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

Abstract

The application relates to a high-fidelity point cloud completion method and system based on a double-path attention and fractal structure, which comprises an iterative farthest point sampling algorithm for processing input point clouds; an extended combined multi-layer perception machine is combined with a double-path attention mechanism to extract and construct a joint feature vector; a pyramid point generator is used to hierarchically reconstruct a missing area point cloud; a composite loss function and a dynamic weight scheduling strategy are designed to optimize a reconstruction target; an adversarial training framework of a discriminator based on an embedded spectrum normalization and gradient penalty mechanism is designed; and a point cloud reconstruction result is generated through reasoning. The application solves the dependence of a traditional point cloud completion method on geometric prior and a matching template, alleviates the lack of calibration of different dimensional semantic information, the excessive dependence on maximum pooling, and the limitations of abnormal discriminator gradient and loss function difficulty in ensuring uniformity, and obviously improves the geometric detail recovery capability and visual authenticity in a complex structure scene.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and 3D reconstruction technology, and in particular to a high-fidelity point cloud completion method and system based on dual-path attention and fractal structure. Background Technology

[0002] In the field of 3D visual perception, point cloud data, as a discretized carrier of object surface geometric information, has become a core input for applications such as industrial quality inspection and autonomous driving environment modeling. However, in actual acquisition processes, point clouds often suffer from local structural defects and topological breaks due to sensor field-of-view occlusion, dynamic target motion blur, and complex surface reflection characteristics. These data defects severely restrict the reliability of downstream tasks: for example, in the inspection of aero-engine blades, incomplete cooling hole point clouds can lead to distorted airflow simulation; in autonomous driving scenarios, incomplete obstacle outlines may cause collision prediction errors.

[0003] Traditional point cloud completion methods mainly rely on geometric priors or template matching mechanisms. Interpolation methods based on geometric priors assume that the surface is continuous and differentiable, but when dealing with non-smooth geometries such as thin-walled structures and sharp edges, they often produce overfilling or structural collapse, such as the curvature distortion that occurs when completing the leading edge of an engine blade. Registration methods based on template matching require searching for similar models for rigid alignment, which is not only computationally inefficient, but also difficult to accurately match when local features of the component are missing (such as missing teeth on a gear), resulting in semantically contradictory reconstruction results.

[0004] In recent years, deep learning point cloud completion methods have made significant progress. Methods represented by global feature generation networks (such as PCN) achieve overall reconstruction through end-to-end encoding and decoding, but due to neglecting the correlation of local structures, they are prone to incorrectly filling in solid blocks when completing the hollow area of ​​a chair back. Local region prediction methods (such as PF-Net) retain known structures and predict missing parts, but their multi-scale feature fusion mechanism lacks adaptive weighting capabilities, still resulting in grid breaks when reconstructing the grid-like lampshade of a lamp. Iterative optimization methods (such as PMPNet++) improve the continuity of details through multi-step displacement refinement, but the coordinate prediction error in the early stage is amplified in the hierarchical propagation, causing the hub center to drift when completing bicycle wheel spokes.

[0005] Current technical bottlenecks are mainly manifested 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 dynamic calibration mechanisms, such as the excessive smoothing of gradient changes at hole boundaries; Local geometric aggregation operations rely too much on fixed strategies such as max pooling, resulting in insufficient representation capabilities of high-variance structures (such as thread gaps), and noise or collapse frequently appears on the completed surface; In the adversarial training framework, gradient anomalies of the discriminator are prone to causing pattern collapse, and the generated point cloud often degenerates into a single topological form (such as uniformly completing multiple table legs into cylinders), and existing loss functions are difficult to constrain the uniformity of point set distribution at the micro level.

[0006] Although SA-Net optimizes feature selection through cross-layer attention and SnowflakeNet proposes fractal decoding to enhance detail generation, these methods still have inherent limitations: the former has insufficient spatial attention response to sparse and missing regions, and the latter still produces local ambiguity in complex topological scenes. Therefore, there is an urgent need to construct a novel completion architecture that integrates adaptive feature weighting, hierarchical geometric reconstruction, and stable training mechanisms to meet the dual requirements of structural fidelity and detail realism for high-precision 3D perception applications. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a high-fidelity point cloud completion method and system based on dual-path attention and fractal structure. It 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 in different dimensions, excessive reliance on max pooling, discriminator gradient anomalies, and loss functions failing to guarantee point uniformity.

[0008] To address the aforementioned technical problems, this 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:

[0009] S1. Construct a dataset containing residual point clouds. Process the residual point clouds of the input model through an iterative farthest point sampling algorithm to generate a three-layer point cloud subset containing the original resolution, medium resolution, and low resolution, respectively preserving complete geometric details, main structural features, and global topological skeleton.

[0010] S2. For the three-layer point cloud subsets, multi-layer multi-granularity features are extracted using an extended combined multi-layer perceptron. Each multi-granularity feature is input into a dual-path attention mechanism for processing. The weighted features of each layer after processing are independently max-pooled and then concatenated to form a three-layer joint feature vector corresponding to the three-layer point cloud subsets. Finally, the concatenation operation and the multi-layer perceptron are used to form the final feature vector.

[0011] 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 minutiae point respectively. Finally, the points are stitched together to form the predicted point cloud of the missing area and accurately fit the local geometric features.

[0012] S4. Design a composite loss function between the predicted point cloud and the real point cloud in the missing region, and design a dynamic weight scheduling strategy to simultaneously optimize multiple reconstruction objectives.

[0013] S5. An adversarial training framework is adopted, the discriminator applies spectral normalization to constrain the spectral norm of the weight matrix, and the training stability is enhanced by a gradient penalty mechanism.

[0014] 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 stitched together with the input residual point cloud to output a complete reconstruction result.

[0015] Furthermore, in step S2, the specific process includes the following steps:

[0016] S21. The extended combined multilayer perceptron includes a five-layer structure, each layer containing a set of fully connected layers, with the dimensions of each set of fully connected layers increasing in sequence. This five-layer structure can combine low-dimensional geometric details with high-dimensional semantic information.

[0017] S22, Extended combined multilayer perceptron extracts multi-granularity features from each layer The input is processed using a dual-path attention mechanism, which includes two concatenated paths: a channel attention path and a spatial attention path.

[0018] S23, the channel attention path, multi-granularity features Channel descriptors are generated through adaptive global average pooling, nonlinear channel relationships are learned through dynamically scaled fully connected layers, and finally, the output channel weights are activated by Sigmoid. This enables dynamic selection of key geometric features;

[0019] S24, the spatial attention path, is determined by channel weights. Processing multi-granularity features The resulting channel weighted features As input, average pooling and max pooling are performed in parallel along the channel dimension to generate a bimodal spatial graph. After concatenation and fusion, the spatial weights are output. Precisely locate spatial saliency;

[0020] S25. Channel weights output from the two paths mentioned above. and spatial weights Multi-granularity features The process involves multiplying the feature maps element-wise to generate weighted features. To achieve synergistic enhancement:

[0021] ;

[0022] in This represents element-wise multiplication, with weighted features at all levels. After independent max pooling, the vectors are concatenated in ascending order of dimension to form a joint feature vector. ;

[0023] S26. The three-layer point cloud subsets generated in step S1 are used as inputs to the extended combined multilayer perceptron. The joint feature vectors of the three layers are then concatenated and passed through an MLP layer to generate the final feature vector. .

[0024] Furthermore, the specific process in step S3 includes the following steps:

[0025] S31. The main center point is generated by mapping low-dimensional features through a fully connected layer, serving as the global skeleton;

[0026] S32. The secondary center points are generated by diffusing from the main center points through a three-layer MLP network based on the mid-dimensional features, forming a medium-density topology.

[0027] S33. The detailed points are synthesized from high-dimensional features through a deep MLP network in the neighborhood of the secondary center point to construct a high-density surface point cloud.

[0028] S34. The main center point, secondary center point, and detail points are sequentially pieced together to form the predicted point cloud of the missing area.

[0029] 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.

[0030] Furthermore, in step S4, the design of the composite loss function specifically includes the following steps:

[0031] S41. The composite loss function includes nearest neighbor distance loss for constraining global shape alignment, extreme value distance loss for suppressing local geometric deviations, transmission distance loss for optimizing point distribution uniformity, and multi-stage constraint loss for forcing skeleton topology to converge preferentially in the multi-stage constraint mechanism.

[0032] S42. The nearest neighbor distance loss is to traverse each point in the predicted point cloud to find its nearest point in the real point cloud and accumulate the squared distance. At the same time, it is to traverse each point in the real point cloud to find its nearest point in the predicted point cloud and accumulate the squared distance. The sum of the two constitutes the global shape alignment constraint.

[0033] S43. The extreme distance loss is the maximum value of the maximum and minimum distance between the predicted point cloud and the real point cloud and the maximum and minimum distance between the real point cloud and the predicted point cloud, which is specifically used to suppress the deformation of thin-walled structures and sharp edges.

[0034] S44. The transmission distance loss is measured based on the optimal transmission theory to establish a point-to-point mapping relationship, effectively eliminating surface holes and point cluster artifacts.

[0035] Furthermore, the specific process in step S5 includes the following steps:

[0036] S51. The discriminator has all its convolutional layer weights processed by spectral normalization.

[0037] S52. The gradient penalty mechanism calculates the gradient penalty term by linear interpolation between the predicted point cloud and the real point cloud. The discriminator and the generator alternately update parameters using a preset ratio to maintain training stability.

[0038] This invention also provides a high-fidelity point cloud completion system based on dual-path attention and fractal structure, used to apply a high-fidelity point cloud completion method based on dual-path attention and fractal structure, comprising:

[0039] The multi-scale feature extraction module performs iterative farthest point sampling and feature encoding;

[0040] The dual-path attention fusion module enables channel weighting and spatial saliency localization;

[0041] The pyramid point generation module outputs a predicted point cloud in three stages: main center point, secondary center point, and detail point.

[0042] The spectral normalization discriminator receives the predicted point cloud and outputs a truth score.

[0043] The hybrid loss calculation unit simultaneously optimizes global shape alignment constraints, local geometric deviation constraints, point distribution uniformity constraints, and skeleton topology-priority convergence constraints.

[0044] 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 topological relationship of the main center point and refines the mesoscopic structure, and a detail point synthesis unit that generates a surface point cloud in the neighborhood of the secondary center point that adapts to the local curvature.

[0045] Furthermore, the hybrid loss calculation unit adopts a dynamic weighting mechanism, which emphasizes global shape alignment constraints in the early stage of training, enhances local geometric deviation constraints in the middle stage of training, and increases the weight of point distribution uniformity constraints in the later stage of training.

[0046] By employing the above technical solutions, 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:

[0047] (1) This invention achieves adaptive fusion of multi-scale features through an innovative dual-path attention mechanism, which solves the problem of geometric information decay in traditional methods;

[0048] (2) This invention ensures the continuity of complex topological structures by designing a pyramid point generator and using a fractal-guided hierarchical generation architecture.

[0049] (3) This invention effectively balances the reconstruction quality of global shape and local details by designing a composite loss function and a dynamic weight scheduling strategy;

[0050] (4) Experiments show that this method can achieve better completion results than existing technologies in challenging scenarios such as thin-walled structures and porous areas. Attached Figure Description

[0051] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0052] Figure 1 This is a diagram illustrating the overall architecture of a high-fidelity point cloud completion method based on dual-path attention and fractal structure according to the present invention.

[0053] Figure 2 This is a schematic diagram illustrating the structural innovation of the extended combined multilayer perceptron of the present invention;

[0054] Figure 3 This is a schematic diagram of the dual-path attention mechanism of the present invention;

[0055] Figure 4 This is a flowchart of the fractal generation process of the pyramid point generator of the present invention;

[0056] Figure 5 This is a qualitative analysis diagram of point cloud completion for various types of objects in this embodiment of the invention;

[0057] Figure 6 This is a diagram showing the ablation experiment results in an embodiment of the present invention. Detailed Implementation

[0058] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0059] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can 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.

[0060] Please refer to Figures 1-6 This illustration demonstrates a specific implementation of this embodiment. This embodiment provides the geometric information basis for the model input through an iterative farthest point sampling algorithm; designs an extended combined multilayer perceptron incorporating dual-path attention to achieve dynamic enhancement of key geometric features; designs a pyramid point generator to ensure geometric coherence from macro to micro using a progressive generation method; designs a composite loss function and a dynamic weight scheduling strategy to achieve stable multi-objective optimization; and designs an adversarial training framework with built-in spectral normalization and gradient penalty mechanisms to achieve good training balance, significantly improving the point cloud completion quality in complex structural scenes and providing reliable technical support for 3D vision applications.

[0061] Please refer to Figure 1 This embodiment proposes a high-fidelity point cloud completion method based on dual-path attention and fractal structure. The method includes the following steps:

[0062] S1. The input residual point cloud is processed by the iterative farthest point sampling algorithm to generate a three-layer point cloud subset containing the original resolution, medium resolution and low resolution, respectively preserving complete geometric details, main structural features and global topological skeleton.

[0063] Specifically, the input residual point cloud is first processed by the Iterative Farthest Point Sampling (IFPS) algorithm. This algorithm recursively selects the point farthest from the existing point set based on the geodesic distance priority principle. The specific calculation process is as follows: in the point cloud set... Find the point that satisfies the minimum distance maximization condition. ,Right now ,in Belongs to the selected point set The output consists of three subsets with decreasing resolution: the original resolution subset (1024 points) preserves complete geometric details, the medium resolution subset (512 points) extracts main structural features, and the low resolution subset (256 points) focuses on constructing the global topological skeleton. This hierarchical sampling mechanism provides a hierarchical geometric information foundation for subsequent feature learning.

[0064] In this embodiment, the point cloud data is preprocessed to form the dataset required for training the model. Specifically, the ShapeNet 3D model library is used as the basic dataset, from which 16 categories of objects containing typical geometric structures such as furniture and vehicles are selected. A baseline point cloud is generated by uniformly sampling the complete 3D model, with a default sampling number of 2048 points. To simulate occlusion defects in real sensor data acquisition, a multi-view sphere culling strategy is implemented: the spatial center position of a sphere is randomly generated, and point clouds within a preset radius (default 0.3 units) are used as mask areas for deletion. The missing ratio is controlled between 25% and 75% by adjusting the number of spheres. All point cloud coordinates are linearly transformed and normalized to the [-1, 1] interval to eliminate the influence of model size differences. The preprocessed data is divided into training and test sets in an 80%:20% ratio.

[0065] S2. For the three-layer point cloud subsets, multi-layer multi-granularity features are extracted using an extended combined multi-layer perceptron. Each multi-granularity feature is input into a dual-path attention mechanism for processing. The weighted features of each layer after processing are independently max-pooled and then concatenated to form a three-layer joint feature vector corresponding to the three-layer point cloud subsets. Finally, the concatenation operation and the multi-layer perceptron are used to form the final feature vector.

[0066] As a preferred embodiment of step S2, the specific process includes the following steps:

[0067] S21. The three-layer point cloud subsets obtained in step S1 are input in parallel to the Extended Combined Multilayer Perceptron (E-CMLP). (See also...) Figure 2 The extended combined multilayer perceptron comprises a five-layer structure, each layer containing a set of fully connected layers, with the dimensions of each set of fully connected layers increasing progressively (64-dimensional / 128-dimensional / 256-dimensional / 512-dimensional / 1024-dimensional). This five-layer structure can combine low-dimensional geometric details with high-dimensional semantic information.

[0068] S22. The multi-granularity features extracted from each layer of the extended multilayer perceptron are processed by a dual-path attention mechanism. The dual-path attention mechanism includes two connected paths, namely the channel attention path and the spatial attention path.

[0069] S23, the channel attention path, such as Figure 3 As shown in the upper part of the process diagram, Channel weight matrix, multi-granularity features Channel descriptors are generated by compressing spatial information through adaptive global average pooling, and nonlinear channel relationships are learned through a fully connected layer with dynamic scaling. The scaling ratio is calculated using the following formula:

[0070] ;

[0071] in Indicates the scaling ratio. Given the number of input channels, this design can adaptively adjust the feature compression ratio. This represents the floor function, which applies the result calculated within the parentheses. The largest integer not greater than the calculated result is taken. Finally, the output channel weights are activated using a Sigmoid function. Strengthen the response of discriminative features such as edges and corners to achieve dynamic selection of key geometric features; Figure 3 Medium-grained features The width and height are respectively represented as and ;

[0072] S24, the spatial attention path, such as Figure 3 The following part of the process is shown in the diagram. The spatial weight matrix is ​​composed of channel weights. Processing multi-granularity features The resulting channel weighted features As input, average pooling is performed in parallel along the channel dimension. ) and max pooling ( The system captures region consistency features and local extremum features respectively, generating a bimodal spatial map. These features are then concatenated to generate bimodal features, which are then fused using a 7×7 large kernel convolution to output spatial weights. Precisely locate high curvature areas and structural fracture zones;

[0073] S25. Channel weights output from the two paths mentioned above. and spatial weights Multi-granularity features The process involves multiplying the feature maps element-wise to generate weighted features. To achieve synergistic enhancement:

[0074] ;

[0075] in This represents element-wise multiplication, with weighted features at all levels. After independent max pooling, the vectors are concatenated in ascending order of dimension to form a 1984-dimensional joint feature vector. The process can be found in [reference]. Figure 2 . Figure 3 Weighted features Features generated by applying max pooling Its width, height, and number of channels are respectively expressed as , , .

[0076] S26. The three-layer point cloud subsets generated in step S1 are used as inputs to the extended combined multilayer perceptron in the multi-scale feature extractor (MFE). The output joint feature vectors of the three layers are used as processed feature vectors, which are then concatenated to generate the final processed feature vector. Finally, the final feature vector is generated through an MLP layer. .

[0077] In this embodiment, the extended combined multilayer perceptron incorporates an innovative dual-path attention mechanism, achieving adaptive fusion of multi-scale features and solving the problem of geometric information attenuation in traditional methods.

[0078] S3. Perform hierarchical reconstruction using a pyramid point generator, resulting in the final feature vector. The low-dimensional, mid-dimensional, and high-dimensional features are used to generate principal center points, secondary center points, and detail points respectively, which are then stitched together to form a predicted point cloud of the missing region, accurately fitting the local geometric features; the hierarchical reconstruction process is as follows: Figure 4 As shown in the figure This represents the number of principal centroids at the coarsest level. This represents the number of secondary center points at the intermediate granularity level. This represents the number of detail points in the final output at the finest granular level.

[0079] As a preferred embodiment of step S3, the specific process includes the following steps:

[0080] S31, Final Feature Vector low-dimensional sub-eigenvectors (i.e., low-dimensional features) 512 principal centroids, or sparse point clouds, are generated through a single-layer fully connected network mapping, convolution, and reconstruction operations, forming the global skeleton of the missing region. For example, in a chair back completion task, the output of this layer forms the main framework structure, mathematically expressed as:

[0081] ;

[0082] in, Indicates the main center point. Indicates a fully connected operation. Corresponding final feature vector The front low-dimensional components (i.e., low-dimensional features) carry geometric skeleton information;

[0083] S32, Mid-dimensional sub-eigenvectors of the final eigenvector (i.e., mid-dimensional features) Input a three-layer MLP network (256 hidden units) and convolution and reconstruction operations to predict coordinate offsets. The coordinates are superimposed on the main center point coordinates, expanding to form 128 secondary points, i.e., secondary center points:

[0084] ;

[0085] in, This indicates the secondary center point. This step achieves the diffusion of the skeleton towards a medium-density topology, forming a medium-density topological structure. It precisely maintains the continuity of the openwork in the reconstruction of the chair back's hollowed-out boundary. (See also...) Figure 5 Qualitative analysis diagram of point cloud completion for input point clouds, points after completion, and real point clouds for various types of objects.

[0086] S33, Final Feature Vector High-dimensional sub-eigenvectors (i.e., high-dimensional features) predict micro-displacements through a deep MLP (512 hidden units) and convolution and reconstruction operations. 64 high-density surface points, i.e., detail points, are generated by expanding the neighborhood of the secondary point:

[0087] ;

[0088] in, This layer represents details. It can accurately fit local geometric changes, such as the circular profile of a chair leg cross-section or the axial curvature of a lamp holder.

[0089] S34. The main center point, secondary center point, and detail points are reconstructed and added in sequence to form the predicted point cloud of the missing region.

[0090] 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 topology skeleton.

[0091] In this embodiment, this progressive generation method of "skeleton-branch-skin" ensures geometric coherence from macro to micro.

[0092] S4. Design a composite loss function between the predicted point cloud and the real point cloud in the missing region, and design a dynamic weight scheduling strategy to simultaneously optimize multiple reconstruction objectives.

[0093] As a preferred embodiment of step S4, the design of the composite loss function specifically includes the following steps:

[0094] S41. The composite loss function includes nearest neighbor distance loss (CD) to constrain global shape alignment, extreme distance loss (HD) to suppress local geometric deviations, transmission distance loss (EMD) to optimize point distribution uniformity, and multi-stage constraint loss in the multi-stage constraint mechanism that forces skeleton topology to converge first.

[0095] S42. The nearest neighbor distance loss is to traverse each point in the predicted point cloud to find its nearest point in the real point cloud and accumulate the squared distance. At the same time, it is to traverse each point in the real point cloud to find its nearest point in the predicted point cloud and accumulate the squared distance. The sum of the two constitutes the global shape alignment constraint.

[0096] S43. The extreme distance loss is the maximum value of the maximum and minimum distance between the predicted point cloud and the real point cloud and the maximum and minimum distance between the real point cloud and the predicted point cloud, which is specifically used to suppress the deformation of thin-walled structures and sharp edges.

[0097] S44. The transmission distance loss is measured based on the optimal transmission theory to establish a point-to-point mapping relationship, effectively eliminating surface holes and point cluster artifacts.

[0098] In this embodiment, nearest neighbor distance loss ensures the alignment and coverage of the global shape; extreme distance loss specifically suppresses local geometric deviations and enhances the reconstruction accuracy of thin-walled structures and sharp edges; transmission distance loss optimizes the uniformity of point distribution based on optimal transmission theory; and multi-stage constraint loss obtains feature point sets by downsampling the real point cloud, forcing the primary / secondary center points to be hierarchically aligned with the corresponding feature point sets, ensuring the priority accuracy of the skeleton structure.

[0099] In this embodiment, a dynamic weight scheduling strategy is introduced to coordinate the conflicts of multi-objective optimization. Specifically, during the first 30 rounds of training, the nearest neighbor distance loss of 0.4 is given a dominant weight to accelerate global shape convergence. During the training from the 30th to the 80th round, the extreme distance loss weight is gradually increased to 0.3 to focus on optimizing local geometric accuracy. After the 80th round, the transmission distance loss weight is increased to 0.3 to improve the uniformity of point distribution.

[0100] S5. An adversarial training framework is adopted, the discriminator applies spectral normalization to constrain the spectral norm of the weight matrix, and the training stability is enhanced by a gradient penalty mechanism.

[0101] As a preferred embodiment of step S5, the specific process includes the following steps:

[0102] S51. The discriminator, in which all convolutional layer weights are spectral normalized, that is, the weight matrix is ​​divided by its largest singular value, and the Lipschitz constant is strictly constrained:

[0103] ;

[0104] in, Represents the weight matrix. yes The maximum singular value, It is the weight matrix after spectral normalization;

[0105] S52. The gradient penalty mechanism adopts the Wasserstein GAN with Gradient Penalty (WGAN-GP) framework, which performs linear interpolation sampling between the real point cloud and the predicted point cloud to generate interpolated sampling points. :

[0106] ;

[0107] in, Represents a true point cloud. Represents the predicted point cloud. These are the interpolation weight parameters. The discriminator is calculated at these interpolation sampling points. Define the gradient norm at the given point and define the gradient penalty term. Constrain the degree to which it deviates from 1:

[0108] ;

[0109] in, Indicates the interpolation sampling points We calculate the expectation of the distribution to ensure that the penalty term applies across the entire distribution. This represents the gradient of the discriminator. This indicates the calculation of the L2 norm.

[0110] In this embodiment, the discriminator and generator alternately update parameters at a ratio of 5:1, 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 after removing the dual-path attention mechanism in step S2; (c) is the point cloud completion result after removing the extended combined multilayer perceptron in step S2; (d) is the point cloud completion result after removing the gradient penalty mechanism in step S5; (e) is the point cloud completion result when the method is fully applied; and (f) is the real point cloud.

[0111] 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 stitched together with the input residual point cloud to output a complete reconstruction result.

[0112] In this embodiment, during actual deployment, the input residual point cloud is first processed by the IFPS algorithm to generate a subset of three resolutions (1024 points / 512 points / 256 points). These subsets are then input in parallel into the E-CMLP module to extract features weighted by a dual-path attention mechanism, thereby obtaining a joint feature vector. The joint feature vector is then input into a pyramid point generator, which sequentially generates 512 main center points (representing the global skeleton), 128 secondary center points (constructing topological branches), and 64 minutiae points (fitting surface geometry). Finally, these are merged to form the predicted point cloud of the missing region. By concatenating the predicted point cloud with the input residual point cloud, a complete reconstruction result can be output. The proposed method significantly improves the point cloud completion quality in complex structural scenes, providing reliable technical support for 3D vision applications.

[0113] This embodiment also proposes a high-fidelity point cloud completion system based on dual-path attention and fractal structure, used to apply a high-fidelity point cloud completion method based on dual-path attention and fractal structure, including:

[0114] The multi-scale feature extraction module performs iterative farthest point sampling and feature encoding;

[0115] The dual-path attention fusion module enables channel weighting and spatial saliency localization;

[0116] The pyramid point generation module outputs a predicted point cloud in three stages: main center point, secondary center point, and detail point.

[0117] The spectral normalization discriminator receives the predicted point cloud and outputs a truth score.

[0118] The hybrid loss calculation unit simultaneously optimizes global shape alignment constraints, local geometric deviation constraints, point distribution uniformity constraints, and skeleton topology-priority convergence constraints.

[0119] 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 topological relationship of the main center point and refines the mesoscopic structure, and a detail point synthesis unit that generates a surface point cloud in the neighborhood of the secondary center point that adapts to the local curvature.

[0120] Furthermore, the hybrid loss calculation unit adopts a dynamic weighting mechanism, which emphasizes global shape alignment constraints in the early stage of training, enhances local geometric deviation constraints in the middle stage of training, and increases the weight of point distribution uniformity constraints in the later stage of training.

[0121] In this embodiment, the proposed method and system demonstrate significant advantages in industrial scenarios: in the field of autonomous driving, it successfully repairs road defects caused by vehicle occlusion in LiDAR point clouds; in the industrial quality inspection process, it achieves precision control of tooth profile error within 0.3mm for the reconstruction of single missing teeth of gears with a module of 2; and in the field of cultural relic digitization, it accurately completes the decorative patterns at the broken handle in the scanning of bronze artifacts.

[0122] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0123] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can 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 processor-included system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0124] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A high-fidelity point cloud completion method based on dual-path attention and fractal structure, characterized in that, Includes the following steps: S1. Construct a dataset containing residual point clouds. Process the residual point clouds of the input model through an iterative farthest point sampling algorithm to generate a three-layer point cloud subset containing the original resolution, medium resolution, and low resolution, respectively preserving complete geometric details, main structural features, and global topological skeleton. S2. For each of the three-layer point cloud subsets, an extended combined multilayer perceptron is used to extract multilayer multigranular features. Each multilayer multigranular feature is input into a dual-path attention mechanism for processing. The weighted features of each layer after processing are independently max-pooled and then concatenated to form a three-layer joint feature vector corresponding to the three-layer point cloud subset. The final feature vector is then formed by concatenation and multilayer perceptron. The dual-path attention mechanism includes two concatenated paths, namely the channel attention path and the spatial attention path. 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 minutiae point respectively. Finally, the points are stitched together to form the predicted point cloud of the missing area and accurately fit the local geometric features. S4. Design a composite loss function between the predicted point cloud and the real point cloud in the missing region, and design a dynamic weight scheduling strategy to simultaneously optimize multiple reconstruction objectives. S41. The composite loss function includes nearest neighbor distance loss for constraining global shape alignment, extreme value distance loss for suppressing local geometric deviations, transmission distance loss for optimizing point distribution uniformity, and multi-stage constraint loss for forcing skeleton topology to converge preferentially in the multi-stage constraint mechanism. S42. The nearest neighbor distance loss is to traverse each point in the predicted point cloud to find its nearest point in the real point cloud and accumulate the squared distance. At the same time, it is to traverse each point in the real point cloud to find its nearest point in the predicted point cloud and accumulate the squared distance. The sum of the two constitutes the global shape alignment constraint. S43. The extreme distance loss is the maximum value of the maximum and minimum distance between the predicted point cloud and the real point cloud and the maximum and minimum distance between the real point cloud and the predicted point cloud, which is specifically used to suppress the deformation of thin-walled structures and sharp edges. S44. The transmission distance loss is measured based on the optimal transmission theory to establish a point-to-point mapping relationship, effectively eliminating surface holes and point cluster artifacts. S45. Specifically, the dynamic weight scheduling strategy assigns a dominant weight of 0.4 to the nearest neighbor distance loss during the first 30 rounds of training to accelerate global shape convergence; during the 30 to 80 rounds of training, the extreme distance loss weight is gradually increased to 0.3 to focus on optimizing local geometric accuracy; after 80 rounds, the transmission distance loss weight is increased to 0.3 to focus on improving the uniformity of point distribution. S5. An adversarial training framework is adopted, the discriminator applies spectral normalization to constrain the spectral norm of the weight matrix, and the training stability is enhanced by a 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 stitched together 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: Step S2 specifically includes the following steps: S21. The extended combined multilayer perceptron includes a five-layer structure, each layer containing a set of fully connected layers, with the dimensions of each set of fully connected layers increasing in sequence. This five-layer structure can combine low-dimensional geometric details with high-dimensional semantic information. S22, Extended combined multilayer perceptron extracts multi-granularity features from each layer Input is processed using a dual-path attention mechanism; S23, the channel attention path, multi-granularity features Channel descriptors are generated through adaptive global average pooling, nonlinear channel relationships are learned through dynamically scaled fully connected layers, and finally, the output channel weights are activated by Sigmoid. This enables dynamic selection of key geometric features; S24, the spatial attention path, is determined by channel weights. Processing multi-granularity features The resulting channel weighted features As input, average pooling and max pooling are performed in parallel along the channel dimension to generate a bimodal spatial graph. After concatenation and fusion, the spatial weights are output. Precisely locate spatial saliency; S25. Channel weights output from the two paths mentioned above. and spatial weights Multi-granularity features The process involves multiplying the feature maps element-wise to generate weighted features. To achieve synergistic enhancement: ； in This represents element-wise multiplication, with weighted features at all levels. After independent max pooling, the vectors 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 inputs to the extended combined multilayer perceptron. The joint feature vectors of the three layers are concatenated and then passed through an MLP layer to generate the final feature vector. .

3. The high-fidelity point cloud completion method based on dual-path attention and fractal structure according to claim 1, characterized in that: Step S3 specifically includes the following steps: S31. The main center point is generated by mapping low-dimensional features through a fully connected layer, serving as the global skeleton; S32. The secondary center points are generated by diffusing from the main center points through a three-layer MLP network based on the mid-dimensional features, forming a medium-density topology. S33. The detailed points are synthesized from 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, secondary center point, and detail points are sequentially pieced together to form the 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 topology skeleton.

5. The high-fidelity point cloud completion method based on dual-path attention and fractal structure according to claim 1, characterized in that: Step S5 specifically includes the following steps: S51. The discriminator has all its convolutional layer weights processed by spectral normalization. S52. The gradient penalty mechanism calculates the gradient penalty term by linear interpolation between the predicted point cloud and the real point cloud. The discriminator and the generator alternately update parameters using a preset ratio to maintain training stability.

6. A high-fidelity point cloud completion system based on dual-path attention and fractal structure, used to apply the high-fidelity point cloud completion method based on dual-path attention and fractal structure as described in any one of claims 1-5, characterized in that, include: The multi-scale feature extraction module performs iterative farthest point sampling and feature encoding; The dual-path attention fusion module enables channel weighting and spatial saliency localization; The pyramid point generation module outputs a 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 truth score. The hybrid loss calculation unit simultaneously optimizes global shape alignment constraints, local geometric deviation constraints, point distribution uniformity constraints, and skeleton topology-priority convergence constraints.

7. A high-fidelity point cloud completion system based on dual-path attention and fractal structure according to claim 6, 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 topological relationship of the main center point and refines the mesoscopic structure, and a detail point synthesis unit that generates surface point clouds that adapt to local curvature in the neighborhood of the secondary center point.

8. A high-fidelity point cloud completion system based on dual-path attention and fractal structure according to claim 6, characterized in that, The hybrid loss calculation unit adopts a dynamic weighting mechanism, which focuses on global shape alignment constraints in the early stage of training, enhances local geometric deviation constraints in the middle stage of training, and increases the weight of point distribution uniformity constraints in the later stage of training.

Images