Point cloud surface reconstruction method and device fusing geometric radio and gated attention

Abstract

The application discloses a point cloud surface reconstruction method and device fusing geometric radiation and gated attention. The method acquires a local block point set with a query point as a center and a global subsampling point set, and performs translation alignment on the global subsampling point set; coordinates of the local block point set are first subjected to adaptive normalization and a learnable affine transformation by a geometric affine modulation module, input into a point-by-point feature extraction network, and a local point-by-point feature matrix of a reserved point dimension is obtained; then the local point-by-point feature matrix is subjected to another point-by-point feature extraction network and symmetric pooling to obtain a global feature vector; after dimension reduction and projection, the two are spliced into a fusion feature matrix; an aggregated feature vector is obtained through a gated attention aggregation decoding module, and then the query point is output through a fully connected decoding network. The application improves the consistency and generalization ability of local feature expression through geometric affine modulation; the precision, detail fidelity and robustness of implicit surface reconstruction are improved by replacing symmetric pooling and simple splicing with gated attention aggregation.

Description

Technical Field

[0001] This invention relates to the field of intelligent transformation of CNC machining programs in intelligent manufacturing technology, and in particular to a method and apparatus for implicit surface reconstruction of point clouds that integrates geometric affine modulation and gated attention aggregation. Background Technology

[0002] Existing implicit surface reconstruction methods for point clouds typically decompose the input point cloud into two parts: a local block point set around a query point and a global subsampled point set. These are encoded by two identical point-by-point feature extraction networks: the local branch learns the absolute distance information at the query point, and the global branch learns the sign information of the point relative to the closed surface. After extracting point-by-point features, both branches are compressed into a single feature vector using symmetric max pooling. The pooling vectors from the two branches are then directly concatenated and input into a multi-layer fully connected network for decoding to obtain the signed distance value.

[0003] The aforementioned technological status quo has the following shortcomings:

[0004] (1) Both local and global branches perform symmetric pooling immediately after obtaining point-by-point features. This operation treats the channel features of each point in the point set equally in terms of numerical value, and fails to distinguish the differences in contribution of different points to the SDF prediction of the query point. This results in the model's limited detail fidelity and robustness on noisy, sparse or non-uniformly sampled point clouds.

[0005] (2) The scale and density distribution of the local block point set input by the local branch varies greatly on different objects and different local regions. If it is directly fed into the feature extraction network without adaptive normalization modulation, it is easy to cause inconsistent feature distribution between different local blocks, which weakens the cross-class generalization ability of the model.

[0006] (3) In the feature fusion stage, the local vector after pooling is directly concatenated with the global vector. The decoding network has difficulty in adaptively adjusting the contribution weights of different feature channels or different spatial locations according to the importance of the local structure, which limits the ability to reconstruct fine-grained geometric details.

[0007] Therefore, there is an urgent need for a new method that can improve the accuracy and generalization ability of implicit surface reconstruction of point clouds by using adaptive geometric normalization and attention-weighted aggregation while preserving point-level feature distribution information. (Summary of the Invention)

[0008] To address the problems of inconsistent local branch feature distribution and loss of point-level importance information in existing point cloud implicit surface reconstruction methods, this invention provides a point cloud implicit surface reconstruction method and apparatus that integrates geometric affine modulation and gated attention aggregation. Geometric affine modulation improves the consistency of local feature representation, and adaptive attention aggregation replaces symmetric pooling and simple stitching, thereby improving the accuracy, detail fidelity, and cross-class generalization ability of implicit surface reconstruction.

[0009] The first aspect of this invention provides a method for implicit surface reconstruction of point clouds, the method comprising the following steps:

[0010] Step S1: Obtain a local block point set centered on the query point x, the local block point set consisting of the n points in the point cloud closest to x; simultaneously obtain a global sub-sample point set obtained by sampling from the overall point cloud according to the density gradient decreasing from the distance to x; subtract the query point coordinates from the coordinates of each point in the global sub-sample point set to achieve coordinate translation and alignment with the query point as the origin;

[0011] Step S2: Input the coordinate features of the local block point set into the geometric affine modulation module (GAM) for adaptive normalization and learnable affine transformation to obtain the modulated coordinate features; input the modulated coordinate features into the local branch point-by-point feature extraction network to obtain the local point-by-point feature matrix that retains the point dimension.

[0012] Step S3: Input the global subsample point set after translation and alignment into the global branch point-by-point feature extraction network to obtain global point-by-point features. Then, perform symmetric pooling along the point dimension to obtain the global feature vector. The global branch point-by-point feature extraction network has the same structure as the local branch point-by-point feature extraction network but the parameters are independent, and it does not contain the geometric affine modulation module (GAM).

[0013] Step S4: Project and reduce the dimensionality of the local pointwise feature matrix and the global feature vector respectively; copy and expand the dimensionality-reduced global feature vector along the point dimension so that its number of points is consistent with the dimensionality-reduced local feature matrix; concatenate the expanded global feature matrix and the dimensionality-reduced local feature matrix along the channel dimension to obtain the fused feature matrix;

[0014] Step S5: Input the fused feature matrix into the gated attention aggregation decoding module, and sequentially execute gated weight generation, feature modulation, query feature and numerical feature mapping, attention weight calculation and weighted aggregation operation to obtain the aggregated feature vector;

[0015] S6: Input the aggregated feature vector into the fully connected decoding network and output the implicit surface prediction value corresponding to the query point. The implicit surface prediction value includes the absolute distance prediction value and the symbolic Logits prediction value.

[0016] Furthermore, the execution process of the geometric affine modulation module (GAM) includes: calculating the mean of each channel along the point dimension and performing mean-removal processing; calculating the joint standard deviation of the mean-removed features in the channel dimension and the point dimension, and normalizing the mean-removed features using the joint standard deviation; and performing an affine transformation on the normalization result using independent learnable scaling coefficients and bias coefficients for each channel.

[0017] Furthermore, the execution process of the gated attention aggregation decoding module includes: transforming the fused feature matrix using one-dimensional convolution and then activating it with Sigmoid to generate a gated weight matrix; multiplying the gated weight matrix element-wise with the fused feature matrix to obtain the modulation feature; reducing the channel dimension of the modulation feature to the target dimension using two parameter-independent one-dimensional convolutions to generate a query feature matrix and a numerical feature matrix; performing Softmax normalization on the query feature matrix along the point dimension to obtain the attention weight matrix, and taking the maximum value along the channel dimension to obtain the importance weight of each point; and performing a weighted summation of the numerical feature matrix and the importance weights to obtain the aggregated feature vector.

[0018] Furthermore, for the input local point cloud feature group ,in This represents the number of neighboring points within the local neighborhood. For each point's feature dimension, Indicates the first The first center point Feature vectors of neighboring points:

[0019] Channel-by-channel mean calculation: For each feature channel Calculate the mean along the dimension of neighboring points. , Representing the eigenvector In the The values ​​on each channel For the first The average of each channel;

[0020] Feature decentralization: computation , These are the decentralized feature values;

[0021] Calculation of combined standard deviation: ;

[0022] Global normalization: ,in It is a very small constant;

[0023] Affine transformations can be learned: ,in , For channel-by-channel learnable parameters.

[0024] Furthermore, the output of the gated attention aggregation decoding module in step S5 is the implicit surface parameter prediction vector at the query point x. ,in This is the predicted value of absolute distance. The predicted value is the symbolic Logits.

[0025] The fully connected decoding network includes: a first fully connected layer that aggregates feature vectors. From the perspective of dimensions The data is compressed to the intermediate dimension D of the decoding process, then batch normalized and ReLU activated to obtain the hidden layer feature vector h; the second fully connected layer then... Mapping to the two-dimensional output space yields the prediction vector y.

[0026] Furthermore, the method also includes a training phase:

[0027] Repeat steps S1 to S6 for multiple point cloud-surface sample pairs and their query points in the training set, employing a multi-task joint learning strategy, with the total loss function... ,in For based on Norm-based distance regression loss, The binary cross-entropy sign classification loss is used. To balance the weight coefficients, the backpropagation algorithm is used to update all trainable parameters end-to-end until convergence.

[0028] It also includes an inference phase: repeatedly executing steps S1 to S6 for each sampling point in the voxel grid to estimate the signed distance field, and performing a synthesis operation. The signed distance estimate is calculated and the zero isosurface is extracted by applying the moving cube algorithm to obtain the final reconstructed 3D surface mesh.

[0029] Furthermore, the local branch point-by-point feature extraction network described in step S2 consists of multiple layers of one-dimensional convolution, batch normalization, and ReLU activation function. Each layer of one-dimensional convolution is followed by batch normalization and ReLU activation to achieve progressive dimensionality increase and abstraction of the feature dimension.

[0030] A second aspect of the present invention provides a point cloud implicit surface reconstruction device, characterized in that it comprises:

[0031] The point set acquisition and alignment module is used to acquire a local block point set and a global subsample point set centered on the query point x, and to perform coordinate translation and alignment of the global subsample point set with the query point as the origin.

[0032] The geometric affine modulation module (GAM) is used to adaptively normalize and learnable affine transformation the coordinate features of the local block point set, including: a channel-wise mean calculation and mean removal unit, a cross-channel and cross-point joint standard deviation calculation unit, a global normalization unit, and a channel-wise learnable affine transformation unit.

[0033] The local branch encoder, consisting of multiple layers of one-dimensional convolution, batch normalization, and ReLU activation, is used to receive modulated coordinate features and output a local point-by-point feature matrix that retains the point dimension.

[0034] The global branch encoder has the same structure as the local branch encoder but with independent parameters and does not contain the geometric affine modulation module (GAM). It is used to receive the global sub-sample point set after translation and alignment and output the global feature vector after symmetric pooling.

[0035] The feature fusion module is used to project and reduce the dimensionality of the local pointwise feature matrix and the global feature vector respectively, and then concatenate them to obtain a fused feature matrix.

[0036] The gated attention aggregation decoding module includes: a gated weight generation unit, which generates a gated weight matrix using one-dimensional convolution and sigmoid activation; a feature modulation unit, which multiplies the gated weights and fused features element-wise; a query / numerical mapping unit, which generates a query feature matrix and a numerical feature matrix using two sets of independent one-dimensional convolutions; an attention weight calculation unit, which performs Softmax normalization along the point dimension and takes the maximum value along the channel dimension to obtain the importance weight; and a weighted aggregation unit, which sums the numerical features and importance weights to obtain the aggregated feature vector.

[0037] A fully connected decoding network is used to map the aggregated feature vector to the implicit surface prediction value corresponding to the query point, and the output includes the absolute distance prediction value and the symbolic Logits prediction value.

[0038] Furthermore, the device also includes a training phase module:

[0039] Multiple point cloud-surface sample pairs and their query points in the training set are sequentially input into the point set acquisition and alignment module, the geometric affine modulation module (GAM), the local branch encoder, the global branch encoder, the feature fusion module, the gated attention aggregation decoding module, and the fully connected decoding network. A multi-task joint learning strategy is adopted, and the total loss function is... ,in For based on Norm-based distance regression loss, The binary cross-entropy sign classification loss is used. To balance the weight coefficients, the backpropagation algorithm is used to update all trainable parameters end-to-end until convergence.

[0040] The device further includes an inference module: a module for acquiring and aligning repeated input point sets for each sampling point in the voxel grid, a geometric affine modulation module (GAM), a local branch encoder, a global branch encoder, a feature fusion module, a gated attention aggregation decoding module, and a fully connected decoding network to estimate the signed distance field, and then synthesizes it. The signed distance estimate is calculated and the zero isosurface is extracted by applying the moving cube algorithm to obtain the final reconstructed 3D surface mesh.

[0041] A third aspect of the present invention provides a computer device, comprising:

[0042] processor;

[0043] Memory for storing computer programs executable by the processor;

[0044] The processor executes the computer program to implement the steps of the above-described point cloud implicit surface reconstruction method.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. By using the Geometric Affine Modulation (GAM) module, adaptive mean-variance normalization and learnable affine transformation are performed on the local block point cloud coordinates before feature extraction. This alleviates the feature distribution shift caused by scale and density differences between different local blocks, enhances the consistency and discriminative power of local features, and improves the model's generalization ability across different object categories.

[0047] 2. Local branches retain point-by-point features without global pooling, enabling subsequent aggregation modules to process based on point-level spatial distribution information. Compared to direct symmetric pooling, this retains richer geometric structure information, providing the necessary information foundation for gated attention aggregation.

[0048] 3. The gated attention aggregation decoding module is used to replace symmetric pooling and feature concatenation. This enables the network to adaptively assign different aggregation weights to different points according to the content of the fused features. This enhances the sensitivity to local geometric details (such as holes, depressions, thin-walled structures, etc.), improves reconstruction accuracy and detail fidelity, and shows better robustness, especially under noisy and sparse sampling conditions. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1This is a schematic diagram of the overall structure of the point cloud implicit surface reconstruction network provided in an embodiment of the present invention;

[0051] Figure 2 This is a flowchart of the GAM calculation process of the geometric affine modulation module in an embodiment of the present invention;

[0052] Figure 3 This is a logic diagram of the gated attention aggregation decoding module in an embodiment of the present invention. Detailed Implementation

[0053] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are used to illustrate the present invention and are not intended to limit the scope of protection of the present invention.

[0054] Example 1: Specific Implementation of the Point Cloud Implicit Surface Reconstruction Method

[0055] This embodiment provides a detailed description of the internal structure, computational process, and embedding method of the Geometric Affine Modulation (GAM) module in the feature extraction network. This module, as a plug-and-play coordinate preprocessing unit, is deployed at the front end of the local branch feature extraction network. Its purpose is to adaptively correct the distribution of local point features in the input, eliminate translation, scale, and density differences, and enhance the model's ability to perceive geometric details.

[0056] Input: A local point cloud feature set. in, This represents the number of neighboring points within the local neighborhood. Feature dimensions for each point (if the input is three-dimensional coordinates) ).here Indicates the first The first center point The feature vectors of the neighboring points.

[0057] Output: Features after geometric affine modulation The output dimension strictly matches the input dimension, achieving dimension order preservation without requiring changes to the subsequent network structure.

[0058] Step 1: Calculate the mean of each channel and perform local centering.

[0059] First, for each feature channel in the local feature group, along the dimension of neighboring points... Calculate the mean to obtain the local feature center. Subsequently, the central value is subtracted from the features of each neighboring point to eliminate translation redundancy in local point blocks.

[0060] Formula (1): Calculation of the mean of local features:

[0061] ;

[0062] in, Representing the eigenvector In the The values ​​on each channel For the first The average value of each channel.

[0063] Formula (2): Feature decentralization:

[0064] ;

[0065] in, These are the decentralized feature values. This step ensures that feature learning focuses on the relative positions of local geometry, rather than absolute coordinates.

[0066] Step 2: Calculate the combined standard deviation across channels and across points.

[0067] Combining feature dimensions and neighborhood points Calculate the standard deviation of a global scalar. This statistic reflects the overall dispersion of the current local sample, providing a unified scale benchmark for normalization.

[0068] Formula (3): Calculation of joint standard deviation:

[0069] ;

[0070] in, It is a scalar. This cross-dimensional joint computation method (corresponding to the operation dim=[-1, -2] in the code) makes the model more robust to scale changes in local regions.

[0071] Step 3: Global normalization operation.

[0072] Using the global standard deviation obtained in step 2, the decentralized features are normalized to balance the differences in feature distribution in different local regions.

[0073] Formula (4): Global normalization:

[0074] ;

[0075] in, For example, a very small constant (e.g.) (), used for numerical stability to prevent the denominator from being zero; These are the normalized eigenvalues.

[0076] Step 4: Channels can learn affine transformations.

[0077] Introducing learnable scaling parameters With bias parameters The normalized features are linearly modulated channel by channel. This allows the network to adaptively adjust the distribution of different feature channels, restoring or enhancing the expressive power of key geometric features.

[0078] Formula (5): Learnable affine transformation:

[0079] ;

[0080] During initialization, Set to 1, Set it to 0 to ensure the network's initial state is stable.

[0081] Local point cloud feature groups processed by the GAM module ,in This represents the number of points within the local neighborhood. The feature dimensions of each point are used as input and fed into the local branch encoder. The encoder adopts a multilayer perceptron (MLP) structure, and the specific implementation steps are as follows:

[0082] 1. Point-by-point feature mapping: The features of each point in the local point cloud are nonlinearly mapped layer by layer through a multilayer perceptron (MLP). Each MLP layer is followed by batch normalization (BN) and ReLU activation function to achieve progressive dimensionality increase and abstraction of features, so as to extract deep feature representations of local geometric details.

[0083] 2. Feature Aggregation: For the local point cloud features after point-by-point feature mapping, perform channel max pooling along the point dimension to aggregate the point features in the local neighborhood into a local feature vector. ,in This is the encoded feature dimension, and the vector represents the geometric feature distribution of the local region.

[0084] Global subsampled point cloud (where n is the number of global subsampling points and d is the feature dimension of each point) is used as input and fed into the global branch encoder. This encoder has the same structure as the local branch encoder but with independent parameters. The specific implementation steps are as follows:

[0085] 1. Point-by-point feature mapping: The multilayer perceptron (MLP) structure, consistent with the local branch encoder structure, is adopted to perform layer-by-layer nonlinear mapping on the features of each point in the global subsampled point cloud. Each MLP layer is followed by batch normalization (BN) and ReLU activation function to realize the progressive dimensionality increase and abstraction of global features, so as to extract the coarse-grained global structural features of the object.

[0086] 2. Feature Aggregation: For the global point cloud features after point-by-point feature mapping, perform channel max pooling along the point dimension to output a global feature vector. This vector and the local branch Dimensionality is consistent, representing the global semantic information of an object.

[0087] After obtaining the feature representations of the local and global branches, the following fusion steps are performed:

[0088] Local feature dimensionality reduction: dimensionality reduction of the feature vectors output from local branches. Convolutional operations are used to reduce the channel dimension and output a local feature matrix. ,in The feature dimension is the dimension after dimensionality reduction.

[0089] Global Feature Dimensionality Reduction and Expansion: Global Feature Vectors Output by the Global Branch (in (The encoder outputs a feature dimension), which is then reduced in dimension by a fully connected layer and expanded and broadcast along the edge point dimension to directly obtain a global feature matrix with the same dimension as the local feature matrix. .

[0090] Feature concatenation: combining the dimensionality-reduced local feature matrix With the expanded global feature matrix The features are concatenated along the channel dimension to obtain the fused feature matrix. This matrix is ​​then fed as input into the subsequent SHS attention module for further feature aggregation and prediction.

[0091] This embodiment details the specific structure, operational logic, and implementation process of the gated attention aggregation decoding module, which replaces the traditional symmetric pooling concatenation decoding scheme. This module introduces a gated adaptive modulation mechanism, abandoning the traditional paradigm of fixed symmetric pooling and direct feature concatenation in decoding networks, to achieve adaptive weighted aggregation and high-precision decoding of point features. The module sets intermediate dimension parameters. This is equivalent to a dimensional benchmark for unified feature mapping and aggregation.

[0092] Input: fused feature matrix ,in This represents the number of points within the local neighborhood. This represents the channel dimension after concatenation. This matrix integrates local geometric details and prior global structural features.

[0093] Output: Query point Implicit surface parameter prediction vector at the location The vector contains two predictions: the absolute distance prediction. : Indicates the query point Unsigned distance to the implicit surface. Signed Logits prediction value. : Indicates the query point The logarithmic probability of being located outside (positive) or inside (negative) a surface.

[0094] Step 1: Gated modulation and feature preprocessing.

[0095] This step adaptively modulates the fused features using a gating mechanism and generates the query and numerical features required for attention computation. One-dimensional convolution is employed. right A linear transformation is performed, and the result is activated by a Sigmoid function to generate a gated weight matrix. This is used for adaptive selection of feature channels and sampling points. Then, the gating weights are... and Element-wise multiplication yields the modulation characteristics. This enhances the response of key features and suppresses redundant information.

[0096] Formula (8): Generation of gating weights.

[0097] ;

[0098] in, This is the Sigmoid activation function.

[0099] Formula (9): Adaptive feature modulation.

[0100] ;

[0101] in, This is for the Hadamard product operation.

[0102] Finally, two independent sets of one-dimensional convolution operators for feature maps are set. , Modulation features The channel dimension is determined by Down to Generate query feature matrix With numerical characteristic matrix .in, Model the association weights of sampling points. Retain valid feature information.

[0103] Formula (10): Query and numerical feature mapping.

[0104] , ;

[0105] Step 2: Attention weight calculation and aggregation.

[0106] This step calculates the importance weight of each sampling point through an attention mechanism, realizes adaptive weighting of features, and generates the final feature representation through weighted aggregation.

[0107] First, define the query features. Perform Softmax normalization along the point dimension to obtain the channel attention weight matrix. Then, take the maximum value along the channel dimension to obtain the single-point importance weight. The contribution of each sampling point to the surface reconstruction is quantified.

[0108] Formula (12): Solving for attention weights.

[0109] ;

[0110] ;

[0111] Next, the numerical feature matrix single-point importance weight Multiplying by the transposes yields the aggregated feature vector. This enables differentiated weighted aggregation.

[0112] Formula (14): Adaptive weighted aggregation.

[0113] ;

[0114] Step 3: Implicit surface prediction.

[0115] This step will aggregate feature vectors. The input is processed by a multilayer perceptron (MLP) decoder through a nonlinear mapping, ultimately outputting the absolute distance and the sign logits. This decoder consists of two fully connected layers, as detailed below:

[0116] The first step is through a fully connected layer. Batch Normalized Normalization (BN) and ReLU activation functions reduce the feature dimension from... Compression to the intermediate dimension of decoding The hidden layer feature vector is obtained. Extract deep geometric semantics.

[0117] Formula (15):

[0118] ;

[0119] The second step involves using a second-layer fully connected layer. Hidden layer features Mapping to the output space yields the prediction vector. .in, The first component is the query point. absolute distance prediction value The second component is the predicted value of the symbolic Logits. .

[0120] Formula (16):

[0121] ;

[0122] Finally, the probability of the query point being located outside the surface is calculated using the Sigmoid function based on the symbolic Logits. Combined with the absolute distance, the complete signed distance field (SDF) can be reconstructed, which is used to extract the final 3D surface mesh.

[0123] This embodiment details the complete training and inference process based on the aforementioned network architecture, focusing on how to synthesize the absolute distance and symbolic Logits output by the network into the final implicit symbolic distance function (SDF) and efficiently extract the 3D surface mesh.

[0124] During network training, for each surface sample in the point cloud dataset, several query points are sampled, including points sampled on the surface and offset along the normal, and points randomly sampled within the bounding box. For each query point, a local block point set and a global sub-sampled point set are constructed from its respective point cloud. The predicted implicit surface value is calculated according to the aforementioned steps S1 to S6, and the query point is finally obtained. Predicted absolute distance at location Binary cross-entropy loss Logits for symbol classification .

[0125] Training employs a multi-task joint learning strategy, with a total loss function. It is composed of a weighted average of distance regression loss and symbol classification loss, and its mathematical expression is shown in formula (17):

[0126] Formula (17) – Joint training loss:

[0127]

[0128] in, For based on Norm-based distance regression loss (e.g., applied to predicted and true distances) The function is weighted to focus on the region near the surface, which is used to constrain the numerical accuracy of the predicted absolute distance; The binary cross-entropy symbol classification loss is used to constrain the correctness of symbol prediction; To balance the weighting coefficients contributing to the two types of losses, the total loss gradient is propagated throughout the entire network using backpropagation, updating all trainable parameters end-to-end until the model converges.

[0129] During grid inference, the signed distance estimate at the grid point is calculated through a synthesis operation. : This involves multiplying the absolute distance by a sign function of the sign function Logits, thus obtaining the complete SDF value with internal and external information. At this point, the entire voxel mesh is endowed with a complete signed distance field.

[0130] After obtaining the complete signed distance field (or the filled TSDF), the classic Moving Cubes algorithm is applied to traverse all voxel units, and isosurfaces with algebraic values ​​of zero are extracted through linear interpolation. This isosurface is the final reconstructed, watertight 3D surface mesh, which can be directly used for downstream rendering, analysis, and editing tasks.

[0131] Example 2: Specific Implementation of the Point Cloud Implicit Surface Reconstruction Device

[0132] This embodiment corresponds to the second aspect of the present invention, and describes in detail the specific structure of a point cloud implicit surface reconstruction device. This device is a hardware / software functional modular implementation of the method described in Embodiment 1 above.

[0133] The device includes the following modules:

[0134] (1) Point set acquisition and alignment module: configured to receive input point cloud and query point, acquire local block point set centered on query point x (composed of the k points closest to x) and global subsample point set (sampled according to the density gradient of decreasing distance), and translate and align by subtracting the query point coordinates from the coordinates of the global subsample point set.

[0135] (2) Geometric Affine Modulation Module (GAM): Configured for adaptive normalization and learnable affine transformation of the coordinate features of local block point sets. This module includes: a channel-wise mean calculation unit for calculating the mean of each channel and performing centering; a joint standard deviation calculation unit for calculating the scalar standard deviation across channels and points; a global normalization unit for performing normalization using this standard deviation; and a learnable affine transformation unit containing channel-wise learnable parameters α and β for performing affine transformations. The output dimension is strictly consistent with the input dimension.

[0136] (3) Local branch encoder: It consists of multi-layer one-dimensional convolution, batch normalization and ReLU activation. It receives the features after GAM modulation and outputs a local point-by-point feature matrix that retains the point dimension.

[0137] (4) Global branch encoder: The structure is the same as the local branch encoder but the parameters are independent and there is no GAM module. It receives the global sub-sample point set after translation and alignment, and outputs the global feature vector after point-by-point feature extraction and symmetric pooling.

[0138] (5) Feature fusion module: includes local dimensionality reduction unit (one-dimensional convolution), global dimensionality reduction and broadcast expansion unit (fully connected layer + copy expansion) and channel concatenation unit, outputting a fused feature matrix. .

[0139] (6) Gated attention aggregation decoding module (SHS): includes a gated weight generation unit (Conv+Sigmoid), a feature modulation unit (Hadamard product), a query / numerical mapping unit (two independent Conv), an attention weight calculation unit (Softmax+channel maximum value) and a weighted aggregation unit (matrix multiplication+summation), which outputs an aggregated feature vector.

[0140] (7) Fully connected decoding network: It contains two fully connected layers. The first layer compresses the aggregated features to the intermediate dimension of the decoding and then activates them through BN and ReLU. The second layer maps to the two-dimensional output space and outputs the absolute distance prediction value and the symbolic Logits prediction value.

[0141] The modules described above can be implemented in software and run on general-purpose computing devices (such as GPU servers), or they can be implemented as dedicated acceleration devices using firmware or hardware circuits (such as FPGAs and ASICs). The data flow and cooperation relationships between the modules are consistent with the description of the corresponding steps in Example 1.

[0142] Example 3: Specific Implementation of Computer Device

[0143] This embodiment corresponds to the third aspect of the present invention, and describes in detail a computer device for performing the aforementioned point cloud implicit surface reconstruction method. The computer device includes a processor (such as a CPU, GPU, TPU, etc.) and a memory (such as DDR memory, SSD, hard disk, etc.). The memory stores a computer program containing instruction code for performing steps S1 to S6 as described in Embodiment 1, specifically including:

[0144] Point set construction and alignment instructions: used to select local block point sets and global subsampled point sets from the input point cloud and perform coordinate translation and alignment;

[0145] GAM modulation instructions: used to perform channel-by-channel mean calculation, joint standard deviation normalization, and learnable affine transformation;

[0146] Dual-branch encoding instructions: used to perform point-by-point feature extraction in local branches (preserving point dimensions) and global branches (symmetric pooling) respectively;

[0147] Feature fusion instructions: used to perform dimensionality reduction, broadcast expansion, and channel splicing;

[0148] Gated attention aggregation instructions: used to perform gating generation, feature modulation, query / numerical mapping, attention weight calculation and weighted aggregation;

[0149] Decoding output instructions: used to perform forward inference of a fully connected decoding network;

[0150] Training instructions: Used to perform multi-task joint loss calculation and backpropagation parameter updates;

[0151] Inference and Mesh Extraction Commands: Used to estimate the SDF field on a voxel mesh and invoke the moving cube algorithm to extract isosurfaces.

[0152] The computer device can be embodied as a workstation, server, embedded computing platform, or mobile computing device, as long as it has sufficient computing power and storage capacity to support forward inference and / or training of neural networks.

[0153] This invention has many specific applications. The above description is only a preferred embodiment of this invention. It should be noted that for those skilled in the art, several improvements can be made without departing from the principle of this invention, and these improvements should also be considered within the scope of protection of this invention.

Claims

1. A point cloud surface reconstruction method integrating geometrical radiality and gated attention, characterized in that, The method includes the following steps: Step S1: Obtain a local block point set centered on the query point x, the local block point set consisting of the n points in the point cloud closest to x; simultaneously obtain a global sub-sample point set obtained by sampling from the overall point cloud according to the density gradient decreasing from the distance to x; subtract the query point coordinates from the coordinates of each point in the global sub-sample point set to achieve coordinate translation and alignment with the query point as the origin; Step S2: Input the coordinate features of the local block point set into the geometric affine modulation module (GAM) for adaptive normalization and learnable affine transformation to obtain the modulated coordinate features; The modulated coordinate features are input into the local branch point-by-point feature extraction network to obtain a local point-by-point feature matrix that retains the point dimension; Step S3: Input the global subsample point set after translation and alignment into the global branch point-by-point feature extraction network to obtain global point-by-point features. Then, perform symmetric pooling along the point dimension to obtain the global feature vector. The global branch point-by-point feature extraction network has the same structure as the local branch point-by-point feature extraction network but the parameters are independent, and it does not contain the geometric affine modulation module (GAM). Step S4: Project and reduce the dimensionality of the local pointwise feature matrix and the global feature vector respectively; copy and expand the dimensionality-reduced global feature vector along the point dimension so that its number of points is consistent with the dimensionality-reduced local feature matrix; concatenate the expanded global feature matrix and the dimensionality-reduced local feature matrix along the channel dimension to obtain the fused feature matrix; Step S5: Input the fused feature matrix into the gated attention aggregation decoding module, and sequentially execute gated weight generation, feature modulation, query feature and numerical feature mapping, attention weight calculation and weighted aggregation operation to obtain the aggregated feature vector; S6: Input the aggregated feature vector into the fully connected decoding network and output the implicit surface prediction value corresponding to the query point. The implicit surface prediction value includes the absolute distance prediction value and the symbolic Logits prediction value.

2. The point cloud surface reconstruction method fusing geometrical radiality and gated attention as described in claim 1, characterized in that, The execution process of the geometric affine modulation module (GAM) includes: calculating the mean of each channel along the point dimension and performing mean-removal processing; calculating the joint standard deviation of the mean-removed features in the channel dimension and the point dimension, and normalizing the mean-removed features using the joint standard deviation; and performing an affine transformation on the normalization result using independent learnable scaling coefficients and bias coefficients for each channel.

3. The point cloud surface reconstruction method fusing geometrical radiality and gated attention as described in claim 1, characterized in that, The execution process of the gated attention aggregation decoding module includes: transforming the fused feature matrix using one-dimensional convolution and then activating it with Sigmoid to generate a gated weight matrix; multiplying the gated weight matrix and the fused feature matrix element by element to obtain the modulation feature; reducing the channel dimension of the modulation feature to the target dimension using two parameter-independent one-dimensional convolutions to generate a query feature matrix and a numerical feature matrix; performing Softmax normalization on the query feature matrix along the point dimension to obtain the attention weight matrix, and taking the maximum value along the channel dimension to obtain the importance weight of each point; and weighting and summing the numerical feature matrix with the importance weights to obtain the aggregated feature vector.

4. The point cloud surface reconstruction method fusing geometrical radiality and gated attention as described in claim 1, characterized in that, For the input local point cloud feature group ,in This represents the number of neighboring points within the local neighborhood. For each point's feature dimension, Indicates the first The first center point Feature vectors of neighboring points: Channel-by-channel mean calculation: For each feature channel Calculate the mean along the dimension of neighboring points. , Representing the eigenvector In the The values ​​on each channel For the first The average of each channel; Feature decentralization: computation , These are the decentralized feature values; Calculation of combined standard deviation: ; Global normalization: ,in It is a very small constant; Affine transformations can be learned: ,in , For channel-by-channel learnable parameters.

5. The point cloud surface reconstruction method fusing geometrical radiality and gated attention according to claim 4, characterized in that, The output of the gated attention aggregation decoding module in step S5 is the implicit surface parameter prediction vector at query point x. ,in This is the predicted value of absolute distance. The predicted value is the symbolic Logits. The fully connected decoding network includes: a first fully connected layer that aggregates feature vectors. From the perspective of dimensions The hidden layer feature vector h is obtained by compressing to the intermediate dimension D of the decoding, followed by batch normalization and ReLU activation. The second fully connected layer will Mapping to the two-dimensional output space yields the prediction vector y.

6. The point cloud surface reconstruction method fusing geometrical radiality and gated attention according to claim 1, characterized in that, The method also includes a training phase: Repeat steps S1 to S6 for multiple point cloud-surface sample pairs and their query points in the training set, employing a multi-task joint learning strategy, with the total loss function... ,in For based on Norm-based distance regression loss, The binary cross-entropy sign classification loss is used. To balance the weight coefficients, the backpropagation algorithm is used to update all trainable parameters end-to-end until convergence. It also includes an inference phase: repeatedly executing steps S1 to S6 for each sampling point in the voxel grid to estimate the signed distance field, and performing a synthesis operation. The signed distance estimate is calculated and the zero isosurface is extracted by applying the moving cube algorithm to obtain the final reconstructed 3D surface mesh.

7. The point cloud surface reconstruction method fusing geometrical radiality and gated attention according to claim 1, characterized in that, The local branch point-by-point feature extraction network described in step S2 consists of multiple layers of one-dimensional convolution, batch normalization, and ReLU activation function. Each layer of one-dimensional convolution is followed by batch normalization and ReLU activation to achieve progressive dimensionality increase and abstraction of features.

8. A point cloud surface reconstruction device integrating geometrical radial and gated attention methods, characterized in that, include: The point set acquisition and alignment module is used to acquire a local block point set and a global subsample point set centered on the query point x, and to perform coordinate translation and alignment of the global subsample point set with the query point as the origin. The geometric affine modulation module (GAM) is used to adaptively normalize and learnable affine transformation the coordinate features of the local block point set, including: a channel-wise mean calculation and mean removal unit, a cross-channel and cross-point joint standard deviation calculation unit, a global normalization unit, and a channel-wise learnable affine transformation unit. The local branch encoder, consisting of multiple layers of one-dimensional convolution, batch normalization, and ReLU activation, is used to receive modulated coordinate features and output a local point-by-point feature matrix that retains the point dimension. The global branch encoder has the same structure as the local branch encoder but with independent parameters and does not contain the geometric affine modulation module (GAM). It is used to receive the global sub-sample point set after translation and alignment and output the global feature vector after symmetric pooling. The feature fusion module is used to project and reduce the dimensionality of the local pointwise feature matrix and the global feature vector respectively, and then concatenate them to obtain a fused feature matrix. ; The gated attention aggregation decoding module includes: a gated weight generation unit, which generates a gated weight matrix using one-dimensional convolution and sigmoid activation; a feature modulation unit, which multiplies the gated weights and fused features element-wise; a query / numerical mapping unit, which generates a query feature matrix and a numerical feature matrix using two sets of independent one-dimensional convolutions; an attention weight calculation unit, which performs Softmax normalization along the point dimension and takes the maximum value along the channel dimension to obtain the importance weight; and a weighted aggregation unit, which sums the numerical features and importance weights to obtain the aggregated feature vector. A fully connected decoding network is used to map the aggregated feature vector to the implicit surface prediction value corresponding to the query point, and the output includes the absolute distance prediction value and the symbolic Logits prediction value.

9. The point cloud surface reconstruction apparatus integrating geometrical radiality and gated attention according to claim 8, characterized in that, The device also includes a training phase module: Multiple point cloud-surface sample pairs and their query points in the training set are sequentially input into the point set acquisition and alignment module, the geometric affine modulation module (GAM), the local branch encoder, the global branch encoder, the feature fusion module, the gated attention aggregation decoding module, and the fully connected decoding network. A multi-task joint learning strategy is adopted, and the total loss function is... ,in For based on Norm-based distance regression loss, The binary cross-entropy sign classification loss is used. To balance the weight coefficients, the backpropagation algorithm is used to update all trainable parameters end-to-end until convergence. The device further includes an inference module: a module for acquiring and aligning repeated input point sets for each sampling point in the voxel grid, a geometric affine modulation module (GAM), a local branch encoder, a global branch encoder, a feature fusion module, a gated attention aggregation decoding module, and a fully connected decoding network to estimate the signed distance field, and then synthesizes it. The signed distance estimate is calculated and the zero isosurface is extracted by applying the moving cube algorithm to obtain the final reconstructed 3D surface mesh.

10. A computer device, characterized in that, include: processor; Memory for storing computer programs executable by the processor; Wherein, when the processor executes the computer program, it implements the steps of the point cloud surface reconstruction method that integrates geometrical radiation and gated attention as described in any one of claims 1 to 7.

Images