level token features, and the line-level token features are interacted with the point cloud proxy point features. Step S5: The point cloud completion model is trained and optimized using a feature line-weighted chamfer distance loss function, and points closer to the feature lines are given higher weights in the loss calculation.

[0010] In a further embodiment, in step S1, the multi-source feature line acquisition and standardization system supports manual interactive generation of quadrilateral grid layout. The user specifies key feature points or feature lines on the triangular grid, and the algorithm automatically generates a geodesic network to form a coherent quadrilateral layout based on the input. In the geodesic calculation process, a local optimization strategy based on edge flipping is introduced to minimize the path length by continuously flipping the edges of the triangular grid so that the path approximates the geodesic.

[0011] In a further embodiment, step S1 also includes a symmetry-sensitive mechanism, whereby the user manually marks symmetry points on the model to guide the generation of the model's symmetry plane or symmetry axis. When the user marks or draws geodesics on one side, the symmetry module automatically copies the corresponding operation on the other side to maintain symmetry consistency.

[0012] In a further embodiment, step S1 also includes automated quadrilateral mesh layout generation. After the user provides the curvature alignment information of some facets of the model, a smooth cross field and corresponding singularity are generated according to the alignment constraints, and a seamless surface parameterization is generated. Guided by the seamless parameterization result, the virtual particles start from the singularity and travel along the boundary line direction in the cross field, leaving a trajectory on the model to form a motorcycle diagram. The motorcycle diagram divides the mesh model into several quadrilateral mesh regions to form a feature line frame.

[0013] In a further embodiment, step S2, which involves consistency assessment, cleaning, repair, and standardization of the feature line data, specifically includes: establishing a local coordinate system and error metric for each curve using its analytical definition; calculating the residual distribution on the grid vertices corresponding to the curve; determining the reliability using robust statistical methods; and classifying the curves into three categories: OK, SUSPECT, and BAD. For OK curves, the error distribution is concentrated, and clean, equal-arc-length point sequences are obtained by resampling directly according to the analytical curve. For SUSPECT curves, which are generally consistent but have a small number of outliers, the outliers are removed or projected back onto the curve without destroying the topology. For BAD curves, if there is an overall mismatch, the analytical parameters are abandoned, and a polyline approximation on the grid or a local fitting is performed instead.

[0014] In a further embodiment, in step S3, the geometric relationship metric is calculated by the distance between the target point and its neighboring points in three-dimensional space, and the high-dimensional feature relationship metric is calculated by the feature similarity between the neighboring points and the target point. The contribution weight is dynamically determined based on the weighted combination of the geometric relationship metric and the high-dimensional feature relationship metric. The geometric relationship metric reflects the spatial positional differences between points, while the high-dimensional feature relationship metric reflects the feature representation differences between points. In a further embodiment, the workflow of the SplineLineEncoder in step S4 includes: constructing a local frame for each point of the feature line that has been parameterized by arc length; projecting the relative displacement onto the local coordinate system; building a point-level initial embedding containing the three-dimensional coordinates, curvature, tangency, parameterized values, and closure information of each feature line point; and edge-level geometric embedding for modeling the relationship between adjacent points, including arc length, rate of change of curvature, rotation angle, and frame information; using SplineGATBlock to aggregate information on the expanded adjacency of each point using geometric attention; implementing the dilated TCN branch in parallel to capture longer-range patterns; finally, combining the TCN branch with the SplineGAT main branch to project and obtain the feature representation for the next round; after multiple iterations, obtaining the final feature representation of all points; using CLS line-level attention aggregation to obtain preliminary line-level token features; and then incorporating the closure markers and global statistics through MLP to obtain the final line-level features.

[0015] In a further embodiment, step S4, which involves exchanging information between the line-level token features and the point cloud proxy point features, specifically includes: calculating the shortest distance from each PoinTr proxy point to the multiple feature lines with the shortest distance, then using RBF to weight these distances, aggregating the feature vectors of each feature line according to the weighted distance to obtain the adjustment features of each proxy point, and then using feature-level linear modulation to perform the final information exchange between the point proxy and the line-level token.

[0016] In a further embodiment, the feature line weighted chamfer distance loss function in step S5 assigns a weight to each point in the point cloud based on its distance from the nearest feature line. Points closer to the feature line have a higher weight, and points farther from the feature line have a lower weight. The contribution ratio of the feature line region in the loss calculation is adjusted by the weight coefficient.

[0017] This invention also provides a feature-driven point cloud completion method for CAD parametric models, comprising: The multi-source feature line acquisition and standardization module is used to acquire feature line data from CAD models and perform consistency assessment, cleaning, repair and standardization processing; The local feature extraction module is used to extract local features from incomplete point clouds based on multi-relation adaptive weights. The feature line information injection module is used to fuse feature line geometric information into the point cloud completion deep learning model; The loss optimization module is used to train and optimize the point cloud completion model using a feature line-weighted chamfer distance loss function.

[0018] The beneficial effects of this invention are as follows: First, by constructing a multi-source feature line acquisition and standardization system, this system uniformly supports multiple acquisition paths, including parsed CAD information extraction, automatic detection based on mesh point clouds, and interactive editing and rule generation. It also performs consistency evaluation, cleaning and repair, and scale control on the acquisition results to form standardized exported results, providing high-quality, reproducible geometric priors for subsequent model completion. Even when facing complex boundaries, local missing features, and noise interference, this system can still maintain the structural rationality of feature lines through cleaning and repair, thus providing a high-quality data foundation for improving the accuracy and robustness of the completed model under geometric constraints.

[0019] Secondly, by designing a local feature extraction method based on multi-relationship adaptive weights, this method dynamically adjusts the weights according to the relationship between the target point and its neighboring points in 3D space and high-dimensional feature space, building upon the DGCNN framework. It considers not only the geometric relationships between points but also their similarity in high-dimensional feature space and their geometric relationship with surrounding splines, enabling more accurate capture of local geometric features. This method maintains computational efficiency while allowing for more refined feature extraction within local regions, enhancing feature expressiveness and thus more accurately capturing the geometric details of the point cloud sampled from the CAD model, ultimately improving the structural integrity of the completed point cloud.

[0020] Third, by proposing a deep information injection technique for feature lines, a SplineLineEncoder is used to encode feature lines to obtain line-level token features. These line-level token features are then interacted with point cloud proxy point features through a cross-attention mechanism. This allows the information of the feature lines to be continuously strengthened during the point cloud completion process, effectively preserving the geometric details of the feature lines. This technique not only improves the overall accuracy of point cloud completion but also enhances the geometric continuity and structural consistency of the completion results, particularly in complex CAD models and highly symmetric models.

[0021] Fourth, by adopting a feature line-weighted chamfer distance loss function, points close to the feature lines are given higher weights in the loss calculation, which significantly improves the fidelity of the boundary and acute angle regions, making the reconstructed model more accurate. Especially in the case of missing or occluded features, it can better recover the feature line structure, avoid common bulges and collapses, and improve the usability and accuracy of the completion results. Attached Figure Description

[0022] Figure 1 This is a schematic diagram illustrating the use of the model's mirror symmetry to draw a quadrilateral mesh layout. Figure 2 This is a schematic diagram illustrating the rendering of a quadrilateral mesh layout using the model's rotational symmetry. Figure 3 A schematic diagram of automatically generated cross fields and subdivision lines on the model; Figure 4 Generate a schematic diagram of the quadrilateral layout and its corresponding spline, surface, and control point structure on the model; Figure 5 A statistical chart showing the types of curves, the proportion of sharp edges, and the number of feature surfaces for data block 0. Figure 6 Example diagram of point cloud and feature line frame of CAD model in ABC; Figure 7 Architecture diagram for multi-metric-aware local feature extraction; Figure 8 The comparison chart shows that the partial model with the introduction of multiple metrics has better completion performance compared to the basic PoinTr. Figure 9 Example diagram of the dot-line hit relationship and field amplitude of a hexagonal nut; Figure 10 This is a Top-K statistical chart of the most recent curve hits for the nut model. Figure 11 Example diagrams of distance distribution histograms and weight distribution histograms for the nut model; Figure 12The flowchart illustrates the feature-driven point cloud completion method for CAD parametric models provided in this embodiment of the invention. Detailed Implementation

[0023] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0024] This invention mainly focuses on feature-driven point cloud completion methods for parametric CAD models, and emphasizes three core issues: a CAD model feature extraction system based on a quadrilateral mesh layout, local feature extraction methods, and deep information injection technology for feature lines. The specific content is as follows: (1) Multi-source feature line acquisition and standardization system for CAD parametric models In CAD model processing, feature line extraction and representation are crucial steps connecting geometric semantics with subsequent learning tasks. Since feature lines may originate from different sources, such as parsed CAD information, mesh / point cloud geometric cues, and interactive editing and rule generation, and these sources differ in sampling density, noise level, and topological consistency, this invention proposes a feature line acquisition system integrating manual, automatic, and spline generation methods. This system follows a "multi-source acquisition—consistency assessment—cleaning and repair—standardized export" approach: on the one hand, it comprehensively utilizes geometric constraints and structural priors to obtain candidate feature lines, supplemented by structured mesh generation and spline fitting when necessary; on the other hand, it suppresses the influence of anomalies and outliers through robust consistency checks and quality grading, ensuring the usability and reproducibility of the feature line data.

[0025] This system supports automatic and interactive feature line acquisition and correction. It can combine the geometric information and feature line data of CAD models to output feature representations in a unified format (such as fixed-length point sequences, topological relationships, and metadata), providing stable geometric priors for subsequent feature-driven point cloud completion. Even when facing complex boundaries, local missing features, and noise interference, this method can maintain the structural rationality of feature lines through cleaning and repair, thus providing a high-quality data foundation for improving the accuracy and robustness of the completed model under geometric constraints.

[0026] (2) Local feature extraction method One of the key aspects of point cloud completion is the efficient and accurate extraction of local features. Traditional local feature extraction methods mostly rely on simple convolutional-pooling networks or K-nearest neighbor algorithms, but these methods have limitations when dealing with complex geometric features. To overcome these problems, this invention proposes a local feature extraction method based on multiple relationships. This method quantifies the contribution of each neighboring point to the local features of the target point by establishing a finer-grained relationship metric between the target point and its neighboring points.

[0027] Specifically, this invention introduces an adaptive contribution weight mechanism. Based on the DGCNN framework, it dynamically adjusts weights according to the relationship between the target point and its neighboring points in 3D space and high-dimensional feature space. This method not only considers the geometric relationships between points but also combines their similarity in high-dimensional feature space and their geometric relationship with surrounding splines, thus enabling more accurate capture of local geometric features. In this way, the feature extractor can more accurately mine local feature information such as boundaries and curvature from the point cloud collected from the original CAD model, thereby significantly improving the performance of subsequent modules.

[0028] (3) Research on feature line information injection technology Feature lines are important geometric features in CAD models, typically used to represent model boundaries and curvature variations. To further improve the accuracy and geometric detail of point cloud completion, this invention proposes a deep feature line information injection technique. By deeply fusing the geometric information of feature lines into the point cloud completion process, this technique enhances the model's ability to perceive complex geometric structures.

[0029] The core innovation of this invention lies in how to effectively inject feature line information into a deep learning model for point cloud completion. By designing a multi-level information injection mechanism based on feature lines, combined with geometric constraints and deep neural networks, the information of feature lines can be continuously strengthened during the point cloud completion process, thereby effectively preserving the geometric details of the feature lines. This technique not only improves the overall accuracy of point cloud completion but also enhances the geometric continuity and structural consistency of the completion results, especially in complex CAD models and highly symmetric models.

[0030] Specifically, the present invention includes the following: like Figure 12 As shown, a feature-driven point cloud completion method for CAD parametric models includes: Step S1: Construct a multi-source feature line acquisition and standardization system to obtain feature line data from the CAD model. The sources of the feature line data include CAD information parsing and extraction, automatic detection based on grids or point clouds, and interactive editing and rule generation. Step S2 involves performing consistency evaluation, cleaning, repair, and standardization on the feature line data obtained in step S1, and outputting a feature representation in a unified format, which includes a fixed-length point sequence, topological relationships, and metadata. Step S3: Based on the local feature extraction method of multi-relationship adaptive weight, local features are extracted from the incomplete point cloud. Under the DGCNN framework, the contribution weight of each neighboring point to the local features of the target point is dynamically adjusted according to the geometric relationship measurement between the target point and its neighboring points in the three-dimensional space and the similarity relationship measurement in the high-dimensional feature space. Step S4: The feature line information obtained in step S2 is fused into the point cloud completion deep learning model through the feature line deep information injection technology. The feature lines are encoded by the SplineLineEncoder to obtain line-level token features, and the line-level token features are interacted with the point cloud proxy point features. Step S5: The point cloud completion model is trained and optimized using a feature line-weighted chamfer distance loss function, and points closer to the feature lines are given higher weights in the loss calculation.

[0031] A feature-driven point cloud completion system for CAD parametric models includes: The multi-source feature line acquisition and standardization module is used to acquire feature line data from CAD models and perform consistency assessment, cleaning, repair and standardization processing; The local feature extraction module is used to extract local features from incomplete point clouds based on multi-relation adaptive weights. The feature line information injection module is used to fuse feature line geometric information into the point cloud completion deep learning model; The loss optimization module is used to train and optimize the point cloud completion model using a feature line-weighted chamfer distance loss function.

[0032] Example 1: CAD Model Feature Extraction System Based on Quadrilateral Grid Layout like Figure 1 and Figure 2 As shown, quad layout and feature line extraction are closely related in 3D geometric modeling. QuadLayout is a technique that reorganizes a coarse, complex initial triangular mesh structure into a regularized topological framework composed of quadrilateral units. By generating a coherent network of geodesics or streamline structures on the surface, it makes the model more controllable and consistent in subsequent parameterization, texture mapping, and surface fitting. These streamline structures are often generated with reference to the model's inherent morphological contours and functional structures, corresponding to significant changes in the model's shape, such as edges, polylines, and curvature extrema, thus revealing the object's characteristics.

[0033] To this end, this invention integrates a variety of feature extraction methods based on quadrilateral grid layouts to facilitate users in generating models and their feature line data according to their own needs, thereby facilitating subsequent pipeline work.

[0034] First, this invention integrates a method for manually and interactively generating quadrilateral mesh layouts. The core idea of ​​this method is to allow users to manually specify key feature points or feature lines on a triangular mesh, and the algorithm automatically generates a corresponding geodesic network based on these inputs. These geodesics combine to form a coherent quadrilateral layout. During the geodesic calculation process, the system introduces a local optimization strategy based on edge flipping. By continuously flipping the edges of the triangular mesh, the path length is minimized, allowing the path to gradually approach the geodesics. Compared to traditional iterative optimization or global energy minimization methods, this edge flipping strategy can quickly obtain the local shortest path while maintaining the path topology. The quadrilateral mesh generated in this way better fits the geometric features of the model, especially in sharp edges or high curvature regions, naturally capturing the structural orientation of the model.

[0035] In particular, this invention integrates a symmetry-sensitive mechanism: users can manually annotate several symmetry points on the model to guide the generation of the model's symmetry plane or axis of symmetry, thereby identifying mirror-symmetric or rotationally symmetric regions within the model. When the user annotates or draws geodesics on one side, the symmetry module automatically replicates the corresponding operation on the other side, thus maintaining symmetry consistency across the entire model. This design significantly reduces manual intervention while ensuring the geometric and topological symmetry of the generated mesh. This is especially important for highly symmetric objects such as CAD models or mechanical parts, as these objects often require strict symmetry to be maintained in subsequent parametric or NURBS fitting processes.

[0036] At the same time, such as Figure 3 As shown, this invention also provides an automated quadrilateral mesh layout generation tool. Users can provide curvature alignment information for partial facets of the model and generate smooth cross fields and corresponding singularities based on relevant alignment constraints, thereby generating seamless surface parameterization. Guided by the seamless parameterization result, several virtual particles start from their singularities, travel along the boundary lines in their cross fields, and leave corresponding trajectories on the model. When a particle encounters a trajectory left by any particle, it stops. The graph structure formed by this trajectory is the motorcycle graph. The motorcycle graph divides the mesh model into several quadrilateral mesh regions, forming the feature line framework of the model. Integer programming methods can be introduced to iteratively add, delete, and cross edges in the motorcycle graph, thereby obtaining a quadrilateral mesh layout without T nodes.

[0037] Given a quadrilateral layout, the core of generating spline curves lies in transforming the regular quadrilateral blocks within the layout into continuous, parameterizable surfaces. Each quadrilateral block can be considered a local parameter domain, with its four vertices and adjacent blocks collectively defining the surface's boundary conditions. By parametrically mapping these local blocks, a set of control points can be established for each block, forming the basic structure for surface fitting. Subsequently, the algorithm uses these control points to generate B-splines or NURBS curves. The control points determine the local shape of the curve, while the basis functions ensure a smooth transition between the control points. By interpolating or fitting these control points within each quadrilateral block, a globally continuous and locally adjustable spline surface can be obtained. Introducing NURBS weight parameters can further control the curvature of the surface, making it more closely match the geometric features of the CAD model.

[0038] Ultimately, as Figure 4 As shown, these spline curves collectively define the characteristic structure of the surface. They are typically distributed along the principal curvature directions or feature edges of the model, ensuring that the generated surface maintains both shape smoothness and accurate geometric representation. In other words, the quadrilateral layout provides structural constraints for spline generation, while the parameterization of the splines endows the surface with greater continuity and expressive power.

[0039] Example 2: Model Geometric Data Cleaning Method To more controllably and objectively evaluate the point cloud completion pipeline results, this invention prefers to collect data from publicly available geometric learning datasets. Therefore, this invention will consider using the ABC dataset for high-quality data cleaning. In the ABC dataset, each curve and surface of each valid model is given an explicit parameterized form, along with information on the correspondence between discrete grid sampling points and parameter domains. The dataset contains five types of feature lines: Line: Straight line The parameter is the location vector (location( ) and direction vector ( ).

[0040] Circle The parameter is the center location( radius () ), plane normal z_axis, and local coordinate axes x_axis(x), y_axis(y).

[0041] Ellipse: an ellipse The parameters include the two focal points focus1 / focus2, the major and minor axes x_axis / y_axis, the plane normal z_axis, and the principal / secondary radii x_radius(rx) and y_radius(ry).

[0042] BSpline: A spline curve defined by control points and node vectors. Parameters include rational / closed / continuity / degree / poles / knots. If it is a rational curve, it also includes weights.

[0043] Other: Other curve types that are not subdivided or supported, typically only retaining general fields corresponding to the samples. Since they account for a very small percentage of the model and cannot be parsed for resampling, models containing this type of curve are ignored by default.

[0044] In addition, such as Figure 5 As shown, each curve simultaneously records indicators such as type, whether it is sharp, vertex index (vert_indices) on the mesh model, and parameter values ​​(vert_parameters) for each vertex.

[0045] However, the feature line data in the ABC dataset contains errors, making the analytical information unreliable. These errors mainly come from the intermediate conversion link from CAD / B-Rep to triangular mesh: on the one hand, analytical curves / surfaces in STEP / B-Rep usually have trimming, periodic parameter domains, and seams, which will result in unpacking, segmentation marking, or parameter encoding during export; on the other hand, tessellation will make discrete approximations of analytical geometry under a given tolerance, and shared endpoints / intersections often produce "compromise vertices" in order to maintain topological connectivity, causing a small number of vertices to no longer fall precisely on the analytical curves; coupled with a few mismatches / mislabelings in the index or type, the final result is phenomena such as "most points are aligned but outliers are mixed in" or "analytical parameters are inconsistent with the overall mesh points".

[0046] Therefore, as Figure 6As shown, when using the feature line data of ABC, this invention does not assume that the analytical parameters (such as circle center / radius, spline nodes / control points, surface parameters, etc.) are geometric true values ​​that are "strictly aligned with the triangular mesh" by default. Instead, it adopts a strategy of "curve-by-curve quality assessment + diversion and cleaning" to construct usable feature line data. The specific approach is as follows: For each curve, first establish a local coordinate system and error metrics (such as the normal distance from a point to the curve, the deviation from the plane containing the curve, and the consistency of the parameter domain) using its analytical definition, and calculate the residual distribution on the grid vertices corresponding to the curve's vert_indices (including possible segmentation markers); then use robust statistics (such as median, MAD, p90, inlierratio) instead of max / min to determine the confidence level, thereby classifying the curves into three categories: OK, SUSPECT, and BAD. OK curves have a concentrated error distribution and can be directly resampled according to the analytical curve / surface to obtain a clean sequence of points with equal arc lengths; SUSPECT curves are generally consistent but have a small number of outliers, so the outliers are removed or projected back to the curve without destroying the topology; BAD curves are mismatched overall, so the analytical parameters are abandoned, and a polyline approximation on the grid is used instead (resampled by arc length) or local fitting / recovery is performed with the help of higher-level information. By using this cleaning and diversion method with robust consistency as its core, we can maximize the retention of the analytical geometric advantages available in ABC while suppressing the pollution caused by a small number of incorrect labels or local mismatches in subsequent feature line modeling and information propagation.

[0047] Example 3: Point Cloud Local Feature Extraction Method Unlike traditional point cloud completion work that focuses on performance under simple point cloud input, this invention focuses on point cloud completion tasks based on geometric priors such as feature line injection, rather than feature engineering at the deep neural network level. This approach is suitable for research on adding pluggable modules on a fixed, classic baseline, allowing for objective evaluation of the effects of introducing geometric priors. This invention will improve upon the PoinTr framework or its variants to meet this requirement.

[0048] To address this invention, a sufficiently effective local feature extraction method is needed. While existing technologies such as PointNet and PointNet++ have made significant progress in feature extraction from raw point clouds, these methods typically rely on simple K-NN search, making it difficult to achieve a balance between global and local features. To overcome these problems, this invention plans to introduce a novel relation-based local feature extraction method, the specific implementation of which is as follows: like Figure 7As shown, traditional graph neural network local feature extraction methods only consider establishing neighborhoods for each feature point in the geometric or feature space for perception. This invention, however, further explores the establishment of a relationship metric between the target point and its neighboring points to quantify the contribution of each neighboring point to the local features of the target point in more detail. This method not only considers the geometric relationships between points in detail but also their similarity in the high-dimensional feature space.

[0049] Specifically, for a target point Its neighborhood points can be determined through K-NN search. Each neighborhood point The relationship with the target point can be measured simultaneously through geometric relationships. Relationship measurement with high-dimensional features To quantify. Geometric relation measurement. It can be calculated using the Manhattan distance between the target point and its neighboring points in three-dimensional space, while high-dimensional feature relationship measurement This is calculated using the average relative relationship between the target point and its neighboring points. These relationship metrics are used to determine the contribution weight of each neighboring point to the target point's features, as shown in the following formula:

[0050]

[0051] in This represents the vector difference between two points in the feature space. From a lower-level perspective, It is a statistical measure of the absolute value of spatial differences, which can be regarded as a simplified estimate of local geometric changes and helps to better distinguish between actual edges and symmetrical structures. It can be viewed as a statistic of first-order difference, a simplified estimate of local curvature, and can be used to guide the network to focus on high-curvature regions, achieving an edge enhancement effect similar to the Laplacian operator. In the novel feature extraction graph neural network, the target point... The features are calculated by combining the features of its neighboring points and contribution weights. The extracted features are then concatenated into the original graph neural network feature extractor. This method, while maintaining computational efficiency, allows the model to perform more refined feature extraction within local regions, enhancing the expressive power of the features and thus more accurately capturing the geometric details of the point cloud sampled from the CAD model, ultimately improving the structural integrity of the completed point cloud. Experimental results on a classic point cloud completion dataset are shown below, demonstrating that the multi-metric-aware local feature extraction architecture is beneficial for feature extraction:

[0052] Table 1: Multi-metric perceptual feature extraction performance Based on this, for cases where feature lines exist, a similar approach can be used to access the field-aware features of the nearby feature lines in the original point cloud, which is beneficial for subsequent information exchange.

[0053] Specifically, for any feature line, it is first sampled, and its arc length is parameterized simultaneously. The feature line is then transformed into a sequence of sampling points and a sequence of broken line segments. During the sampling process, features such as curvature at these geometric units are also collected.

[0054] After sampling, as shown in the figure, points in each incomplete point cloud set will have more geometric features added to their 3D coordinates: the distance between the nearest feature lines. (The search can be accelerated using structures such as KD-trees), projected arc length Nearest line curvature This method involves techniques such as line ID embedding. The combined geometric features are mapped to a high dimension via an MLP, and then processed through the EdgeConv structure in the local feature extractor. This method does not involve deep information exchange, and its impact on performance requires further experimental verification in conjunction with the following feature line deep information injection techniques.

[0055] Example 4: Feature Line Information Injection Technology In this invention, incomplete point clouds are denoted as Class A point sets (partial point clouds from scanning / sampling, with a default number of 2048 points), and feature line sampling points are denoted as Class B point sets (curve / boundary sampling points from CAD models, containing three-dimensional coordinates and possible additional attributes). It is necessary to fuse these two types of heterogeneous information.

[0056] To evaluate the effectiveness of the most basic injection method of "simply concatenating feature line points as homogeneous points," this invention constructs a set of homogeneous point injection baselines that do not introduce heterogeneous structures, do not employ line topology, and do not use differential geometric features. All baselines maintain the PoinTr main network structure and training process consistent with each other, only changing the way the input point set is constructed.

[0057] (1) A-only baseline (without using feature line points) PoinTr(A-only, scratch): Trains from scratch on the ABC dataset, taking only incomplete point clouds of class A as input, and is used to measure the basic performance under the condition of training from scratch.

[0058] PoinTr(A-only,zero-shot): Directly loads pre-trained weights from public datasets such as ShapeNet55, performs no training on ABC, and directly tests on A-class point clouds, using only A-class point clouds as input, to measure the lower bound and reference for cross-dataset generalization.

[0059] PoinTr(A-only, finetune): After loading pre-trained weights, perform standard fine-tuning on ABC, with only A-class point cloud as input, used to measure the improvement brought by pre-training.

[0060] (2) Injecting baseline with homogeneous points: UnionPoinTr(A∪B,xyz-only) To ensure the interpretability of the "homogeneous point splicing" itself, the Union baseline only uses point coordinates (x, y, z), without providing additional information such as line sequence order, line segment connections, or differential geometric quantities like tangent and curvature, and without altering the network structure. This type of baseline exists in three forms, which contrast one-to-one with A-only in terms of training settings: UnionPoinTr(A∪B,scratch): Train from scratch, with the input being the concatenation result of homogeneous points of A and B; UnionPoinTr(A∪B,zero-shot): Loads pre-trained weights, tests directly on ABC without training, and takes the set of splicing points as input. UnionPoinTr(A∪B, finetune): Fine-tunes on ABC after loading pre-trained weights, with the input being the set of splicing points.

[0061] (3) Two implementation strategies and design intentions for homogeneous point injection Considering that "how to control the point count budget after concatenating B-class points" will significantly affect the model's computational cost and information ratio, the Union baseline is further subdivided into two implementations (both do not change the network, only the point set construction): Budget-matched (Union-2048): After merging the point sets of A and B, FPS is applied to the union, downsampling to 2048 points that are completely identical to A-only. This is to evaluate whether "replacing some ordinary points with some feature line points" brings gains under the condition that the number of input points and computational cost are completely comparable, thereby isolating the two types of factors: "line points are more informative" and "more points / more computational power".

[0062] Budget-augmented (Union-3072): First, FPS the B-class points separately to 1024 points (if the original number of points is less than 1024, the original number of points remains unchanged), then directly concatenate it with the 2048 points of the A-class, resulting in an input of up to 3072 points. This is used to evaluate whether explicitly allocating a fixed budget to the B-class points without sacrificing the integrity of the A-class points is more effective in completing the model, and to observe the model's sensitivity to the proportion of line points or the expansion of the number of points. It can be regarded as an upper limit or enhanced version of homogeneous point injection.

[0063] This invention selects approximately 3226 models from ABCblock0 to form an experimental subset. The selection criteria are: single component, number of vertices ≤ 20000, and number of feature lines ≤ 256, and the subset is divided into a train:test ratio of 4:1. This size is similar to the "maximum number of samples per class" of many common point cloud datasets, facilitating rapid iteration during data cleaning and method validation. It should be noted that the training at this stage does not fully approach the upper limit of the model's capabilities; therefore, it is more suitable for observing the relative trends and gain patterns between different strategies, rather than the final upper limit performance. The experimental results of this invention are shown below:

[0064] Table 2: A-only baseline results, with the default difficulty set to normal.

[0065] Table 3: Baseline of Homogeneous Point Injection, Default Difficulty is Normal

[0066] Table 4: Results of U2048+FT and U3072+FT at three difficulty levels in: ZS: Zero-shot, which means loading ShapeNet55 pre-trained weights and testing directly on ABC without training; FT: Finetune, which means fine-tuning on ABC after loading ShapeNet55; U2048: After A and B are spliced ​​together, the overall FPS drops back to 2048 (budget alignment). U3072: B first reduces its FPS to less than or equal to 1024, then concatenates it with A (2048), resulting in a maximum of 3072 (budget expansion). Easy, Median, and Hard: These are different test difficulties, corresponding to a cropping percentage of 25%, 50%, and 75%, respectively.

[0067] As can be seen, under the condition of only inputting incomplete point clouds, pre-training and fine-tuning bring stable and significant overall improvement: from zero training to zero samples, the F-Score increased from 0.426 to 0.493, and the CD-L1 decreased from 17.802 to 14.867, indicating that pre-training has a certain cross-domain generalization ability; on this basis, after further fine-tuning the ABC subset, the F-Score increased to 0.546 and the CD-L1 decreased to 11.865, indicating that "pre-training + fine-tuning" is still one of the key means to obtain better completion quality, and provides a stable reference baseline for all subsequent injection strategies.

[0068] Without altering the network structure or providing line topology or differential geometric properties, injecting feature line sampling points as homogeneous point injection input still yields reliable gains. In particular, U2048 shows improvement even with the same number of input points as A-only: the F-Score improves from 0.426 to 0.449 in Scratch and from 0.546 to 0.558 in FT, while CD-L1 also significantly decreases. This indicates that the gains do not stem from "more points / more computing power," but rather from the higher information density of line points in structural representation, such as providing more concentrated structural cues near sharp edges and boundaries. This verifies the effectiveness of "homogeneous point injection" as the most basic line information fusion method.

[0069] The budget-enhanced version U3072 delivers more significant gains under the Fourier Transform (FT) condition: the F-Score reaches 0.591, CD-L1 decreases to 9.675, and CD-L2 further decreases to 0.422, outperforming U2048+FT overall. This indicates that explicitly allocating additional budget to line points without sacrificing A-point coverage can further enhance the model's ability to utilize structural priors. Meanwhile, U3072 can also be considered an upper limit for homogeneous point injection: if more complex line features or heterogeneous fusion modules are introduced subsequently, their improvement needs to be compared with U3072 to avoid misinterpreting the increased point budget as a fundamental benefit from structural fusion.

[0070] It's worth noting that the injection strategy doesn't always yield monotonic benefits under zero-sample conditions: the ZS for A-only is 0.493, while U2048's ZS is only 0.472; however, U3072's ZS is 0.510, outperforming A-only. This phenomenon suggests that pre-trained models are sensitive to the input point distribution: the overall regression after concatenation in U2048 may have altered the point distribution or sampling structure, making it difficult for pre-trained weights to generalize directly; while U3072, by concatenating additional points while preserving the complete A-point distribution, may have a smaller distribution shift and thus benefit more easily. To fully leverage the advantages of pre-training, the impact of the injection process on the input distribution needs to be explicitly considered, such as through hybrid sampling / augmentation during the training phase.

[0071] In PoinTr's easy / median / hard clipping difficulty tests, both Fourier Transform (FT) models showed a reasonable trend of performance decline with increasing difficulty. However, U3072+FT outperformed U2048+FT at all three difficulty levels, with the advantage being more pronounced in the Hard level: the F-Score improved from 0.544 to 0.585, and the CD-L1 score decreased from 11.242 to 10.346 in Hard mode. This result indicates that when there are fewer visible points and more severe missing points, the additional line point budget can better play the role of structural constraints and shape cues, improving the model's robustness to severely incomplete inputs, which is consistent with the motivation for feature line injection.

[0072] In summary, the current interim results offer direct insights for data cleaning and subsequent experiments: Firstly, the improvement in U2048 under conditions of comparable point counts proves that line points do indeed possess effective information, supporting the priority retention of higher-quality feature line data during cleaning. Secondly, U3072 shows greater gains under Hard difficulty, suggesting that future efforts should focus more on long-tailed complex samples and scenarios with high missing data to more fully demonstrate the value of the injection strategy. Finally, since the training has not yet reached its limit, these conclusions are more suitable as "trend verification and control." Further testing and amplification of the true gain of the injection strategy are still possible through more thorough training, larger-scale samples, or relaxed screening criteria.

[0073] However, under actual feature line constraints, spline sampling points are not ordinary scattered points: they are ordered along the curve and contain strong geometric priors, such as adjacency, arc length, bends, and closure. Therefore, an efficient feature extractor must be provided for the spline sampling point sequence. To this end, this invention proposes a geometry-aware spline encoder: SplineLineEncoder (SLE). SLE can condense the point-level representation of each open / closed curve spline while maintaining robustness to variable-length sequences. And aggregated into a single line-level token. It can depict local bends, represent the direction of the entire line, and serve as a proxy point / query point for the prior deconditioning of PoinTr.

[0074] For a feature line whose arc length has been parameterized, construct a local frame for each point. Projecting the relative displacement onto the local coordinate system:

[0075] Subsequently, a point-level initial embedding is constructed, including the 3D coordinates, curvature, tangent, parameter values, and closure information of each Class B point:

[0076] And edge-level geometric embedding, used for modeling relationships between adjacent points, including information such as arc length, rate of change of curvature, rotation angle, and frame:

[0077] Subsequently, using the proposed SplineGATBlock, the adjacency of each point is expanded (here, a 2-neighborhood is taken, i.e., numbered as...). Information aggregation is achieved using geometric attention at points (i.e., points). in This approach utilizes a lightweight MLP to map edge features to attention biases. RBF or Fourier encoding representing the relative position of the arc length provides prior knowledge of the position along the curve.

[0078] Simultaneously, an inflated TCN branch is implemented in parallel, with inflated rates set to 1 / 2 / 4 respectively, to capture longer-range patterns:

[0079] Finally, the TCN branch is combined with the SplineGAT main branch and projected to obtain the feature representation for the next round:

[0080] go through After one round of iterations, the final feature representation of all points is obtained. Subsequently, CLS line-level attention aggregation can be used to obtain preliminary line-level token features. Then, the closure markers and global statistics, such as mean curvature and standard deviation of curvature, are concatenated using an MLP. This yields the final line-level features. .

[0081] Subsequently, the calculation of each PoinTr agent point can be performed. To the shortest distance The shortest distance of each feature line Then, these distances are weighted using RBF, as shown in the formula: in is the RBF smoothing coefficient.

[0082] Finally, the eigenvectors of each feature line are... Based on this distance-weighted aggregation, the adjustment features of each proxy point are obtained. Subsequently, feature-level linear modulation is used to achieve the final information exchange between point proxies and line-level tokens. Furthermore, structures such as feature-line-feature-line node heterogeneous graph networks and dual-tower Transformers can be used to achieve closer interaction between line token information and point proxy information.

[0083] Furthermore, it is noted that assigning higher weights to points closer to feature lines in the chamfer distance significantly improves the fidelity of boundary and acute-angle regions, resulting in a more accurate reconstructed model. This is especially true in cases of missing or occluded features, where the feature line structure can be better recovered. Therefore, this invention proposes Edge-weighted chamfer distance (EWCD). By emphasizing feature lines, the model can provide more reliable geometric constraints, avoiding common bulges and collapses. This is particularly beneficial in engineering applications such as CAD, simulation, and design, improving the usability and accuracy of the completion results.

[0084] in The distance between the point and the line. This is a parameter for controlling the width of the line's neighborhood. This is the width control coefficient.

[0085] Therefore, in the actual parameter setting process, this invention needs to statistically analyze the spatial distribution characteristics of points and lines of each model in the data block to determine suitable hyperparameters. This will be explained separately for all models and a single model: This invention calculates the Euclidean distance from each of the 8192 FPS sampled point clouds of each model in data block 0 to the nearest sharp feature line, and records the sequence number of the nearest curve. To ensure comparability of distance scales between different models, the distances are uniformly statistically analyzed in a unit sphere normalized coordinate system to obtain normalized distances. Therefore, the present invention has a distance attenuation kernel:

[0086] in For reference attenuation radius, use Weighted contribution to each point.

[0087] Distance statistics from global sampling show: The mean is approximately 0.132, the median (p50) is approximately 0.064, while p90≈0.342, p95≈0.462, and p99≈0.814, indicating that "most points are not far from the nearest feature line, but there are still obvious long tails and a small number of points that are extremely far from the feature line."

[0088] Regarding the construction of EWCD weights, this data block uses List = {0.02, 0.05, 0.1}, and with As a reference standard; and simultaneously setting the target weight mean. and upper limit .

[0089] "Make" obtained from global sampling estimation q90≈0.1 "Approximately 0.049, compared with..." The basic consistency indicates that the reference It falls within a reasonable range that "can distinguish between near and far lines without being excessively sparse".

[0090] Furthermore, with Smaller, kernel function Sharper, meaning points farther from the feature line decay to near 0 faster, in order to maintain... Global recommendations It will increase accordingly; conversely When it grows bigger It becomes smaller. This phenomenon occurs globally within this data block. The recommended values ​​show a consistent trend and can be used for subsequent EWCD hyperparameter selection and stability analysis.

[0091]

[0092] Table 5: Differences Global kernel distribution and suggestion Taking a single hexagonal nut model as an example, such as Figures 9 to 11 As shown in the feature line hit distribution diagram, the "nearest curve affiliation" of the hexagonal nut exhibits a clear structured partitioning: a large number of points are dominated by a few main curves, such as curves numbered 25 and 21, which have the highest hit count. This usually corresponds to the sharpest edge with the largest geometric coverage, in this case, the sharp edge of the inner hole annulus. Meanwhile, this model hit a total of 32 different curves, indicating that the set of sharp curves itself is not scarce; however, the hit distribution shows a concentrated trend, namely the typical case of "most curves participating, a few curves dominating".

[0093] The distance heatmap reveals that the nut body is mostly planar or cylindrical, with many points not close to sharp edges. Therefore, the distance to the nearest sharp curve forms a ring-shaped / regional gradient. Statistically, this model... The mean is 0.110, the median is 0.104, p90 = 0.200, p99 = 0.274, and the maximum value is 0.278. This means that in unit spherical coordinates, a large number of points are more than 0.1 away from the sharp edge, while there are still a number of points near the sharp edge that are close to 0, corresponding to the sampling points near the actual edge / ring edge.

[0094] Meanwhile, in this example The histogram shows a pattern of "large accumulations near 0 and a few peaks near 1": when a point is more than approximately [missing information] away from the sharp edge... hour, It will decay rapidly to near 0, so large surfaces far from the sharp edge essentially return to "approximately normal CD"; while points close to the sharp edge... This will be significantly amplified. Statistically, this model... The mean is 0.228. If we use global fitting... (To make the average weight at the block level close to the target value) the calculation shows that the average weight of this model is:

[0095] It can be seen that the overall increase in point cloud weight is relatively mild, but the maximum weight increase for points near sharp edges can reach [amount missing]. This aligns perfectly with the design intent of preserving the shape of most points while providing extra emphasis to sharp-edge neighborhoods.

[0096] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A feature-driven point cloud completion method for CAD parametric models, characterized in that, Includes the following steps: Step S1: Construct a multi-source feature line acquisition and standardization system to obtain feature line data from the CAD model. The sources of the feature line data include CAD information parsing and extraction, automatic detection based on grids or point clouds, and interactive editing and rule generation. Step S2 involves performing consistency evaluation, cleaning, repair, and standardization on the feature line data obtained in step S1, and outputting a feature representation in a unified format, which includes a fixed-length point sequence, topological relationships, and metadata. Step S3: Based on the local feature extraction method of multi-relationship adaptive weight, local features are extracted from the incomplete point cloud. Under the DGCNN framework, the contribution weight of each neighboring point to the local features of the target point is dynamically adjusted according to the geometric relationship measurement between the target point and its neighboring points in the three-dimensional space and the similarity relationship measurement in the high-dimensional feature space. Step S4: The feature line information obtained in step S2 is fused into the point cloud completion deep learning model through the feature line deep information injection technology. The feature lines are encoded by the SplineLineEncoder to obtain line-level token features, and the line-level token features are interacted with the point cloud proxy point features. Step S5: The point cloud completion model is trained and optimized using a feature line-weighted chamfer distance loss function, and points closer to the feature lines are given higher weights in the loss calculation.

2. The feature-driven point cloud completion method for CAD parametric models according to claim 1, characterized in that, In step S1, the multi-source feature line acquisition and standardization system supports manual interactive generation of quadrilateral grid layout. Users specify key feature points or feature lines on the triangular grid, and the algorithm automatically generates a geodesic network to form a coherent quadrilateral layout based on the input. In the geodesic calculation process, a local optimization strategy based on edge flipping is introduced. By continuously flipping the edges of the triangular grid, the path length is minimized so that the path approximates the geodesic.

3. The feature-driven point cloud completion method for CAD parametric models according to claim 2, characterized in that, Step S1 also includes a symmetry-sensitive mechanism, where users manually mark symmetry points on the model to guide the generation of the model's symmetry plane or symmetry axis. When the user marks or draws geodesics on one side, the symmetry module automatically copies the corresponding operation on the other side to maintain symmetry consistency.

4. The feature-driven point cloud completion method for CAD parametric models according to claim 1, characterized in that, Step S1 also includes automated quadrilateral mesh layout generation. After the user provides the curvature alignment information of some facets of the model, a smooth cross field and corresponding singularity are generated according to the alignment constraints, and a seamless surface parameterization is generated. Under the guidance of the seamless parameterization result, the virtual particles start from the singularity and travel along the boundary line direction in the cross field, leaving a trajectory on the model to form a motorcycle diagram. The motorcycle diagram divides the mesh model into several quadrilateral mesh regions to form a feature line frame.

5. The feature-driven point cloud completion method for CAD parametric models according to claim 1, characterized in that, The consistency assessment, cleaning, repair, and standardization of the feature line data in step S2 specifically includes: establishing a local coordinate system and error metric for each curve using its analytical definition; calculating the residual distribution on the grid vertices corresponding to the curve; determining the reliability using robust statistical methods; and classifying the curves into three categories: OK, SUSPECT, and BAD. For OK curves with concentrated error distribution, the error distribution is directly resampled according to the analytical curve to obtain a clean sequence of points with equal arc lengths. For SUSPECT curves, which are generally consistent but have a small number of outliers, the outliers are removed or projected back onto the curve without destroying the topology. For BAD curves, if the overall configuration is mismatched, the analytical parameters are abandoned and approximated by a polyline on the grid or by local fitting.

6. The feature-driven point cloud completion method for CAD parametric models according to claim 1, characterized in that, In step S3, the geometric relationship metric is calculated by the distance between the target point and its neighboring points in three-dimensional space, and the high-dimensional feature relationship metric is calculated by the feature similarity between the neighboring points and the target point. The contribution weight is dynamically determined based on the weighted combination of the geometric relationship metric and the high-dimensional feature relationship metric. The geometric relationship metric reflects the spatial positional differences between points, and the high-dimensional feature relationship metric reflects the feature representation differences between points.

7. The feature-driven point cloud completion method for CAD parametric models according to claim 1, characterized in that, The workflow of the SplineLineEncoder in step S4 includes: constructing a local frame for each point of the feature line that has been parameterized by arc length; projecting the relative displacement onto the local coordinate system; building a point-level initial embedding containing the three-dimensional coordinates, curvature, tangency, parameterized values, and closure information of each feature line point; and edge-level geometric embedding for modeling the relationship between adjacent points, including arc length, rate of change of curvature, rotation angle, and frame information; using SplineGATBlock to aggregate information on the expanded adjacency of each point using geometric attention; implementing the dilated TCN branch in parallel to capture longer-range patterns; and finally, combining the TCN branch with the SplineGAT main branch to project and obtain the feature representation for the next round. After multiple iterations, the final feature representation of all points is obtained. Using CLS line-level attention aggregation, preliminary line-level token features are obtained. Then, the closure markers and global statistics are infused through MLP to obtain the final line-level features.

8. The feature-driven point cloud completion method for CAD parametric models according to claim 1, characterized in that, The information interaction between the line-level token features and the point cloud proxy point features in step S4 specifically includes: calculating the shortest distance from each PoinTr proxy point to multiple feature lines with the shortest distance, then using RBF to weight these distances, aggregating the feature vectors of each feature line according to the weighted distance to obtain the adjustment features of each proxy point, and then using feature-level linear modulation to perform the final information interaction between the point proxy and the line-level token.

9. The feature-driven point cloud completion method for CAD parametric models according to claim 1, characterized in that, In step S5, the feature line weighted chamfer distance loss function assigns a weight to each point in the point cloud based on its distance from the nearest feature line. Points closer to the feature line have a higher weight, while points farther from the feature line have a lower weight. The contribution ratio of the feature line region in the loss calculation is adjusted by the weight coefficient.

10. A feature-driven point cloud completion method for CAD parametric models, characterized in that, include: The multi-source feature line acquisition and standardization module is used to acquire feature line data from CAD models and perform consistency assessment, cleaning, repair and standardization processing; The local feature extraction module is used to extract local features from incomplete point clouds based on multi-relation adaptive weights. The feature line information injection module is used to fuse feature line geometric information into the point cloud completion deep learning model; The loss optimization module is used to train and optimize the point cloud completion model using a feature line-weighted chamfer distance loss function.

Images