High-fidelity point cloud completion method and system based on double-path attention and fractal structure

Abstract

The invention relates to a high-fidelity point cloud completion method and system based on a double-path attention and fractal structure. The method comprises the following steps: processing an input point cloud by an iterative farthest point sampling algorithm; extracting and constructing a joint feature vector by combining an extended combined multi-layer perceptron with a double-path attention mechanism; reconstructing missing region point clouds in a layered manner by using a pyramid point generator; designing a composite loss function and a dynamic weight scheduling strategy to optimize a reconstruction target; designing an adversarial training framework of a discriminator based on a built-in spectrum normalization and gradient penalty mechanism; and reasoning to generate a point cloud reconstruction result. According to the method, the problem that a traditional point cloud completion method depends on geometric prior and a matching template is solved, the limitation that calibration on semantic information of different dimensions is lacked, excessive dependence on maximum pooling is achieved, and the uniformity of discriminator gradient anomaly and a loss function is difficult to guarantee is relieved; and the geometric detail recovery capability and the visual authenticity in a complex structure scene are obviously improved.

Description

Technical Field

[0001] The present invention relates to the field of computer vision and three-dimensional reconstruction technology, and in particular to a high-fidelity point cloud completion method and system based on dual-path attention and fractal structure. Background Art

[0002] In the field of three-dimensional visual perception, point cloud data, as a discretized carrier of geometric information about the surface of an object, has become a core input for applications such as industrial quality inspection and autonomous driving environment modeling. However, during the actual acquisition process, the acquired point clouds are often affected by sensor field of view obstructions, motion blur of dynamic targets, and complex surface reflection characteristics. Local structure loss and topological fractures are common in these data defects. These data defects seriously restrict the reliability of downstream tasks: for example, in the inspection of aircraft engine blades, incomplete cooling hole point clouds can lead to distortion of airflow simulation; in autonomous driving scenarios, incomplete obstacle outlines can lead to collision prediction errors.

[0003] Traditional point cloud completion methods primarily rely on geometric priors or template matching mechanisms. Interpolation methods based on geometric priors assume that surfaces are continuous and differentiable. However, when dealing with non-smooth geometries such as thin-walled structures and sharp edges, they often result in overfilling or structural collapse, such as the curvature distortion seen when completing the leading edge of an engine blade. Registration algorithms based on template matching require the retrieval of similar models for rigid alignment, which is computationally inefficient and difficult to accurately match when local features are missing (such as missing teeth on a gear), leading to semantically contradictory reconstruction results.

[0004] In recent years, deep learning point cloud completion methods have made significant progress. Methods such as global feature generation networks (PCN) achieve holistic reconstruction through end-to-end encoding and decoding. However, they ignore local structural correlations, making them prone to incorrectly filling in hollowed-out areas of seatbacks as solid blocks. Local region prediction methods (such as PF-Net) preserve known structures and predict missing parts, but their multi-scale feature fusion mechanism lacks adaptive weighting capabilities, resulting in grid discontinuities when reconstructing the mesh-like lampshade of a lamp. Iterative optimization methods (such as PMPNet++) improve detail continuity through multi-step displacement refinement, but coordinate prediction errors in the early stages are amplified during the hierarchical propagation, resulting in hub center drift when completing bicycle spokes.

[0005] The current technical bottlenecks are concentrated in three aspects: in the process of multi-scale feature fusion, low-dimensional geometric details and high-dimensional semantic features suffer from key information attenuation due to the lack of a dynamic calibration mechanism, a typical example being the excessive smoothing of the gradient changes at the hole boundary; local geometric aggregation operations rely too much on fixed strategies such as maximum pooling, resulting in insufficient representation capabilities of high-variance structures (such as thread gaps), and frequent noise or collapse in the completed surface; the gradient anomaly of the discriminator in the adversarial training framework can easily lead to mode collapse, and the generated point cloud often degenerates into a single topological form (such as uniformly completing various table legs into cylinders), and the existing loss function is difficult to constrain the uniformity of point set distribution at the micro level.

[0006] While SA-Net optimizes feature selection through cross-layer attention, and SnowflakeNet proposes fractal decoding to enhance detail generation, these approaches still have inherent limitations: the former suffers from insufficient spatial attention response to sparse, missing regions, while the latter still produces local ambiguity in complex topological scenarios. Therefore, a new completion architecture that integrates adaptive feature weighting, hierarchical geometric reconstruction, and a stable training mechanism is urgently needed to meet the dual requirements of structural fidelity and detail authenticity for high-precision 3D perception applications. Summary of the Invention

[0007] In response to the shortcomings of the existing technology, the present invention provides a high-fidelity point cloud completion method and system based on dual-path attention and fractal structure, which solves the problems of traditional point cloud completion methods such as excessive reliance on geometric priors and matching templates, lack of effective calibration of semantic information of different dimensions, excessive reliance on maximum pooling, abnormal discriminator gradients, and inability of loss functions to ensure point uniformity.

[0008] To solve the above technical problems, the present invention provides the following technical solution: a high-fidelity point cloud completion method based on dual-path attention and fractal structure, comprising the following steps: S1. Construct a dataset containing residual point clouds. The residual point clouds of the input model are processed by an iterative farthest point sampling algorithm to generate three-layer point cloud subsets including original resolution, medium resolution, and low resolution, respectively retaining complete geometric details, main structural features, and global topological skeleton; S2. For each of the three point cloud subsets, an extended combined multi-layer perceptron is used to extract multi-layer and multi-granular features. Each layer of multi-granular features is input into a dual-path attention mechanism for processing. The weighted features of each layer are independently max-pooled and then spliced ​​to form a three-layer joint feature vector corresponding to the three point cloud subsets. The final feature vector is then formed through the splicing operation and the multi-layer perceptron. S3. Use the pyramid point generator to perform hierarchical reconstruction. The low-dimensional, medium-dimensional, and high-dimensional features in the final feature vector are used to generate the main center point, secondary center point, and detail point respectively. Finally, the predicted point cloud of the missing area is formed by splicing, accurately fitting the local geometric features. S4. Design a composite loss function between the predicted point cloud and the real point cloud in the missing area, and design a dynamic weight scheduling strategy to simultaneously optimize multiple reconstruction objectives; S5. Using the adversarial training framework, the discriminator applies spectral normalization to constrain the spectral norm of the weight matrix and enhances training stability through the gradient penalty mechanism; S6. During the inference process, the residual point cloud is input. After steps S1-S3, the pyramid point generator outputs high-quality 3D predicted point cloud data of the missing area. This predicted point cloud data can be spliced ​​with the input residual point cloud to output a complete reconstruction result.

[0009] Furthermore, in step S2, the specific process includes the following steps: S21, the extended combined multilayer perceptron comprises a five-layer structure, each layer comprising a set of fully connected layers, the dimensions of each set of fully connected layers increasing in sequence, and the five-layer structure is capable of combining low-dimensional geometric details with high-dimensional semantic information; S22. Extended combination of multi-layer perceptron to extract multi-granularity features from each layer The input is processed by a dual-path attention mechanism, which includes two paths in series, namely a channel attention path and a spatial attention path; S23, the channel attention path, multi-granularity features Generate channel descriptors through adaptive global average pooling, learn nonlinear channel relationships through a fully connected layer with dynamic scaling ratio, and finally activate the output channel weights through Sigmoid. , achieving dynamic selection of key geometric features; S24, the spatial attention path, with channel weights Processing multi-granularity features , the channel weighted features formed As input, average pooling and maximum pooling are performed in parallel along the channel dimension to generate a bimodal spatial map, and the spatial weights are output after splicing and convolution fusion. , accurately locate spatial saliency; S25, channel weights output by the above two paths and spatial weights Multi-granularity features Processing is performed to generate weighted features by element-by-element multiplication of feature maps , achieving synergistic enhancement: ; in Represents element-by-element multiplication operation, all levels of weighted features After independent maximum pooling, they are concatenated in ascending order of dimension to form a joint feature vector ; S26, the three-layer point cloud subsets generated in step S1 are used as the input of the extended combined multi-layer perceptron, and the joint feature vectors corresponding to the three layers are output and then spliced, and then the final feature vector is generated through a layer of MLP. .

[0010] Furthermore, the specific process in step S3 includes the following steps: S31, the main center point is generated by mapping the low-dimensional features through the fully connected layer as the global skeleton; S32, the secondary center points are generated by diffusion based on the medium-dimensional features through a three-layer MLP network and the main center points to form a medium-density topological structure; S33, the detail points are synthesized by high-dimensional features through a deep MLP network in the neighborhood of the secondary center point to construct a high-density surface point cloud; S34: The main center point, the secondary center point, and the detail points are sequentially spliced ​​to form a predicted point cloud of the missing area.

[0011] Furthermore, the hierarchical reconstruction described in step S3 introduces a multi-stage constraint mechanism: the main center point and the secondary center point are respectively aligned with the downsampling results of the corresponding layers of the real point cloud to ensure the accuracy of the network's priority learning of the topological skeleton.

[0012] Furthermore, in step S4, the composite loss function is designed, and the specific process includes the following steps: S41, the composite loss function includes a nearest neighbor distance loss for constraining global shape alignment, an extreme value distance loss for suppressing local geometric deviations, a transmission distance loss for optimizing point distribution uniformity, and a multi-stage constraint loss for enforcing skeleton topology priority convergence in a multi-stage constraint mechanism; S42, the nearest neighbor distance loss, traversing each point in the predicted point cloud to find its nearest point in the real point cloud and accumulating the square of the distance, and traversing each point in the real point cloud to find its nearest point in the predicted point cloud and accumulating the square of the distance. The sum of the two items constitutes the global shape alignment constraint; S43, the extreme distance loss is calculated by taking the maximum value of the maximum and minimum distances between the predicted point cloud and the real point cloud and the maximum and minimum distances between the real point cloud and the predicted point cloud, specifically suppressing the deformation of thin-walled structures and sharp edges; S44. The transmission distance loss is measured by establishing a point-to-point mapping relationship based on the optimal transmission theory, effectively eliminating surface holes and point cluster aggregation artifacts.

[0013] Furthermore, the specific process in step S5 includes the following steps: S51, all convolutional layer weights of the discriminator are spectrally normalized; S52, the gradient penalty mechanism calculates the gradient penalty term by linear interpolation between the predicted point cloud and the real point cloud, and the discriminator and generator use a preset ratio to alternately update parameters to maintain training stability.

[0014] The present invention also provides a high-fidelity point cloud completion system based on dual-path attention and fractal structure, which is used to apply the high-fidelity point cloud completion method based on dual-path attention and fractal structure, including: The multi-scale feature extraction module performs iterative farthest point sampling and feature encoding; The dual-path attention fusion module realizes channel weighting and spatial saliency positioning; The pyramid point generation module outputs the predicted point cloud in three stages: main center point, secondary center point, and detail point; The spectral normalization discriminator receives the predicted point cloud and outputs a realism score; The hybrid loss calculation unit simultaneously optimizes the global shape alignment constraint, local geometric deviation constraint, point distribution uniformity constraint, and skeleton topology priority convergence constraint.

[0015] Furthermore, the pyramid point generation module includes a main center point generation unit that maps the global skeleton structure, a secondary center point diffusion unit that inherits the main center point topological relationship and refines the mesoscopic structure, and a detail point synthesis unit that generates a surface point cloud that adapts to the local curvature in the secondary center point neighborhood.

[0016] Furthermore, the hybrid loss calculation unit adopts a dynamic weighting mechanism, focusing on global shape alignment constraints in the early stage of training, enhancing local geometric deviation constraints in the middle stage of training, and improving the point distribution uniformity constraint weight in the late stage of training.

[0017] Through the above technical solution, the present invention provides a high-fidelity point cloud completion method and system based on dual-path attention and fractal structure, which has at least the following beneficial effects: (1) This paper realizes the adaptive fusion of multi-scale features through an innovative dual-path attention mechanism, solving the problem of geometric information attenuation in traditional methods; (2) The present invention ensures the continuity of complex topological structures by designing a pyramid point generator based on a fractal-guided hierarchical generation architecture; (3) This paper effectively balances the reconstruction quality of global shape and local details by designing a composite loss function and a dynamic weight scheduling strategy; (4) Experiments show that this method can achieve better completion effects than existing technologies in challenging scenarios such as thin-walled structures and hole areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings: Figure 1 This is a diagram showing the overall architecture of a high-fidelity point cloud completion method based on dual-path attention and fractal structure in the present invention; Figure 2 This is a schematic diagram of the structural innovation of the expanded combined multi-layer perceptron of the present invention; Figure 3 Schematic diagram of the dual-path attention mechanism of the present invention; Figure 4 This is a fractal generation flow chart of the pyramid point generator of the present invention; Figure 5 This is a qualitative analysis diagram of point cloud completion for various types of objects in an embodiment of the present invention; Figure 6 This is a diagram showing the ablation experiment results in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] To make the above-mentioned objectives, features, and advantages of the present invention more clearly understood, the present invention is further described below in detail with reference to the accompanying drawings and specific embodiments. This will enable a full understanding of how this application uses technical means to solve technical problems and achieve technical effects, and to implement the invention accordingly.

[0020] Those skilled in the art will appreciate that all or part of the steps in the above-mentioned embodiment methods can be accomplished by instructing the relevant hardware through a program. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0021] Please refer to Figures 1-6 , showing a specific implementation method of this embodiment, which provides a geometric information basis for model input through an iterative farthest point sampling algorithm; designs an extended combined multi-layer perceptron combined with dual-path attention to achieve dynamic enhancement of key geometric features; designs a pyramid point generator to ensure geometric coherence from macro to micro in a progressive generation manner; designs a composite loss function and a dynamic weight scheduling strategy to achieve stable multi-objective optimization; designs an adversarial training framework with built-in spectral normalization and gradient penalty mechanism to achieve good training balance, significantly improves the quality of point cloud completion in complex structure scenes, and provides reliable technical support for three-dimensional vision applications.

[0022] Please refer to Figure 1This embodiment proposes a high-fidelity point cloud completion method based on dual-path attention and fractal structure, which includes the following steps: S1. The input residual point cloud is processed by the iterative farthest point sampling algorithm to generate three-layer point cloud subsets including original resolution, medium resolution and low resolution, which respectively retain the complete geometric details, main structural features and global topological skeleton; Specifically, the input residual point cloud is first processed by the Iterative Farthest Point Sampling (IFPS) algorithm, which recursively selects the point farthest from the existing point set based on the geodesic distance priority principle. The specific calculation process is: Find the point that satisfies the minimum distance maximization condition ,Right now ,in Belongs to the selected point set The output is three layers of decreasing resolution subsets: the original resolution subset (1024 points) retains complete geometric details, the medium resolution subset (512 points) extracts the main structural features, and the low resolution subset (256 points) focuses on global topological skeleton construction. This layered sampling mechanism provides a hierarchical geometric information foundation for subsequent feature learning.

[0023] In this embodiment, the point cloud data is preprocessed to form the data set required for the training model. Specifically, the ShapeNet three-dimensional model library is used as the basic data set, from which 16 types of objects with typical geometric structures such as furniture and vehicles are selected. The complete three-dimensional model is uniformly sampled to generate a reference point cloud, and the default number of sampling points is 2048. In order to simulate the occlusion defects in real sensor acquisition, a multi-view sphere culling strategy is implemented: the center position of the spatial sphere is randomly generated, and the point cloud within a preset radius (default 0.3 unit length) is used as the mask area for deletion. The missing ratio control of 25%-75% is achieved by adjusting the number of spheres. All point cloud coordinates are normalized to the [-1,1] interval by linear transformation to eliminate the influence of model size differences. The preprocessed data is divided into a training set and a test set at a ratio of 80%:20%.

[0024] S2. For each of the three point cloud subsets, an extended combined multi-layer perceptron is used to extract multi-layer and multi-granular features. Each layer of multi-granular features is input into a dual-path attention mechanism for processing. The weighted features of each layer are independently max-pooled and then spliced ​​to form a three-layer joint feature vector corresponding to the three point cloud subsets. The final feature vector is then formed through the splicing operation and the multi-layer perceptron. As a preferred implementation of step S2, the specific process includes the following steps: S21, the three-layer point cloud subsets obtained in step S1 are input in parallel to the extended combined multilayer perceptron (E-CMLP), which can be found in Figure 2The extended combined multilayer perceptron comprises a five-layer structure, each layer comprising a set of fully connected layers with increasing dimensions (64D / 128D / 256D / 512D / 1024D). This five-layer structure is capable of combining low-dimensional geometric details with high-dimensional semantic information. S22. Extending and combining the multi-granularity features extracted from each layer of the multi-layer perceptron and inputting them into a dual-path attention mechanism for processing. The dual-path attention mechanism includes two paths connected in series, namely, a channel attention path and a spatial attention path. S23, the channel attention path, such as Figure 3 As shown in the above process, is the channel weight matrix, multi-granularity feature The channel descriptor is generated by compressing spatial information through adaptive global average pooling, and the nonlinear channel relationship is learned through a fully connected layer with a dynamic scaling ratio. The scaling ratio calculation formula is: ; in Indicates the zoom ratio, For the number of input channels, this design can adaptively adjust the feature compression rate. Represents a floor function that acts on the calculation result in the brackets , take the maximum integer that is not greater than the calculated result. Then finally activate the output channel weight through Sigmoid ,strengthening the discriminative feature response of edges, corners, etc., and realizing the dynamic selection of key geometric features; Figure 3 Medium- and multi-granularity features The width and height are expressed as and ; S24, the spatial attention path, such as Figure 3 As shown in the following process, is the spatial weight matrix, which is composed of channel weights Processing multi-granularity features , the channel weighted features formed As input, average pooling is performed in parallel along the channel dimension ( ) and max pooling ( ), respectively capturing regional consistency features and local extreme value features, generating a bimodal spatial map, generating bimodal features through splicing operations, and outputting spatial weights after bimodal features are fused through 7×7 large kernel convolution. , accurately locate high curvature areas and structural fault zones; S25, channel weights output by the above two paths and spatial weights Multi-granularity features Processing is performed to generate weighted features by element-by-element multiplication of feature maps , achieving synergistic enhancement: ; in Represents element-by-element multiplication operation, all levels of weighted features After independent maximum pooling, the dimensions are concatenated in ascending order to form a 1984-dimensional joint feature vector The process can be found in Figure 2 . Figure 3 Weighted features Apply max pooling to generate features , its width, height and number of channels are expressed as 、 、 .

[0025] S26, the three-layer point cloud subsets generated in step S1 are used as the input of the extended combined multi-layer perceptron in the multi-scale feature extractor (MFE), and the joint feature vectors corresponding to the three layers are output as the processed feature vector, which are then concatenated to generate the final processed feature vector, and then passed through a layer of MLP to generate the final feature vector .

[0026] In this embodiment, the extended combined multi-layer perceptron is integrated with an innovative dual-path attention mechanism to achieve adaptive fusion of multi-scale features and solve the problem of geometric information attenuation in traditional methods.

[0027] S3, using the pyramid point generator to perform layered reconstruction, the final feature vector The low-dimensional, medium-dimensional, and high-dimensional features in the image are respectively used to generate the main center point, secondary center point, and detail point, and finally the predicted point cloud of the missing area is formed by splicing to accurately fit the local geometric features; the hierarchical reconstruction process is as follows Figure 4 As shown in the figure Represents the number of main center points at the coarsest granularity level, Represents the number of secondary centers at the intermediate granularity level, The number of minutiae points in the final output representing the finest level of granularity.

[0028] As a preferred implementation of step S3, the specific process includes the following steps: S31, final feature vector The low-dimensional sub-eigenvector of (i.e., low-dimensional features) are mapped through a single-layer fully connected network and then convolved and reconstructed to generate 512 main center points, i.e., a sparse point cloud, which forms the global skeleton of the missing area. For example, in the chair back completion task, the output of this layer forms the main frame structure, which can be expressed mathematically as: ; in, represents the main center point, represents a full connection operation, Corresponding to the final eigenvector The front low-dimensional component (i.e., low-dimensional feature) carries the geometric skeleton information; S32, the intermediate sub-eigenvector of the final eigenvector (i.e., medium-dimensional features) input three-layer MLP network (hidden layer 256 units) and convolution and reconstruction operations to predict coordinate offsets , superimposed on the coordinates of the main center point, expanding to form 128 secondary points, namely secondary center points: ; in, This step realizes the diffusion of the skeleton to the medium-density topology, forming a medium-density topological structure, and accurately maintains the continuity of the hollow boundary of the chair back. Figure 5 Qualitative analysis of point cloud completion for the input point cloud, completed points, and real point cloud of each category of objects.

[0029] S33, final feature vector The high-dimensional sub-eigenvector of (i.e. high-dimensional features) are used to predict micro-displacements through deep MLP (hidden layer 512 units) and convolution and reconstruction operations , expand the secondary point neighborhood to generate 64 high-density surface points, namely detail points: ; in, Represents detail points. This layer accurately fits local geometric variations, such as the circular profile of a chair leg cross section or the axial curvature of a lamp bracket.

[0030] S34: The main center point, the secondary center point, and the detail points are reconstructed and added in sequence to form a predicted point cloud of the missing area.

[0031] More specifically, in step S3, the hierarchical reconstruction process introduces a multi-stage constraint mechanism: the main center point and the secondary center point are respectively aligned with the downsampling results of the corresponding layers of the real point cloud to ensure the accuracy of the network's priority learning of the topological skeleton.

[0032] In this embodiment, this “skeleton-branch-skin” progressive generation method ensures geometric coherence from macro to micro.

[0033] S4. Design a composite loss function between the predicted point cloud and the real point cloud in the missing area, and design a dynamic weight scheduling strategy to simultaneously optimize multiple reconstruction objectives As a preferred implementation of step S4, the design of the composite loss function specifically includes the following steps: S41, the composite loss function includes a nearest neighbor distance loss (CD) for constraining global shape alignment, an extreme distance loss (HD) for suppressing local geometric deviations, an transmission distance loss (EMD) for optimizing point distribution uniformity, and a multi-stage constraint loss for enforcing skeleton topology priority convergence in a multi-stage constraint mechanism; S42, the nearest neighbor distance loss, traversing each point in the predicted point cloud to find its nearest point in the real point cloud and accumulating the square of the distance, and traversing each point in the real point cloud to find its nearest point in the predicted point cloud and accumulating the square of the distance. The sum of the two items constitutes the global shape alignment constraint; S43, the extreme distance loss is calculated by taking the maximum value of the maximum and minimum distances between the predicted point cloud and the real point cloud and the maximum and minimum distances between the real point cloud and the predicted point cloud, specifically suppressing the deformation of thin-walled structures and sharp edges; S44. The transmission distance loss is measured by establishing a point-to-point mapping relationship based on the optimal transmission theory, effectively eliminating surface holes and point cluster aggregation artifacts.

[0034] In this embodiment, the nearest neighbor distance loss ensures the alignment and coverage of global shapes; the extreme value distance loss specifically suppresses local geometric deviations and enhances the reconstruction accuracy of thin-walled structures and sharp edges; the transmission distance loss optimizes the uniformity of point distribution based on the optimal transmission theory; the multi-stage constraint loss obtains the feature point set by downsampling the real point cloud, forcing the primary / secondary center points to be hierarchically aligned with the corresponding feature point set to ensure the priority accuracy of the skeleton structure.

[0035] In this embodiment, in order to coordinate the conflicts among multiple objectives, a dynamic weight scheduling strategy is introduced. Specifically, a dominant weight of 0.4 is assigned to the nearest neighbor distance loss within the initial 30 rounds of training to accelerate global shape convergence; the extreme value distance loss weight is gradually increased to 0.3 during the 30th to 80th rounds of training, focusing on optimizing local geometric accuracy; after 80 rounds, the transmission distance loss weight is enhanced to 0.3, focusing on improving the uniformity of point distribution.

[0036] S5. Using the adversarial training framework, the discriminator applies spectral normalization to constrain the spectral norm of the weight matrix and enhances training stability through the gradient penalty mechanism; As a preferred implementation of step S5, the specific process includes the following steps: S51, the discriminator, all its convolutional layer weights are spectrally normalized, that is, the weight matrix is ​​divided by its maximum singular value, strictly constraining the Lipschitz constant: ; in, represents the weight matrix, yes The maximum singular value of is the weight matrix after spectral normalization; S52, the gradient penalty mechanism uses the Wasserstein GAN with Gradient Penalty (WGAN-GP) framework to linearly interpolate samples between the real point cloud and the predicted point cloud to generate interpolation sampling points : ; in, represents the real point cloud, represents the predicted point cloud, Is the difference weight parameter. Calculate the discriminator at these interpolation sampling points The gradient norm at and define the gradient penalty term , constraining its deviation from 1: ; in, Indicates the interpolation sampling points , and ensure that the penalty term takes effect on the overall distribution. represents the discriminator gradient, Indicates calculation of L2 norm.

[0037] In this embodiment, the discriminator and the generator update parameters alternately in a 5:1 ratio, maintaining a good adversarial balance during training. Figure 6 The ablation experiment results of each component of the model in this embodiment are shown, where (a) is the input point cloud; (b) is the point cloud completion result without the dual-path attention mechanism in step S2; (c) is the point cloud completion result without the extended combined multi-layer perceptron in step S2; (d) is the point cloud completion result without the gradient penalty mechanism in step S5; (e) is the point cloud completion result when this method is fully applied; and (f) is the real point cloud.

[0038] S6. During the inference process, the residual point cloud is input. After steps S1-S3, the pyramid point generator outputs high-quality 3D predicted point cloud data of the missing area. This predicted point cloud data can be spliced ​​with the input residual point cloud to output a complete reconstruction result.

[0039] In this embodiment, during actual deployment, the input residual point cloud is first generated using the IFPS algorithm to generate subsets at three resolution levels (1024 points, 512 points, and 256 points). These subsets are then fed into the E-CMLP module in parallel to extract features weighted by a two-way attention mechanism, thereby obtaining a joint feature vector. The joint feature vector is then fed into the pyramid point generator, which sequentially generates 512 primary center points (representing the global skeleton), 128 secondary center points (building topological branches), and 64 detail points (fitting the surface geometry). These are then merged to form a predicted point cloud for the missing region. The predicted point cloud is then concatenated with the input residual point cloud to produce a complete reconstruction. The proposed method significantly improves the quality of point cloud completion in complex structural scenes, providing reliable technical support for 3D vision applications.

[0040] This embodiment further proposes a high-fidelity point cloud completion system based on dual-path attention and fractal structure, which is used to apply a high-fidelity point cloud completion method based on dual-path attention and fractal structure, including: The multi-scale feature extraction module performs iterative farthest point sampling and feature encoding; The dual-path attention fusion module realizes channel weighting and spatial saliency positioning; The pyramid point generation module outputs the predicted point cloud in three stages: main center point, secondary center point, and detail point; The spectral normalization discriminator receives the predicted point cloud and outputs a realism score; The hybrid loss calculation unit simultaneously optimizes the global shape alignment constraint, local geometric deviation constraint, point distribution uniformity constraint, and skeleton topology priority convergence constraint.

[0041] Furthermore, the pyramid point generation module includes a main center point generation unit that maps the global skeleton structure, a secondary center point diffusion unit that inherits the main center point topological relationship and refines the mesoscopic structure, and a detail point synthesis unit that generates a surface point cloud that adapts to the local curvature in the secondary center point neighborhood.

[0042] Furthermore, the hybrid loss calculation unit adopts a dynamic weighting mechanism, focusing on global shape alignment constraints in the early stage of training, enhancing local geometric deviation constraints in the middle stage of training, and improving the point distribution uniformity constraint weight in the late stage of training.

[0043] In this embodiment, the proposed method and system demonstrate significant advantages in industrial scenarios: in the field of autonomous driving, they successfully repair road loss caused by vehicle occlusion in the LiDAR point cloud; in the industrial quality inspection process, they achieve precision control of the tooth profile error within 0.3mm for the reconstruction of a single missing tooth of a gear with a module of 2; and in the field of cultural relics digitization, they accurately complete the decorative patterns at the broken handle in the scanning of bronze artifacts.

[0044] In the description of this specification, the reference terms "one embodiment," "some embodiments," "example," "specific example," or "some examples" mean that the specific features, structures, materials, or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. Moreover, the specific features, structures, materials, or characteristics described may be combined in any appropriate manner in any one or more embodiments or examples. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification, as well as features of different embodiments or examples, unless they are mutually inconsistent.

[0045] The logic and / or steps represented in the flowchart or otherwise described herein may be considered, for example, as an ordered list of executable instructions for implementing logical functions, and may be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device).

[0046] The above embodiments provide a detailed introduction to the present invention. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. At the same time, for those skilled in the art, according to the ideas of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the contents of this specification should not be understood as limiting the present invention.

Claims

1. A high-fidelity point cloud completion method based on dual-path attention and fractal structure, characterized by: The following steps are involved: S1. Construct a dataset containing residual point clouds. The residual point clouds of the input model are processed by an iterative farthest point sampling algorithm to generate three-layer point cloud subsets including original resolution, medium resolution, and low resolution, respectively retaining complete geometric details, main structural features, and global topological skeleton; S2. For each of the three point cloud subsets, an extended combined multi-layer perceptron is used to extract multi-layer and multi-granular features. Each layer of multi-granular features is input into a dual-path attention mechanism for processing. The weighted features of each layer are independently max-pooled and then spliced ​​to form a three-layer joint feature vector corresponding to the three point cloud subsets. The final feature vector is then formed through the splicing operation and the multi-layer perceptron. S3. Use the pyramid point generator to perform hierarchical reconstruction. The low-dimensional, medium-dimensional, and high-dimensional features in the final feature vector are used to generate the main center point, secondary center point, and detail point respectively. Finally, the predicted point cloud of the missing area is formed by splicing, accurately fitting the local geometric features. S4. Design a composite loss function between the predicted point cloud and the real point cloud in the missing area, and design a dynamic weight scheduling strategy to simultaneously optimize multiple reconstruction objectives; S5. Using the adversarial training framework, the discriminator applies spectral normalization to constrain the spectral norm of the weight matrix and enhances training stability through the gradient penalty mechanism; S6. During the inference process, the residual point cloud is input. After steps S1-S3, the pyramid point generator outputs high-quality 3D predicted point cloud data of the missing area. This predicted point cloud data can be spliced ​​with the input residual point cloud to output a complete reconstruction result.

2. The high-fidelity point cloud completion method based on dual-path attention and fractal structure according to claim 1, characterized in that: The specific process of step S2 includes the following steps: S21, the extended combined multilayer perceptron comprises a five-layer structure, each layer comprising a set of fully connected layers, the dimensions of each set of fully connected layers increasing in sequence, and the five-layer structure is capable of combining low-dimensional geometric details with high-dimensional semantic information; S22. Extended combination of multi-layer perceptron to extract multi-granularity features from each layer The input is processed by a dual-path attention mechanism, which includes two paths in series, namely a channel attention path and a spatial attention path; S23, the channel attention path, multi-granularity features Generate channel descriptors through adaptive global average pooling, learn nonlinear channel relationships through a fully connected layer with dynamic scaling ratio, and finally activate the output channel weights through Sigmoid. , achieving dynamic selection of key geometric features; S24, the spatial attention path, with channel weights Processing multi-granularity features , the channel weighted features formed As input, average pooling and maximum pooling are performed in parallel along the channel dimension to generate a bimodal spatial map, and the spatial weights are output after splicing and convolution fusion. , accurately locate spatial saliency; S25, channel weights output by the above two paths and spatial weights Multi-granularity features Processing is performed to generate weighted features by element-by-element multiplication of feature maps , achieving synergistic enhancement: ； in Represents element-by-element multiplication operation, all levels of weighted features After independent maximum pooling, they are concatenated in ascending order of dimension to form a joint feature vector ; S26, the three-layer point cloud subsets generated in step S1 are used as the input of the extended combined multi-layer perceptron, and the joint feature vectors corresponding to the three layers are output and then spliced, and then the final feature vector is generated through a layer of MLP. .

3. The high-fidelity point cloud completion method based on dual-path attention and fractal structure according to claim 1, characterized in that: The specific process in step S3 includes the following steps: S31, the main center point is generated by mapping the low-dimensional features through the fully connected layer as the global skeleton; S32, the secondary center points are generated by diffusion based on the medium-dimensional features through a three-layer MLP network and the main center points to form a medium-density topological structure; S33, the detail points are synthesized by high-dimensional features through a deep MLP network in the neighborhood of the secondary center point to construct a high-density surface point cloud; S34: The main center point, the secondary center point, and the detail points are sequentially spliced ​​to form a predicted point cloud of the missing area.

4. The high-fidelity point cloud completion method based on dual-path attention and fractal structure according to claim 3, characterized in that: The hierarchical reconstruction described in step S3 introduces a multi-stage constraint mechanism: the main center point and the secondary center point are respectively aligned with the downsampling results of the corresponding layers of the real point cloud to ensure the accuracy of the network's priority learning of the topological skeleton.

5. The high-fidelity point cloud completion method based on dual-path attention and fractal structure according to claim 4, characterized in that: In step S4, the composite loss function is designed, and the specific process includes the following steps: S41, the composite loss function includes a nearest neighbor distance loss for constraining global shape alignment, an extreme value distance loss for suppressing local geometric deviations, a transmission distance loss for optimizing point distribution uniformity, and a multi-stage constraint loss for enforcing skeleton topology priority convergence in a multi-stage constraint mechanism; S42, the nearest neighbor distance loss, traversing each point in the predicted point cloud to find its nearest point in the real point cloud and accumulating the square of the distance, and traversing each point in the real point cloud to find its nearest point in the predicted point cloud and accumulating the square of the distance. The sum of the two items constitutes the global shape alignment constraint; S43, the extreme distance loss is calculated by taking the maximum value of the maximum and minimum distances between the predicted point cloud and the real point cloud and the maximum and minimum distances between the real point cloud and the predicted point cloud, specifically suppressing the deformation of thin-walled structures and sharp edges; S44. The transmission distance loss is measured by establishing a point-to-point mapping relationship based on the optimal transmission theory, effectively eliminating surface holes and point cluster aggregation artifacts.

6. The high-fidelity point cloud completion method based on dual-path attention and fractal structure according to claim 1, characterized in that: The specific process in step S5 includes the following steps: S51, all convolutional layer weights of the discriminator are spectrally normalized; S52, the gradient penalty mechanism calculates the gradient penalty term by linear interpolation between the predicted point cloud and the real point cloud, and the discriminator and generator use a preset ratio to alternately update parameters to maintain training stability.

7. A high-fidelity point cloud completion system based on dual-path attention and fractal structure, for applying the high-fidelity point cloud completion method based on dual-path attention and fractal structure according to any one of claims 1 to 6, characterized in that: include: The multi-scale feature extraction module performs iterative farthest point sampling and feature encoding; The dual-path attention fusion module realizes channel weighting and spatial saliency positioning; The pyramid point generation module outputs the predicted point cloud in three stages: main center point, secondary center point, and detail point; The spectral normalization discriminator receives the predicted point cloud and outputs a realism score; The hybrid loss calculation unit simultaneously optimizes the global shape alignment constraint, local geometric deviation constraint, point distribution uniformity constraint, and skeleton topology priority convergence constraint.

8. The high-fidelity point cloud completion system based on dual-path attention and fractal structure according to claim 7, characterized in that: The pyramid point generation module includes a main center point generation unit that maps the global skeleton structure, a secondary center point diffusion unit that inherits the main center point topological relationship and refines the mesoscopic structure, and a detail point synthesis unit that generates a surface point cloud that adapts to the local curvature in the secondary center point neighborhood.

9. The high-fidelity point cloud completion system based on dual-path attention and fractal structure according to claim 7, characterized in that: The hybrid loss calculation unit adopts a dynamic weighting mechanism, focusing on global shape alignment constraints in the early stage of training, enhancing local geometric deviation constraints in the middle stage of training, and improving the point distribution uniformity constraint weight in the late stage of training.

Images