Point cloud completion method based on DFG-PCN model
Through the adaptive graph connectivity allocation and cross-scale feature fusion mechanism, the problems of unbalanced resource allocation and insufficient feature fusion in the point cloud completion method are solved, high-precision point cloud completion is achieved, and the reconstruction quality of complex areas and the coherence of the overall shape are improved.
Patent Information
- Application Number
- CN202510799721.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-16
- Publication Date
- 2025-09-23
AI Technical Summary
Existing point cloud completion methods fail to effectively distinguish reconstruction requirements when processing different regions, resulting in unbalanced resource allocation and inability to simultaneously maintain the integrity of local details and global structures. In addition, feature fusion methods fail to fully consider the differences in regional importance, resulting in insufficient reconstruction accuracy and geometric fidelity.
Adaptive graph connectivity allocation, local-global dual-graph collaborative modeling and cross-scale feature fusion mechanism are adopted. Graph connectivity is dynamically allocated through detail richness and curvature information. The feature fusion process is optimized by combining the Manhattan distance weighted edge aggregation strategy and cross-scale attention mechanism.
It significantly improves the reconstruction accuracy of complex geometric structures and the consistency of the overall shape, enhances the geometric reconstruction quality of key areas, solves the problems of unbalanced resource allocation and insufficient feature fusion in traditional methods, and achieves high-precision point cloud completion.
Smart Images
Figure CN120689250A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of point cloud completion, and more specifically to a point cloud completion method based on a DFG-PCN model. Background Art
[0002] Point cloud completion is a core task in 3D vision and computer graphics. It aims to reconstruct a complete and dense 3D object model from incomplete or sparse input point cloud data. This technology is valuable in applications such as autonomous driving scene understanding, robot navigation, augmented reality, and digital preservation of cultural relics.
[0003] Currently, mainstream deep learning-based point cloud completion methods generally rely on graph neural networks or attention mechanisms to model point relationships. However, these methods typically use fixed neighbor connection strategies or preset attention windows. This "equal degree" or "equal connectivity" approach fails to effectively distinguish the significant differences in reconstruction requirements across different regions of the point cloud. Regions with simple geometric structures and flat features are allocated the same computing resources and connection weights as regions with rich details, complex structures, or discontinuities. This directly results in insufficient resources for the model to reconstruct key details, while redundant computation occurs in flat areas. Ultimately, the reconstruction of complex regions is blurred, distorted, or structurally broken, severely limiting completion accuracy.
[0004] Furthermore, existing methods have significant limitations in terms of feature extraction and aggregation. On the one hand, methods that rely too much on local neighborhoods can capture fine-grained features, but have difficulty establishing effective long-range dependencies, which can easily lead to loose overall structures, disproportionate proportions, or unreasonable connections between components of the reconstructed object. On the other hand, methods that emphasize global features can ensure overall shape consistency, but inevitably lose a large amount of local geometric details, resulting in overly smooth completed surfaces and blunted sharp features. Existing technologies lack a mechanism to effectively coordinate local fine features with global structural information, and are unable to accurately restore local complex geometry while maintaining overall coherence.
[0005] Furthermore, when fusing features from different layers or graph structures, existing technologies often employ attention fusion methods that involve simple concatenation, addition, or uniform weighting. This "one-size-fits-all" strategy fails to fully consider the varying importance of different spatial regions of the point cloud in contributing to the final reconstruction. In particular, geometric information from highly detailed or structurally critical regions is not adequately addressed and effectively preserved during the fusion process, being treated the same as features from flat areas. This inefficient fusion approach weakens the model's ability to utilize critical information, resulting in decreased accuracy and insufficient geometric fidelity in the completion results in key areas.
[0006] Therefore, how to design a point cloud completion method based on the DFG-PCN model that can effectively coordinate local details and global structural information, and prioritize the integration of geometric clues in high-value areas in feature fusion, and improve the reconstruction accuracy of complex geometric structures and the coherence of the overall shape is a problem that technical personnel in this field urgently need to solve. Summary of the Invention
[0007] In view of this, the present invention provides a point cloud completion method based on the DFG-PCN model to solve the problems of loss of geometric details and insufficient structural coherence caused by fixed connectivity and single-scale feature fusion in traditional point cloud completion methods. Through adaptive graph connectivity allocation, local-global dual-graph collaborative modeling and cross-scale feature fusion mechanism, a high-precision and structurally complete completed point cloud is generated.
[0008] In order to achieve the above object, the present invention adopts the following technical solutions:
[0009] A point cloud completion method based on the DFG-PCN model includes the following steps:
[0010] S1, perform hierarchical feature extraction on the input point cloud P through the feature extraction module to generate the shape feature vector f and the downsampled point cloud P p and point eigenvector F p ;
[0011] S2, fuse the shape feature vector f and the downsampled point cloud P through the seed generation module p and point eigenvector F p , generate low-resolution seed point cloud P0;
[0012] S3, input the seed point cloud P0 into at least one variable degree point map module to generate a step-by-step improved high-resolution point cloud P i ;
[0013] S4. Apply Chamfer distance supervision loss on the multi-stage generated point cloud to gradually optimize the point cloud completion results.
[0014] Preferably, the S1 includes:
[0015] S11, using three-layer SetAbstraction units to perform local to global feature aggregation on the input point cloud P;
[0016] S12, combine the PointTransformer unit to enhance local shape context perception, generate the shape feature vector f through maximum pooling, and output the downsampled point cloud P p , point feature vector F p .
[0017] Preferably, the S2 includes:
[0018] S21, shape feature vector f, downsample point cloud P p and point eigenvector F p Input UpsampleTransformer unit to generate point-level features;
[0019] S22, point-level features and shape encoding are fused and processed by MLP unit to output a rough point cloud P c ;
[0020] S23, for the rough point cloud P c It is stitched with the input point cloud P and downsampled using the FPS algorithm to obtain the seed point cloud P0.
[0021] Preferably, the S3 includes:
[0022] S31, receiving the i-1th level point cloud P i-1 And the transfer characteristic H i-1 , extract point features Q through PointNet unit i-1 ;
[0023] S32, transfer feature H i-1 , point feature Q i-1 Input graph construction layer, based on detail richness metric D i and curvature K i Adaptively assign connection degrees and construct a dual graph structure G consisting of a local graph and a global graph i ;
[0024] S33, transfer feature H i-1 , point feature Q i-1 , dual graph structure G i Input graph aggregation layer, use Manhattan distance weighted edge condition aggregation strategy to update node features, and output aggregated features;
[0025] S34, the aggregated features are input into the graph fusion layer, and the detail richness metric D i Under the guidance of the cross-scale attention mechanism, the local graph and global graph features are fused to generate enhanced features H i ;
[0026] S35. Copy point cloud P i-1 Get the intermediate point set P ' i-1 , will enhance the feature H i Input MLP to generate high-dimensional features, and output the displacement vector ΔP after dimensionality increase through the deconvolution layer i , obtain high-resolution point cloud P i =P ' i-1 +ΔP i .
[0027] Preferably, in S32, the rule for adaptively allocating connection degrees is expressed as:
[0028]
[0029]
[0030] Where B represents the total degree budget, round represents the standard rounding function, α represents the scaling hyperparameter that controls the overall connection budget, n represents the number of points in the point cloud in the region, and D i represents the detail richness measure, K i Represents curvature.
[0031] Preferably, in said S32, the local graph is constructed by a KNN algorithm, with a fixed number of neighbors k=16;
[0032] The global graph samples 512 key points through FPS and connects them with the full graph nodes to capture long-range dependencies.
[0033] Preferably, in S33, the edge condition aggregation strategy is expressed as:
[0034]
[0035] in, represents the Manhattan distance between node j and node k in stage i, γ represents a hyperparameter, represents the three-dimensional coordinates of node j, represents the three-dimensional coordinates of node k, β i represents the linear projection function, Represent the features of the current point and the neighboring point in the i-th stage, δ represents the position encoding vector, represents a parameterized function, C represents a negative constant, Represents the graph G i The logical NOT operation, represents the attention weight, represents the normalized attention weight, N(j) represents the neighbor set of node j, represents the characteristics of neighbor node k in the i-1 stage.
[0036] Preferably, in S34, the cross-scale attention mechanism is expressed as:
[0037]
[0038] Where Q = H l W Q 、K=H l W K 、V=H l WV They represent the embedded representations of query, key, and value after projection, respectively, and W Q , WK, W V learnable projection matrices representing query, key, and value projections, respectively, represents the dimension of the key vector, H l Represents local graph features, Represents the global graph features after linear transformation.
[0039] Preferably, the local graph feature H l Expressed as:
[0040]
[0041] in,
[0042] represents the initial local graph features, Indicates detail richness.
[0043] Preferably, the S4 includes:
[0044] S41. Set multi-stage supervision objectives: Generate point cloud sequence {P0,P1,…,P n} and the corresponding real point cloud S0,S1,…,S n};
[0045] S42. Calculate the Chamfer distance at each stage:
[0046]
[0047] Among them, P i 、S i They represent the point cloud generated in the i-th stage and the corresponding real point cloud, p and s represent P respectively. i 、S i A single point in
[0048] S43. Optimize total loss Until convergence.
[0049] It can be seen from the above technical solution that compared with the prior art, the technical solution of the present invention has the following advantages:
[0050] Beneficial effects:
[0051] This method dynamically allocates graph connectivity degrees by combining detail richness metrics with curvature information, allowing the model to allocate more computational resources to areas with rich detail or structural discontinuities. Compared to traditional fixed-degree KNN methods, this strategy significantly improves geometric reconstruction quality in key areas and alleviates the imbalanced reconstruction demand in point cloud completion.
[0052] 2. A dual-graph construction mechanism, combining local and global graphs, captures long-range dependencies while preserving local fine-grained features. Combined with a Manhattan distance-weighted edge aggregation strategy, it enhances the perception of geometrically discontinuous areas, thereby improving the structural consistency and detail integrity of the overall shape.
[0053] 3. Guided by the richness of details, local and global graph features are adaptively fused through a cross-scale attention mechanism. This breaks through the limitations of traditional unified fusion, prioritizing the integration of regional information that contributes most to reconstruction, significantly enhancing the information density and effectiveness of feature representation, and ultimately achieving multi-stage generation and optimization from coarse to fine. BRIEF DESCRIPTION OF THE DRAWINGS
[0054] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.
[0055] Figure 1 A schematic diagram of the DFG-PCN model structure provided in an embodiment of the present invention;
[0056] Figure 2 A schematic diagram of the structure of a variability point diagram module provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0057] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0058] like Figure 1 As shown, this embodiment provides a point cloud completion method based on the DFG-PCN model, comprising the following steps:
[0059] S1, perform hierarchical feature extraction on the input point cloud P through the feature extraction module to generate the shape feature vector f and the downsampled point cloud P p and point eigenvector F p ;
[0060] S2, fuse the shape feature vector f and the downsampled point cloud P through the seed generation module p and point eigenvector F p , generate low-resolution seed point cloud P0;
[0061] S3, input the seed point cloud P0 into at least one variable degree point map module to generate a step-by-step improved high-resolution point cloud P i ;
[0062] S4. Apply Chamfer distance supervision loss on the multi-stage generated point cloud to gradually optimize the point cloud completion results.
[0063] This method adaptively focuses on key areas through a dynamic degree allocation strategy guided by detail richness and curvature; enhances the ability to express geometric discontinuities by combining a dual-image collaborative mechanism that combines local fine perception with global structure modeling; and optimizes feature fusion efficiency under the guidance of cross-scale attention, significantly improving the comprehensive performance of point cloud completion in detail reconstruction accuracy, structural coherence, and complex shape generalization.
[0064] The following further describes each step in the above method in detail;
[0065] In this embodiment S1, the feature extraction module performs hierarchical feature extraction on the input point cloud P to generate the shape feature vector f and the downsampled point cloud P. p and point eigenvector F p ; Specifically include:
[0066] S11. Use a three-layer SetAbstraction unit to perform local-to-global feature aggregation on the input point cloud P. The SetAbstraction unit uses the three-layer operation of farthest point sampling (FPS), grouping, and point network (PointNet) to gradually downsample and aggregate local features, achieving feature extraction from local to global while preserving key geometric structures.
[0067] S12, combine the PointTransformer unit to enhance local shape context perception, generate the shape feature vector f through maximum pooling, and output the downsampled point cloud P p , point feature vector F p ; It introduces the self-attention mechanism to enhance the context perception ability of local areas of the point cloud and improve the feature expression ability.
[0068] This step not only achieves feature aggregation from local to global through the combination of three-layer SetAbstraction units and PointTransformer units, but also enhances the context perception capability of local areas of the point cloud.
[0069] In this embodiment S2, the shape feature vector f and the down-sampled point cloud P are fused through the seed generation module. p and point eigenvector F p , generate low-resolution seed point cloud P0; specifically including:
[0070] S21, shape feature vector f, downsample point cloud P p and point eigenvector F p The input is an UpsampleTransformer unit that generates point-level features. The UpsampleTransformer unit explicitly models the geometric relationship between the visible part and the missing shape through point-level feature segmentation, providing a strong shape prior for subsequent refinement.
[0071] S22, point-level features and shape encoding are fused and processed by MLP unit to output a rough point cloud P c ;
[0072] S23, for the rough point cloud P c It is stitched with the input point cloud P and downsampled using the FPS algorithm to obtain the seed point cloud P0; this ensures that the seed points evenly cover the observed and missing areas, avoids completion deviation, and improves the spatial rationality of the initial seed points.
[0073] After feeding the shape feature vector, downsampled point cloud, and point feature vector into the UpsampleTransformer unit, the seed generation module generates a low-resolution but structurally complete seed point cloud by explicitly modeling the geometric relationship between the visible and missing shapes. This step ensures that the initial seed points evenly cover both observed and missing areas, effectively avoiding completion bias and paving the way for subsequent refinement.
[0074] In this embodiment S3, the seed point cloud P0 is input into at least one variable degree point map module to generate a step-by-step improved high-resolution point cloud P i ;like Figure 2 As shown, specifically including:
[0075] S31, receiving the i-1th level point cloud P i-1 And the transfer characteristic H i-1 , extract point features Q through PointNet unit i-1 ;
[0076] S32, transfer feature H i-1 , point feature Q i-1 Input graph construction layer, based on detail richness metric D i and curvature K i Adaptively assign connection degrees and construct a dual graph structure G consisting of a local graph and a global graph i ;
[0077] S33, transfer feature H i-1 , point feature Q i-1 , dual graph structure G iInput graph aggregation layer, use Manhattan distance weighted edge condition aggregation strategy to update node features, and output aggregated features;
[0078] S34, the aggregated features are input into the graph fusion layer, and the detail richness metric D i Under the guidance of the cross-scale attention mechanism, the local graph and global graph features are fused to generate enhanced features H i ;
[0079] S35. Copy point cloud P i-1 Get the intermediate point set P ' i-1 , will enhance the feature H i Input MLP to generate high-dimensional features, and output the displacement vector ΔP after dimensionality increase through the deconvolution layer i , obtain high-resolution point cloud P i =P ' i-1 +ΔP i .
[0080] Furthermore, the rule for adaptively allocating connection degrees in S32 is expressed as:
[0081]
[0082] Where B represents the total degree budget, round represents the standard rounding function, α represents the scaling hyperparameter that controls the overall connection budget, n represents the number of points in the point cloud in the region, and D i represents the detail richness measure, K i Represents curvature.
[0083] Furthermore, the local graph in S32 is constructed using the KNN algorithm with a fixed number of neighbors k = 16. The global graph samples 512 key points through FPS and connects them with the nodes of the full graph to capture long-range dependencies. The local graph captures fine-grained geometry, while the global graph models long-range structural dependencies (such as symmetry). This balances local details with global consistency, solving the expression bottleneck of traditional single graph structures.
[0084] Furthermore, the edge condition aggregation strategy in S33 is expressed as:
[0085]
[0086] in, represents the Manhattan distance between node j and node k in stage i, γ represents a hyperparameter, represents the three-dimensional coordinates of node j, represents the three-dimensional coordinates of node k, β i represents the linear projection function, Represent the features of the current point and the neighboring point in the i-th stage, δ represents the position encoding vector, represents a parameterized function, C represents a negative constant, Represents the graph G i The logical NOT operation, represents the attention weight, represents the normalized attention weight, N(j) represents the neighbor set of node j, represents the characteristics of neighbor node k in the i-1 stage.
[0087] It imposes negative weight penalties on non-connected edges, suppresses invalid neighbor interference, improves the signal-to-noise ratio of aggregated features, and retains key geometric clues.
[0088] Furthermore, the cross-scale attention mechanism in S34 is expressed as:
[0089]
[0090] Where Q = H l W Q 、K=H l W K 、V=H l W V They represent the embedded representations of query, key, and value after projection, respectively, and W Q , WK, W V learnable projection matrices representing query, key, and value projections, respectively, represents the dimension of the key vector, H l Represents local graph features, Represents the global graph features after linear transformation.
[0091] Furthermore, the local graph feature H l Expressed as:
[0092]
[0093] in,
[0094] represents the initial local graph features, Indicates the richness of details; the above-mentioned image fusion layer adaptively enhances the feature response of high-detail areas and avoids excessive smoothing of flat areas.
[0095] The variable-degree point graph module employs an adaptive connection mechanism driven by detail richness and curvature, enabling collaborative modeling of local and global graphs. Combining a Manhattan distance-weighted edge aggregation strategy with a cross-scale attention mechanism, it effectively improves the geometric reconstruction accuracy of key regions. This not only enhances the model's perception of complex structures and discontinuous regions (such as edges and sharp corners), but also significantly increases the information density and utilization of feature representations, thereby achieving high-quality reconstruction results from coarse to fine, balancing local detail and global consistency in point cloud completion tasks.
[0096] In this embodiment S4, Chamfer distance supervision loss is applied to the multi-stage generated point cloud to gradually optimize the point cloud completion result; specifically, the following steps are included:
[0097] S41. Set multi-stage supervision objectives: Generate point cloud sequence {P0,P1,…,P n} and the corresponding real point cloud S0,S1,…,S n};
[0098] S42. Calculate the Chamfer distance at each stage:
[0099]
[0100] Among them, P i 、S i They represent the point cloud generated in the i-th stage and the corresponding real point cloud, p and s represent P respectively. i 、S i A single point in
[0101] S43. Optimize total loss Until convergence.
[0102] In this step, a Chamfer distance supervision loss is applied to ensure that the point cloud generated at each stage is as close as possible to the true target point cloud, thereby achieving gradual optimization. This multi-level supervision mechanism helps the network gradually optimize the distribution of points at different resolutions, ultimately achieving high-quality point cloud completion results.
[0103] The following further illustrates the point cloud completion method based on the DFG-PCN model in this embodiment, using a specific industrial inspection scenario. In aircraft engine blade damage inspection, due to the complexity of the blade surface and the difficulty in fully scanning the damaged area, the acquired point cloud data may be missing. The specific implementation steps include:
[0104] 1) Residual cloud input and feature modeling;
[0105] For missing point clouds caused by occlusion or surface reflection during blade scanning (such as root grooves and twisted areas at the blade tip), the residual point cloud is input into the feature extraction module. The three-layer SetAbstraction unit performs local geometric aggregation on the blade aerodynamic surface (preserving the leading / trailing edge curvature features), and the PointTransformer is combined to enhance the contextual perception of high curvature areas, outputting global shape features and downsampled point clouds. The seed generation module explicitly models the geometric continuity between the defect area and the complete blade, generating a low-resolution seed point cloud P0 covering the missing area, providing a structurally reasonable initial topology for reconstruction of the damaged area.
[0106] 2) damage-sensitive adaptive refinement;
[0107] The seed point cloud P0 is input into the cascade variable degree point map module to optimize the parameter configuration according to the requirements of blade damage detection, including:
[0108] Graph construction: The degree of connection is dynamically allocated based on detail richness (gradient change of defect edges) and curvature (continuity of aerodynamic surfaces). The local graph connection density is increased in defect areas such as cracks and pits. The global graph captures the overall bending and torsional deformation characteristics of the blade through 512 FPS key points.
[0109] Feature fusion: A Manhattan distance weighted aggregation strategy (γ = 0.8) is used to suppress scanning noise interference, and a cross-scale attention mechanism prioritizes the fusion of high-frequency features of defect boundaries (such as the microscale geometry of fatigue cracks) to ensure that sharp edges are retained when reconstructing the damaged area.
[0110] Displacement generation: Through the deconvolution layer, the displacement vector ΔPi that conforms to the aerodynamic profile of the blade is generated in the defect area to avoid damage morphological distortion caused by over-smoothing.
[0111] 3) Complete the results and verify the test;
[0112] Output high-resolution point cloud P i Afterwards, a Chamfer distance comparison was performed with the complete blade CAD model (threshold ≤ 0.3 mm).
[0113] The completed point cloud can significantly improve detection reliability. It can fill in the scanning gaps in key areas such as the blade root bolt holes and avoid missed detections due to incomplete data. Through the dual-image collaborative mechanism, it can adapt to the complex curved surfaces of blades of different models (such as wide-chord blades and hollow blades), ensuring that a physically reasonable completion structure can still be generated when the defect shape is unknown.
[0114] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. References to the same or similar parts between the various embodiments are sufficient. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For relevant parts, refer to the method description.
[0115] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A point cloud completion method based on the DFG-PCN model, characterized in that: The following steps are involved: S1, perform hierarchical feature extraction on the input point cloud P through the feature extraction module to generate the shape feature vector f and the downsampled point cloud P p and point eigenvector F p ; S2, fuse the shape feature vector f and the downsampled point cloud P through the seed generation module p and point eigenvector F p , generate low-resolution seed point cloud P0; S3, input the seed point cloud P0 into at least one variable degree point map module to generate a step-by-step improved high-resolution point cloud P i ; S4. Apply Chamfer distance supervision loss on the multi-stage generated point cloud to gradually optimize the point cloud completion results.
2. A point cloud completion method based on the DFG-PCN model according to claim 1, characterized in that: Said S1 comprises: S11, using three-layer SetAbstraction units to perform local to global feature aggregation on the input point cloud P; S12, combine the PointTransformer unit to enhance local shape context perception, generate the shape feature vector f through maximum pooling, and output the downsampled point cloud P p , point feature vector F p .
3. The point cloud completion method based on the DFG-PCN model according to claim 1, characterized in that: The S2 includes: S21, shape feature vector f, downsample point cloud P p and point eigenvector F p Input UpsampleTransformer unit to generate point-level features; S22, point-level features and shape encoding are fused and processed by MLP unit to output a rough point cloud P c ; S23, for the rough point cloud P c It is stitched with the input point cloud P and downsampled using the FPS algorithm to obtain the seed point cloud P0.
4. The point cloud completion method based on the DFG-PCN model according to claim 1, characterized in that: The S3 includes: S31, receiving the i-1th level point cloud P i-1 And the transfer characteristic H i-1 , extract point features Q through PointNet unit i-1 ; S32, transfer feature H i-1 , point feature Q i-1 Input graph construction layer, based on detail richness metric D i and curvature K i Adaptively assign connection degrees and construct a dual graph structure G consisting of a local graph and a global graph i ; S33, transfer feature H i-1 , point feature Q i-1 , dual graph structure G i Input graph aggregation layer, use Manhattan distance weighted edge condition aggregation strategy to update node features, and output aggregated features; S34, the aggregated features are input into the graph fusion layer, and the detail richness metric D i Under the guidance of the cross-scale attention mechanism, the local graph and global graph features are fused to generate enhanced features H i ; S35. Copy point cloud P i-1 Get the intermediate point set P ' i-1 , will enhance the feature H i Input MLP to generate high-dimensional features, and output the displacement vector ΔP after dimensionality increase through the deconvolution layer i , obtain high-resolution point cloud P i =P ' i-1 +ΔP i .
5. The point cloud completion method based on the DFG-PCN model according to claim 4, characterized in that: In S32, the rule for adaptively allocating connection degrees is expressed as: Where B represents the total degree budget, round represents the standard rounding function, α represents the scaling hyperparameter that controls the overall connection budget, n represents the number of points in the point cloud in the region, and D i represents the detail richness measure, K i Represents curvature.
6. The point cloud completion method based on the DFG-PCN model according to claim 4, characterized in that: In S32, the local graph is constructed by the KNN algorithm, and the number of neighbors is fixed at k=16; The global graph samples 512 key points through FPS and connects them with the full graph nodes to capture long-range dependencies.
7. The point cloud completion method based on the DFG-PCN model according to claim 4, characterized in that: In S33, the edge condition aggregation strategy is expressed as: in, represents the Manhattan distance between node j and node k in stage i, γ represents a hyperparameter, represents the three-dimensional coordinates of node j, represents the three-dimensional coordinates of node k, β i represents the linear projection function, Represent the features of the current point and the neighboring point in the i-th stage, δ represents the position encoding vector, represents a parameterized function, C represents a negative constant, Represents the graph G i The logical NOT operation, represents the attention weight, represents the normalized attention weight, N(j) represents the neighbor set of node j, represents the characteristics of neighbor node k in the i-1 stage.
8. The point cloud completion method based on the DFG-PCN model according to claim 4, characterized in that: In S34, the cross-scale attention mechanism is expressed as: Where Q = H l W Q 、K=H l W K 、V=H l W V They represent the embedded representations of query, key, and value after projection, respectively, and W Q , WK, W V learnable projection matrices representing query, key, and value projections, respectively, represents the dimension of the key vector, H l Represents local graph features, Represents the global graph features after linear transformation.
9. The point cloud completion method based on the DFG-PCN model according to claim 8, characterized in that: The local graph feature H l Expressed as: in, represents the initial local graph features, Indicates detail richness.
10. The point cloud completion method based on the DFG-PCN model according to claim 1, characterized in that: The S4 includes: S41. Set multi-stage supervision objectives: Generate point cloud sequence {P0,P1,…,P n } and the corresponding real point cloud S0,S1,…,S n }; S42. Calculate the Chamfer distance at each stage: Among them, P i 、S i They represent the point cloud generated in the i-th stage and the corresponding real point cloud, p and s represent P respectively. i 、S i A single point in S43. Optimize total loss Until convergence.