A point cloud robust classification method based on local geometric change perception and hybrid expert adaptive correction and a classification system thereof

Abstract

A robust point cloud classification method and system based on local geometric change perception and hybrid expert adaptive correction is presented, belonging to the fields of computer vision and artificial intelligence. To address the problems of existing point cloud classification methods facing adversarial geometric perturbations, such as lack of fine-grained perception of local geometric changes, reliance on single correction strategies that are difficult to adapt to diverse attack patterns, and the tendency to destroy the discriminative geometric structure of the point cloud while removing perturbations, the method includes: constructing a local neighborhood for each point in the original point cloud and extracting multi-scale local geometric change perception descriptors; constructing point-by-point conditional features; inputting the point-by-point conditional features into a hybrid expert adaptive correction module to generate point-by-point correction vectors; applying correction magnitude constraints to the point-by-point correction vectors and performing weighted fusion with sample-level global gating factors to update the point cloud coordinates; and inputting the corrected point cloud into a classifier for classification. This method aims to improve the robustness of point cloud classification under adversarial geometric perturbations.

Description

Technical Field

[0001] This invention relates to a robust point cloud classification method and system based on local geometric change perception and hybrid expert adaptive correction, belonging to the fields of computer vision and artificial intelligence technology. Background Technology

[0002] With the rapid development of safety-critical autonomous systems such as autonomous driving, robot navigation, and intelligent monitoring, 3D point cloud, as an important form of perception data, is widely used in tasks such as object classification, semantic segmentation, 3D modeling, and scene understanding due to its advantages such as insensitivity to changes in lighting and ability to directly express spatial geometric structures.

[0003] In recent years, deep learning-based point cloud processing methods have made significant progress. Representative models such as PointNet, PointNet++, and DGCNN have significantly improved the feature extraction and semantic understanding capabilities of irregular and disordered point clouds through permutation-invariant learning, local neighborhood structure modeling, and graph convolutional networks. However, while pursuing high classification accuracy, existing methods generally neglect the issue of robustness to adversarial perturbations.

[0004] Research shows that point cloud models are highly vulnerable to adversarial attacks. Attackers can induce incorrect classifications by introducing small, carefully crafted geometric perturbations (such as slight point displacements, local structural distortions, or subtle changes in neighborhood geometry) with almost no alteration to the overall shape of the object. Unlike pixel perturbations in two-dimensional images, attacks on three-dimensional point clouds directly manipulate geometric coordinates, leading to structural distortions that are often more difficult to repair. This vulnerability poses a serious threat to the security of autonomous systems that rely on point cloud perception.

[0005] To address these challenges, existing defense methods are mainly divided into three categories:

[0006] The first category is defense methods based on preprocessing, such as point cloud filtering, denoising, random point discarding, or statistical outlier removal, which attempt to eliminate adversarial noise before it is input into the classifier. Although these methods are computationally efficient, they often inadvertently degrade classification performance on clean samples and remain susceptible to adaptive attacks.

[0007] The second category involves architectural-level improvements that enhance model robustness by redesigning feature extraction or aggregation mechanisms, such as graph-based network design, robust structured declarative classifiers, or neural architecture search. However, their effectiveness is often limited to specific attack scenarios and is difficult to generalize to diverse adversarial perturbation patterns.

[0008] The third category is adversarial training methods, which promote more robust decision boundary learning by explicitly exposing the model to adversarial examples during training. Existing research shows that adversarial training can significantly enhance the robustness of point cloud classifiers under different attack settings and is currently a relatively effective defense method.

[0009] However, the above-mentioned defense methods share the following common limitations:

[0010] 1. Lack of fine-grained perception of local geometric changes. Most existing methods rely on global features or uniform correction strategies, making it difficult to effectively distinguish between adversarial perturbations and benign geometric variations. Adversarial perturbations are highly localized, exhibiting subtle deformations with direction awareness and structure dependence, but current defense mechanisms are not specifically designed to capture these structured local distortions.

[0011] 2. The correction strategy is singular and lacks adaptability. Existing methods apply a uniform operation to all points, failing to recognize that adversarial perturbations in different local structural regions require customized correction strategies: planar or smooth regions are usually suitable for moderate smoothing constraints, while edges, ridges, or high curvature regions require fine local corrections to prevent unintentional destruction of discriminative geometric features.

[0012] 3. Insufficient ability to preserve geometric structure. While removing perturbations, single correction models can easily destroy the original discriminative geometric structure of point clouds, leading to structural degradation or over-smoothing artifacts, which affect classification performance.

[0013] 4. Difficulty in dealing with diverse adversarial attack patterns. Different attack methods exploit different geometric vulnerabilities: such as local curvature distortion, neighborhood inconsistency, or point density variation. A single defense strategy is insufficient to provide consistent protection against different types of attacks.

[0014] Therefore, there is an urgent need for a point cloud defense method that can both perceive local geometric changes and adaptively select correction strategies based on local features, so as to effectively resist diverse adversarial attacks while maintaining the classification accuracy and geometric integrity of clean samples. Summary of the Invention

[0015] The purpose of this invention is to address the problems of existing point cloud classification methods in the face of adversarial geometric perturbations, such as lack of fine-grained perception of local geometric changes, single correction strategies that are difficult to adapt to diverse attack patterns, and easy destruction of the discriminative geometric structure of point clouds while removing perturbations. This invention provides a robust point cloud classification method and system based on local geometric change perception and hybrid expert adaptive correction.

[0016] The robust point cloud classification method based on local geometric change perception and hybrid expert adaptive correction described in this invention includes the following steps:

[0017] Step 1: Construct a local neighborhood for each point in the original point cloud, and extract a multi-scale local geometric change perception descriptor based on the local neighborhood;

[0018] Step 2: Based on the multi-scale local geometric change perception descriptor, and combined with the original point coordinates, construct point-by-point conditional features; input the point-by-point conditional features into the hybrid expert adaptive correction module, dynamically allocate the weights of multiple expert networks through a gating network, and perform weighted fusion of the correction vectors output by each expert network to generate point-by-point correction vectors;

[0019] Step 3: Apply a correction amplitude constraint to the point-by-point correction vector, and perform weighted fusion with the sample-level global gating factor to update the point cloud coordinates and obtain the corrected point cloud;

[0020] Step 4: Input the corrected point cloud into the classifier for classification, and adopt a joint adversarial training strategy to optimize the classification loss of clean samples and adversarial samples. Train the local geometric change perception module, the hybrid expert adaptive correction module and the classifier end-to-end.

[0021] Preferably, the specific method for constructing a local neighborhood for each point in the original point cloud in step 1 includes:

[0022] The original point cloud input for:

[0023] ;

[0024] in, Represents the first point in the point cloud One point, This represents the total number of points in the point cloud. Represents the three-dimensional real number space;

[0025] Construct a k-nearest neighbor region based on Euclidean distance;

[0026] For any two points in the point cloud and Euclidean distance squared Defined as:

[0027] ;

[0028] point k-nearest neighbor set Defined as:

[0029] ;

[0030] in, This refers to the k-nearest neighbor algorithm, used to find distances. Recent One point, Indicates the number of nearest neighbors;

[0031] yes One of the k nearest neighbors, where the center point is... It was explicitly excluded.

[0032] Preferably, the specific method for extracting multi-scale local geometric change-aware descriptors based on local neighborhoods in step 1 includes:

[0033] For any neighborhood point Calculate its relative to the center point relative displacement vector :

[0034] ;

[0035] in, Point The set of k-nearest neighbors;

[0036] Calculate the intensity measure of local geometric changes :

[0037] ;

[0038] in, Represents extremely small positive numbers;

[0039] Constructing the local covariance matrix based on relative displacement vectors :

[0040] ;

[0041] in, express Transpose of;

[0042] Symmetricize the covariance matrix:

[0043] ;

[0044] in, express Transpose of;

[0045] right Eigenvalue decomposition yields:

[0046] ;

[0047] in, , , express The three eigenvalues;

[0048] The covariance spectrum is normalized using the largest eigenvalue, and the following local structure descriptor is constructed:

[0049] ;

[0050] in, Represents anisotropic descriptors, Represents a planarity descriptor. Represents the sphericity descriptor;

[0051] Repeat the above process at different neighborhood scales k to obtain multi-scale local geometric change perceptual descriptors. :

[0052] .

[0053] Preferably, the specific method for constructing point-by-point conditional features in step 2 includes:

[0054] The multi-scale local geometric change perception descriptor is concatenated with the original point coordinates to form point-by-point conditional features:

[0055] ;

[0056] in, Point Pointwise conditional eigenvectors Point At the neighborhood scale Local geometric change perception descriptor under, Indicates the first Each neighborhood scale.

[0057] Preferably, the hybrid expert adaptive correction module in step 2 includes: E expert networks, a gating network, and a weighted fusion layer;

[0058] The expert network, in which each expert network takes point-by-point conditional features as input and outputs a corresponding point-by-point coordinate correction vector, is described.

[0059] No. The output of an expert for:

[0060] ;

[0061] in, Representing the expert model;

[0062] The gated network takes point-by-point conditional features as input and outputs an E-dimensional weight distribution through a Softmax function:

[0063] ;

[0064] in, Indicates the first The weight distribution of each point on the expert, This represents a point-to-point gating network;

[0065] The weighted fusion layer performs a weighted summation of the correction vectors output by the E expert networks based on the weight distribution of the gated network output, to obtain the final point-by-point correction vector. :

[0066] ;

[0067] in, Indicates the first The point at the th The weight of each expert.

[0068] Preferably, the specific method for applying the correction magnitude constraint to the point-by-point correction vector in step 3 includes:

[0069] The norm of the correction vector at each point is limited to a preset maximum correction range, i.e., the correction vector is bounded. Represented as:

[0070] ;

[0071] in, This indicates the maximum allowable correction range.

[0072] Preferably, the sample-level global gating factor described in step 3 for:

[0073] ;

[0074] in, This represents the Sigmoid function. Indicates the point-by-point weighting coefficient;

[0075] The sample-level global gating factor adjusts the overall correction intensity at the sample level, suppressing overcorrection when the anti-distortion degree of the input point cloud is limited.

[0076] Preferably, the specific method for updating the point cloud coordinates in step 3 includes:

[0077] ;

[0078] in, This represents the updated point cloud coordinates.

[0079] Preferably, the joint adversarial training strategy described in step 4 specifically includes:

[0080] In each round of training, adversarial examples are generated from the input point cloud;

[0081] Simultaneously calculate the classification loss for clean samples and adversarial samples;

[0082] By jointly optimizing the parameters of the local geometric change perception module, the hybrid expert adaptive correction module, and the classifier, the model can enhance its robustness against adversarial examples while maintaining the classification accuracy of clean samples.

[0083] The present invention discloses a robust point cloud classification system based on local geometric change perception and hybrid expert adaptive correction, comprising:

[0084] The local geometric change perception module is used to construct a local neighborhood for each point in the original point cloud and extract a multi-scale local geometric change perception descriptor.

[0085] The hybrid expert adaptive correction module is used to generate point-by-point correction vectors based on multi-scale local geometric change perception descriptors and update point cloud coordinates;

[0086] A classifier is used to classify and identify the corrected point cloud.

[0087] The joint adversarial training module is used to perform end-to-end joint optimization of the entire system.

[0088] Advantages of this invention: This invention extracts multi-scale geometric descriptors through a local geometric change perception module, achieving fine-grained perception of adversarial perturbations; through a hybrid expert adaptive correction module, expert weights are dynamically allocated according to local geometric conditions, achieving differentiated repair of different types of perturbations; through joint adversarial training, robustness to diverse adversarial attacks is significantly enhanced while maintaining the classification accuracy of clean samples. Attached Figure Description

[0089] Figure 1 This is a schematic diagram of the overall process of the present invention, which combines local geometric change perception and adaptive hybrid expert point-by-point correction.

[0090] Figure 2 This is a schematic diagram of the structure of the local geometric change sensing module of the present invention;

[0091] Figure 3 This is a schematic diagram of the adaptive hybrid expert-guided point-by-point correction mechanism of the present invention. Detailed Implementation

[0092] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0093] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0094] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0095] Example 1:

[0096] The following is combined Figures 1-3 This embodiment describes a robust point cloud classification method based on local geometric change perception and hybrid expert adaptive correction. Figure 1 As shown, it includes the following steps:

[0097] Step 1: Construct a local neighborhood for each point in the original point cloud, and extract a multi-scale local geometric change perception descriptor based on the local neighborhood; such as... Figure 2 As shown;

[0098] Furthermore, the specific method for constructing a local neighborhood for each point in the original point cloud as described in step 1 includes:

[0099] The original point cloud input for:

[0100] ;

[0101] in, Represents the first point in the point cloud One point, This represents the total number of points in the point cloud. Represents the three-dimensional real number space;

[0102] Construct a k-nearest neighbor region based on Euclidean distance;

[0103] For any two points in the point cloud and Euclidean distance squared Defined as:

[0104] ;

[0105] point k-nearest neighbor set Defined as:

[0106] ;

[0107] in, This refers to the k-nearest neighbor algorithm, used to find distances. Recent One point, Indicates the number of nearest neighbors;

[0108] yes One of the k nearest neighbors, where the center point is... It was explicitly excluded.

[0109] Furthermore, the specific method for extracting multi-scale local geometric change-aware descriptors based on local neighborhoods as described in step 1 includes:

[0110] For any neighborhood point Calculate its relative to the center point relative displacement vector :

[0111] ;

[0112] in, Point The set of k-nearest neighbors;

[0113] Calculate the intensity measure of local geometric changes :

[0114] ;

[0115] in, Represents extremely small positive numbers;

[0116] Constructing the local covariance matrix based on relative displacement vectors :

[0117] ;

[0118] in, express Transpose of;

[0119] Symmetricize the covariance matrix:

[0120] ;

[0121] in, express Transpose of;

[0122] right Eigenvalue decomposition yields:

[0123] ;

[0124] in, , , express The three eigenvalues;

[0125] The covariance spectrum is normalized using the largest eigenvalue, and the following local structure descriptor is constructed:

[0126] ;

[0127] in, Represents anisotropic descriptors, Represents a planarity descriptor. Represents the sphericity descriptor;

[0128] Repeat the above process at different neighborhood scales k to obtain multi-scale local geometric change perceptual descriptors. :

[0129] .

[0130] Step 2: Based on the multi-scale local geometric change perception descriptor and combined with the original point coordinates, construct point-by-point conditional features; input the point-by-point conditional features into the hybrid expert adaptive correction module, dynamically allocate the weights of multiple expert networks through a gating network, and perform weighted fusion of the correction vectors output by each expert network to generate a point-by-point correction vector; such as Figure 3 As shown;

[0131] Furthermore, the specific method for constructing point-by-point conditional features in step 2 includes:

[0132] The multi-scale local geometric change perception descriptor is concatenated with the original point coordinates to form point-by-point conditional features:

[0133] ;

[0134] in, Point Pointwise conditional eigenvectors Point At the neighborhood scale Local geometric change perception descriptor under, Indicates the first Each neighborhood scale.

[0135] Furthermore, the hybrid expert adaptive correction module described in step 2 includes: E expert networks, a gating network, and a weighted fusion layer;

[0136] The expert network, in which each expert network takes point-by-point conditional features as input and outputs a corresponding point-by-point coordinate correction vector, is described.

[0137] No. The output of an expert for:

[0138] ;

[0139] in, Representing the expert model;

[0140] The gated network takes point-by-point conditional features as input and outputs an E-dimensional weight distribution through a Softmax function:

[0141] ;

[0142] in, Indicates the first The weight distribution of each point on the expert, This represents a point-to-point gating network;

[0143] The weighted fusion layer performs a weighted summation of the correction vectors output by the E expert networks based on the weight distribution of the gated network output, to obtain the final point-by-point correction vector. :

[0144] ;

[0145] in, Indicates the first The point at the th The weight of each expert.

[0146] Step 3: Apply a correction amplitude constraint to the point-by-point correction vector, and perform weighted fusion with the sample-level global gating factor to update the point cloud coordinates and obtain the corrected point cloud;

[0147] Furthermore, the specific method for applying the correction magnitude constraint to the point-by-point correction vector in step 3 includes:

[0148] The norm of the correction vector at each point is limited to a preset maximum correction range, i.e., the correction vector is bounded. Represented as:

[0149] ;

[0150] in, This indicates the maximum allowable correction range.

[0151] Furthermore, the sample-level global gating factor mentioned in step 3... for:

[0152] ;

[0153] in, This represents the Sigmoid function. Indicates the point-by-point weighting coefficient;

[0154] The sample-level global gating factor adjusts the overall correction intensity at the sample level, suppressing overcorrection when the anti-distortion degree of the input point cloud is limited.

[0155] Furthermore, the specific method for updating the point cloud coordinates in step 3 includes:

[0156] ;

[0157] in, This represents the updated point cloud coordinates.

[0158] Step 4: Input the corrected point cloud into the classifier for classification, and adopt a joint adversarial training strategy to optimize the classification loss of clean samples and adversarial samples. Train the local geometric change perception module, the hybrid expert adaptive correction module and the classifier end-to-end.

[0159] Furthermore, the joint adversarial training strategy described in step 4 specifically includes:

[0160] In each round of training, adversarial examples are generated from the input point cloud;

[0161] Simultaneously calculate the classification loss for clean samples and adversarial samples;

[0162] By jointly optimizing the parameters of the local geometric change perception module, the hybrid expert adaptive correction module, and the classifier, the model can enhance its robustness against adversarial examples while maintaining the classification accuracy of clean samples.

[0163] In this embodiment, the core principle lies in the closed-loop mechanism of "perception-decision-correction". First, a local geometric change perception module extracts multi-scale geometric descriptors (including local change intensity, anisotropy, planarity, and sphericity). These descriptors quantify the degree of geometric perturbation and structural type in the neighborhood of each point, providing refined geometric priors for subsequent correction. Second, the gating network in the hybrid expert adaptive correction module dynamically evaluates the damage type and degree of each point based on these geometric priors, assigning weights to different expert networks—experts skilled in handling planar perturbations, experts skilled in handling edge perturbations, and experts skilled in handling noise perturbations are organically combined to achieve accurate repair of different types of perturbations. Finally, through correction amplitude constraints and sample-level global gating, the correction process is ensured to be both sufficient and appropriate, avoiding over-processing of clean samples.

[0164] In this embodiment, experiments on public datasets such as ModelNet40 and ShapeNet demonstrate that the success rate of this method is reduced when facing various adversarial attacks (including gradient-based attacks, iterative attacks, and targeted attacks), significantly outperforming existing defense methods. Particularly when dealing with attacks involving local curvature distortion and neighborhood inconsistencies, this method exhibits unique advantages, proving its effectiveness against diverse attack patterns.

[0165] Example 2:

[0166] This embodiment describes a robust point cloud classification system based on local geometric change perception and hybrid expert adaptive correction, which includes:

[0167] The local geometric change perception module is used to construct a local neighborhood for each point in the original point cloud and extract a multi-scale local geometric change perception descriptor.

[0168] The hybrid expert adaptive correction module is used to generate point-by-point correction vectors based on multi-scale local geometric change perception descriptors and update point cloud coordinates;

[0169] A classifier is used to classify and identify the corrected point cloud.

[0170] The joint adversarial training module is used to perform end-to-end joint optimization of the entire system.

[0171] In this embodiment, the method described in Example 1 is implemented as a complete robust point cloud classification system. Four modules work collaboratively: the local geometric change perception module is responsible for extracting multi-scale geometric features, acting as the system's "perceptual organ"; the hybrid expert adaptive correction module is responsible for making decisions and repairs based on the perceived information, acting as the system's "decision center"; the classifier is responsible for the final recognition task, acting as the system's "executive organ"; and the joint adversarial training module runs throughout the entire training process, ensuring that each module optimizes collaboratively in an adversarial environment. This modular design gives the system good scalability and maintainability, allowing each module to be optimized independently without affecting the overall functionality.

[0172] In this embodiment, the constructed classification system demonstrates excellent robustness in practical applications. Furthermore, the system's modular design allows users to fine-tune individual modules for specific application scenarios (such as low-light, high-noise environments) without retraining the entire system, significantly improving deployment flexibility.

[0173] This invention proposes a robust point cloud classification method and system based on local geometric change perception and hybrid expert adaptive correction. By constructing multi-scale local neighborhoods and extracting multi-dimensional geometric descriptors such as local change intensity, anisotropy, planarity, and sphericity, it achieves refined perception of adversarial perturbations. Unlike traditional methods that rely on global features, this invention can accurately locate the perturbed region and quantify the degree of perturbation, providing accurate geometric priors for subsequent correction. A hybrid expert architecture is adopted, using a gating network to dynamically allocate expert weights based on local geometric conditions, achieving differentiated repair for different types of perturbations. Smoothing constraints are applied to planar regions, conformal correction is applied to edge regions, and moderate filtering is applied to noise points—this "point-specific" strategy effectively removes perturbations while preserving the discriminative geometric structure of the point cloud to the greatest extent. A dual mechanism is introduced: a correction amplitude constraint and a sample-level global gating factor. The former ensures that the displacement of each point does not exceed a preset threshold, preventing over-correction; the latter adaptively adjusts the correction intensity according to the overall perturbation level of the input point cloud, automatically suppressing correction when the adversarial distortion level is low, effectively balancing robustness and clean sample accuracy. This approach integrates local geometric change perception, hybrid expert adaptive correction, and classifiers into a unified framework, employing a joint adversarial training strategy to simultaneously optimize targets for both clean and adversarial samples. This training method enables the modules to co-evolve: the perception module learns to identify more subtle perturbations, the correction module learns to repair more accurately, and the classifier learns more robust decision boundaries.

[0174] Because the method of this invention is not designed for specific attack types, but rather achieves defense by learning general local geometric change patterns, it exhibits stable defense against various adversarial attacks (including untargeted attacks, targeted attacks, gradient-based attacks, and iterative attacks), demonstrating good generalization ability. Through the guidance of multi-scale geometric descriptors and differential correction by expert networks, this invention effectively preserves the original discriminative geometric features of the point cloud, such as edges, corners, and surfaces, while removing adversarial perturbations, avoiding the structural degradation or over-smoothing problems caused by traditional denoising methods.

[0175] In this invention, adversarial perturbations in 3D point clouds often exploit the vulnerability of classifiers to subtle structural deformations. However, traditional "blind" denoising methods often struggle to strike a balance between noise removal and preservation of discriminative geometric details. To address this, we propose an adaptive defense framework that shifts from unified recovery to geometry-aware manifold correction. By introducing a local geometric change sensing module, this framework first quantifies structural anomalies, providing a fine-grained diagnostic prior for the recovery process. This prior guides our designed adaptive hybrid expert (MoE)-guided pointwise correction mechanism, which utilizes a soft-route MoE architecture to dynamically interpolate between different specialized correction strategies. This design enables the model to adaptively project perturbation points back onto their clean manifold while preventing overcorrection of stable geometric regions through a global modulation gate. Extensive experiments on benchmark datasets demonstrate that this collaborative approach not only outperforms existing defenses against strong adversarial attacks but also maintains high fidelity on clean samples. By jointly optimizing the recovery and recognition processes through adversarial training, our work provides a robust and interpretable solution for safeguarding 3D point cloud perception systems.

[0176] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A robust point cloud classification method based on local geometric change perception and hybrid expert adaptive correction, characterized in that, It includes the following steps: Step 1: Construct a local neighborhood for each point in the original point cloud, and extract a multi-scale local geometric change perception descriptor based on the local neighborhood; Step 2: Based on the multi-scale local geometric change perception descriptor, and combined with the original point coordinates, construct point-by-point conditional features; input the point-by-point conditional features into the hybrid expert adaptive correction module, dynamically allocate the weights of multiple expert networks through a gating network, and perform weighted fusion of the correction vectors output by each expert network to generate point-by-point correction vectors; Step 3: Apply a correction amplitude constraint to the point-by-point correction vector, and perform weighted fusion with the sample-level global gating factor to update the point cloud coordinates and obtain the corrected point cloud; Step 4: Input the corrected point cloud into the classifier for classification, and adopt a joint adversarial training strategy to optimize the classification loss of clean samples and adversarial samples. Train the local geometric change perception module, the hybrid expert adaptive correction module and the classifier end-to-end.

2. The robust point cloud classification method based on local geometric change perception and hybrid expert adaptive correction according to claim 1, characterized in that, The specific method for constructing a local neighborhood for each point in the original point cloud as described in step 1 includes: The original point cloud input for: ； in, Represents the first point in the point cloud One point, This represents the total number of points in the point cloud. Represents the three-dimensional real number space; Construct a k-nearest neighbor region based on Euclidean distance; For any two points in the point cloud and Euclidean distance squared Defined as: ； point k-nearest neighbor set Defined as: ； in, This refers to the k-nearest neighbor algorithm, used to find distances. Recent One point, Indicates the number of nearest neighbors; yes One of the k nearest neighbors, where the center point is... It was explicitly excluded.

3. The robust point cloud classification method based on local geometric change perception and hybrid expert adaptive correction according to claim 2, characterized in that, The specific method for extracting multi-scale local geometric change-aware descriptors based on local neighborhoods as described in step 1 includes: For any neighborhood point Calculate its relative to the center point relative displacement vector : ； in, Point The set of k-nearest neighbors; Calculate the intensity measure of local geometric changes : ； in, Represents extremely small positive numbers; Constructing the local covariance matrix based on relative displacement vectors : ； in, express Transpose of; Symmetricize the covariance matrix: ； in, express Transpose of; right Eigenvalue decomposition yields: ； in, , , express The three eigenvalues; The covariance spectrum is normalized using the largest eigenvalue, and the following local structure descriptor is constructed: ； in, Represents anisotropic descriptors, Represents a planarity descriptor. Represents the sphericity descriptor; Repeat the above process at different neighborhood scales k to obtain multi-scale local geometric change perceptual descriptors. : 。 4. The robust point cloud classification method based on local geometric change perception and hybrid expert adaptive correction according to claim 3, characterized in that, The specific method for constructing point-by-point conditional features in step 2 includes: The multi-scale local geometric change perception descriptor is concatenated with the original point coordinates to form point-by-point conditional features: ； in, Point Pointwise conditional eigenvectors Point At the neighborhood scale Local geometric change perception descriptor under, Indicates the first Each neighborhood scale.

5. A robust point cloud classification method based on local geometric change perception and hybrid expert adaptive correction according to claim 3, characterized in that, The hybrid expert adaptive correction module described in step 2 includes: E expert networks, a gating network, and a weighted fusion layer; The expert network, in which each expert network takes point-by-point conditional features as input and outputs a corresponding point-by-point coordinate correction vector, is described. No. The output of an expert for: ； in, Representing the expert model; The gated network takes point-by-point conditional features as input and outputs an E-dimensional weight distribution through a Softmax function: ； in, Indicates the first The weight distribution of each point on the expert, This represents a point-to-point gating network; The weighted fusion layer performs a weighted summation of the correction vectors output by the E expert networks based on the weight distribution of the gated network output, to obtain the final point-by-point correction vector. : ； in, Indicates the first The point at the th The weight of each expert.

6. A robust point cloud classification method based on local geometric change perception and hybrid expert adaptive correction according to claim 5, characterized in that, The specific method for applying correction magnitude constraints to the point-by-point correction vector as described in step 3 includes: The norm of the correction vector at each point is limited to a preset maximum correction range, i.e., the correction vector is bounded. Represented as: ； in, This indicates the maximum allowable correction range.

7. A robust point cloud classification method based on local geometric change perception and hybrid expert adaptive correction according to claim 6, characterized in that, Step 3, sample-level global gating factor for: ； in, This represents the Sigmoid function. Indicates the point-by-point weighting coefficient; The sample-level global gating factor adjusts the overall correction intensity at the sample level, suppressing overcorrection when the anti-distortion degree of the input point cloud is limited.

8. A robust point cloud classification method based on local geometric change perception and hybrid expert adaptive correction according to claim 7, characterized in that, The specific method for updating point cloud coordinates in step 3 includes: ； in, This represents the updated point cloud coordinates.

9. A robust point cloud classification method based on local geometric change perception and hybrid expert adaptive correction according to claim 8, characterized in that, The joint adversarial training strategy described in step 4 specifically includes: In each round of training, adversarial examples are generated from the input point cloud; Simultaneously calculate the classification loss for clean samples and adversarial samples; By jointly optimizing the parameters of the local geometric change perception module, the hybrid expert adaptive correction module, and the classifier, the model can enhance its robustness against adversarial examples while maintaining the classification accuracy of clean samples.

10. A classification system for implementing the robust point cloud classification method based on local geometric change perception and hybrid expert adaptive correction as described in any one of claims 1-9, characterized in that, It includes: The local geometric change perception module is used to construct a local neighborhood for each point in the original point cloud and extract a multi-scale local geometric change perception descriptor. The hybrid expert adaptive correction module is used to generate point-by-point correction vectors based on multi-scale local geometric change perception descriptors and update point cloud coordinates; A classifier is used to classify and identify the corrected point cloud. The joint adversarial training module is used to perform end-to-end joint optimization of the entire system.

Images