Point cloud anomaly segmentation method and device based on pseudo-anomaly distillation discriminative network
By using a pseudo-anomaly distillation discriminant network, the problems of insufficient receptive field and excessive storage consumption in point cloud anomaly segmentation are solved, achieving more efficient and accurate 3D point cloud anomaly segmentation, which is suitable for the detection of real anomalies in industrial scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUAZHONG UNIV OF SCI & TECH
- Filing Date
- 2023-12-14
- Publication Date
- 2026-06-30
AI Technical Summary
Existing point cloud anomaly segmentation methods suffer from problems such as insufficient receptive field and excessive storage space consumption, making it difficult to effectively address the sparsity, disorder, and rotation invariance of 3D point cloud data, resulting in low detection accuracy and low efficiency.
A pseudo-anomaly distillation-based discriminant network approach is adopted. A pseudo-anomaly generation module simulates anomalies on the training point cloud, and a Point Transformer network is used for feature extraction. The student network and the discriminant network are trained end-to-end, and supervised training is carried out in combination with pseudo-anomaly labels to achieve more efficient and accurate anomaly segmentation.
It achieves point cloud anomaly segmentation with a larger receptive field, reduces the false detection rate, improves segmentation accuracy and inference speed, and reduces storage space consumption.
Smart Images
Figure CN117689675B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer vision, and more specifically, relates to a point cloud anomaly segmentation method and apparatus based on a pseudo-anomaly distillation discrimination network. Background Technology
[0002] In industrial production, anomaly segmentation of manufactured products is an essential step. Various problems arising during production or use, such as assembly errors, bumps, and contamination, render products unusable. Various anomaly segmentation algorithms are needed to analyze the product for anomalies and segment out the abnormal parts. For products with anomalies, different repair methods are required depending on the size and location of the anomaly. In traditional 2D industrial anomaly detection, an image of the object is captured by a camera with a fixed viewpoint. The image is then preprocessed and input into a neural network for analysis to detect anomalies and segment out the abnormal regions. However, some anomalies, such as dents, protrusions, and surface deformations, are difficult to detect successfully using only 2D images and suffer from incomplete segmentation. Therefore, it is necessary to segment anomalies using 3D point cloud data with the aid of 3D structural information. Currently, the field of industrial anomaly segmentation in 3D point cloud is in its early stages of development, and many problems need to be solved. Migrating 2D anomaly segmentation algorithms to 3D presents many difficulties. For 3D point cloud data, due to the characteristics of point clouds such as sparsity, disorder, rotation invariance, and translation invariance, there is a significant difference from image data. Most 2D feature extraction methods will fail in 3D. Converting point clouds into voxels to adapt to 3D convolution operations will lead to a significant loss of spatial information. At the same time, the sparsity and disorder of point clouds will make the computational cost of 3D convolution very large.
[0003] Existing point cloud anomaly segmentation methods can be categorized into the following types: ST-based methods use a pre-trained teacher network (such as RandLA-Net) to perform knowledge distillation on a student network with a similar structure, and locate anomalies based on the magnitude of regression error during testing; memory-pool-based methods store reference normal sample features in a memory pool, and identify anomalies by comparing test sample features with features in the memory pool.
[0004] Among the existing point cloud anomaly segmentation methods, the ST architecture-based methods suffer from insufficient receptive field and underfitting of the student network, while the memory pool-based methods suffer from excessive storage space consumption and low inference efficiency. The existing methods cannot solve these problems well. Summary of the Invention
[0005] To address the shortcomings or improvement needs of existing technologies, this invention provides a point cloud anomaly segmentation method based on a pseudo-anomaly distillation discriminant network. This method effectively avoids the storage consumption associated with memory pool-based methods and uses pseudo-anomalies in the point cloud for supervised training, achieving more efficient, less resource-intensive, and more accurate point cloud anomaly segmentation.
[0006] To achieve the above objectives, according to one aspect of the present invention, a point cloud anomaly segmentation method based on a pseudo-anomaly distillation discriminant network is provided, comprising the following steps:
[0007] (1) Use the pseudo-anomaly generation module to simulate anomalies on the training point cloud; including (1.1) setting the pseudo-anomaly position parameters; (1.2) setting the pseudo-anomaly shape parameters; (1.3) implementing convex or concave pseudo-anomalies;
[0008] (2) Use the generated pseudo-anomaly point cloud for teacher-student network training;
[0009] (3) Train the discriminant network using student output features and teacher output features;
[0010] (4) During training, fix the network parameters in the teacher feature extraction module and train the student extraction module and discrimination network end-to-end;
[0011] (5) During the testing phase, the test point cloud is input into the teacher feature extraction module and the student feature extraction module respectively to obtain the teacher output feature and the student output feature. The difference between the teacher output feature and the student output feature is calculated and then input into the discriminant network together with the student output feature to obtain a binary probability point cloud. The last dimension is taken as the anomaly score point cloud. The larger the anomaly score value of a point in the point cloud, the greater the probability that the point is an anomaly. The inverse distance weight is used to interpolate into the original point cloud, and the anomaly segmentation result is obtained according to the preset threshold.
[0012] In one embodiment of the present invention, step (1.1) specifically includes: determining the center point of the pseudo-anomaly generation so that it cannot be close to the edge region; mapping the point cloud onto a 2D image, filling the unmapped area in the 2D image using a 3×3 morphological dilation operation, calculating the position of the object's edge point, randomly selecting a point in the object's foreground region, and setting it as the center point c1 if the 4n neighborhood of that point does not contain the object's edge point, defining the pseudo-anomaly region as the n-neighborhood of point c1, where n is the pseudo-anomaly size; marking the positions of these n points in the point cloud and creating a pseudo-anomaly label g. t ∈R N×1 , where N is the original point cloud size, the n points in the pseudo-anomaly region have a value of 1, and the rest of the region has a value of 0.
[0013] In one embodiment of the present invention, step (1.2) specifically includes: defining the generated pseudo-anomaly types as two kinds, depression and convexity, with each type having an equal probability of generation; the pseudo-anomaly size n is adaptively determined by the original point cloud size N, and follows the following distribution. The maximum value dm for depressions or convexities is determined based on the shape of the 3D bounding box of the object's point cloud. First, calculate the length, width, and height of the bounding box, and sort them from smallest to largest as s, m, and l. Then, calculate dm according to the following formula:
[0014] k ~ U(0.08, 0.15)
[0015]
[0016] This method avoids generating overly distorted pseudo-anomalies. The degree of concavity or convexity (d) at each point in the pseudo-anomaly is calculated using a Gaussian kernel function. First, the geometric center c2 of the n-neighborhood of point c1 is calculated. Then, a point is randomly selected within the n-neighborhood as c3. The coordinates of c2 and c3 are used to calculate c4 in a 1:3 ratio. c4 is then used as the mean μ of the Gaussian kernel function to achieve anomalies with more diverse shapes, where σ∈U(1,5). After calculating the degree of convexity or concavity (d) at each point within the n-neighborhood, it is scaled to [0, dm].
[0017] In one embodiment of the present invention, step (1.3) specifically includes: defining a lifting distance of d for each point in the pseudo-anomaly region, and a lifting distance of 0 for the remaining points; using Poisson image editing at the lifting distance level to make the edge of the pseudo-anomaly fit better with the surface of the object; and performing forward or reverse translation in the normal direction of the points in the pseudo-anomaly region according to the anomaly type to achieve a raised or recessed pseudo-anomaly.
[0018] In one embodiment of the present invention, step (2) specifically includes:
[0019] (2.1) Prepare a Point Transformer network pre-trained on a point cloud dataset as the teacher feature extraction module. Any existing Point Transformer network architecture and pre-trained network parameters can be used. Use a Point Transformer pre-trained on the ShapeNet dataset using the Point-MAE method, fix the parameters, input pseudo-anomaly point clouds, and output the features of layers 3, 7, and 11. After layer standardization, the features are concatenated along the feature dimension to serve as the teacher output features. in C represents the number of points in the point cloud after sampling from the farthest point, and C represents the feature dimension. The student feature extraction module adopts the same network structure as the teacher feature extraction module, only performing initialization without loading pre-trained weights. It takes a pseudo-anomaly point cloud as input and outputs the features of layers 3, 7, and 11. After layer standardization, these features are concatenated along the feature dimension to form the student output features.
[0020] (2.2) Obtain the anomaly label of the corresponding point by sampling the coordinates of the points after sampling from the farthest point in the pseudo-anomaly point cloud. Where 1 represents an outlier and 0 represents a non-outlier; the student feature extraction module is trained while the parameters of the teacher feature extraction module are fixed. The normal feature distribution from the teacher feature extraction module is then distilled into the student feature extraction module, and the loss function is:
[0021]
[0022] Where N is the number of test samples and C is the number of feature channels.
[0023] In one embodiment of the present invention, step (3) specifically includes:
[0024] (3.1) Subtract student output features F from teacher output features t -F s The input to the one-dimensional convolutional layer has C output channels, which are then combined with the student output features F. s The features are concatenated and then fed into another one-dimensional convolutional layer to reduce the dimensionality to C, resulting in the fused product.
[0025] (3.2) The fused features F d The input is a three-layer Point Transformer network with initialized parameters, followed by a two-layer multilayer perceptron. The first layer has C / 3 output channels, and the second layer has 2 output channels. The softmax function is used to transform the network into a system that distinguishes between abnormal and non-abnormal regions. Given a binary probability point cloud of 1 point, interpolate the remaining points in the original point cloud using inverse distance weighted interpolation, and output the binary probability point cloud of the original point cloud, pred∈R. N×2 Optimization is performed based on pseudo-anomaly labels, and the objective function is:
[0026] L d =λ*FL(pred,gt)+SmoothL1(pred,gt)
[0027] Where FL is focal loss, SmoothL1 is smooth L1 loss, gt is pseudo-anomaly label, and λ is the ratio balance coefficient between the two loss functions.
[0028] In one embodiment of the present invention, the value of λ is 5.
[0029] In one embodiment of the present invention, in step (4), the total loss function L = L ST +L d .
[0030] In one embodiment of the present invention, in step (2.1), C = 1152.
[0031] According to another aspect of the present invention, a point cloud anomaly segmentation device based on a pseudo-anomaly distillation discriminant network is also provided, characterized in that it includes at least one processor and a memory, the at least one processor and the memory are connected through a data bus, the memory stores instructions that can be executed by the at least one processor, and the instructions, after being executed by the processor, are used to complete the point cloud anomaly segmentation method based on the pseudo-anomaly distillation discriminant network.
[0032] In summary, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art:
[0033] This invention employs a point transformer to extract features based on the original ST architecture, expanding the receptive field to a complete point cloud to ensure that larger abnormal regions can be completely segmented. Through a carefully designed pseudo-anomaly generation module, it can simulate anomalies when real anomalies are scarce in industrial scenarios. The generated pseudo-anomalies are combined with the discriminative network to simulate supervised training, resulting in more refined anomaly segmentation results and significantly reducing false detections of normal regions. In addition, this invention performs end-to-end training, resulting in fast inference speed, low storage space consumption, and high industrial anomaly segmentation accuracy. Attached Figure Description
[0034] Figure 1 This is a flowchart illustrating the point cloud anomaly segmentation method based on a pseudo-anomaly distillation discrimination network in an embodiment of the present invention. Detailed Implementation
[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0036] To address the problems existing in the prior art, this invention provides a point cloud anomaly segmentation method based on a pseudo-anomaly distillation discrimination network, comprising the following steps:
[0037] (1) Simulate anomalies on the training point cloud using the pseudo-anomaly generation module. The specific steps include:
[0038] (1.1) Set the pseudo-anomaly location parameters. Determine the center point for pseudo-anomaly generation, ensuring it doesn't approach edge regions. Map the point cloud onto a 2D image, using a 3×3 morphological dilation operation to fill unmapped areas in the 2D image. Calculate the positions of object edge points. Randomly select a point in the object's foreground region. If the point's 4n neighborhood does not contain any object edge points, set it as center point c1. The pseudo-anomaly region is defined as the n-neighborhood of point c1, where n is the pseudo-anomaly size. Mark the positions of these n points in the point cloud and create pseudo-anomaly labels gt∈R. N×1 , where N is the original point cloud size, the n points in the pseudo-anomaly region have a value of 1, and the rest of the region has a value of 0.
[0039] (1.2) Set the shape parameters of the pseudo-anomalies. Define two types of pseudo-anomalies: depressions and convexities. Each type has an equal probability of being generated. The size n of the pseudo-anomaly is adaptively determined by the original point cloud size N and follows the distribution as follows: The maximum value dm for depressions or convexities is determined based on the shape of the 3D bounding box of the object's point cloud. First, calculate the length, width, and height of the bounding box, and sort them from smallest to largest as s, m, and l. Then, calculate dm according to the following formula:
[0040] k ~ U(0.08, 0.15)
[0041]
[0042] This method avoids generating overly distorted pseudo-anomalies. The degree of depression or convexity d at each point in the pseudo-anomaly is calculated using a Gaussian kernel function. First, the geometric center c2 of the n-neighborhood of point c1 is calculated. Then, a point is randomly selected in the n-neighborhood as c3. The coordinates of c2 and c3 are used to calculate c4 in a 1:3 ratio. c4 is used as the mean μ of the Gaussian kernel function to achieve anomalies with more shapes, σ∈U(1,5). After the degree of convexity or depression d of each point in the n-neighborhood is calculated, it is scaled to [0, dm].
[0043]
[0044] (1.3) For each point in the pseudo-anomaly region, define the lifting distance as d, and the lifting distance of the other points as 0. Use Poisson image editing on the lifting distance level to make the edge of the pseudo-anomaly fit better with the surface of the object. In the normal direction of the pseudo-anomaly region points, perform forward or reverse translation according to the anomaly type to achieve convex or concave pseudo-anomalies.
[0045] (2) Use the generated pseudo-anomaly point cloud for teacher-student network training. The specific steps include:
[0046] (2.1) Prepare a Point Transformer network pre-trained on a point cloud dataset as the teacher feature extraction module. Any existing Point Transformer network architecture and pre-trained network parameters can be used. In this invention, a Point Transformer pre-trained on the ShapeNet dataset using the Point-MAE method is used with fixed parameters. The input is a pseudo-anomaly point cloud, and the output is the features of layers 3, 7, and 11. After layer standardization, the features are concatenated along the feature dimension to serve as the teacher output features. in The number of points in the point cloud after sampling from the farthest point is, in this invention C is the feature dimension, which is 1152 in this invention. The student feature extraction module adopts the same network structure as the teacher feature extraction module, only undergoing initialization without loading pre-trained weights. It inputs pseudo-anomaly point clouds and outputs features from layers 3, 7, and 11. After layer standardization, these features are concatenated along the feature dimension to form the student output features.
[0047] (2.2) Obtain the anomaly label of the corresponding point by sampling the coordinates of the points after sampling from the farthest point in the pseudo-anomaly point cloud. Where 1 represents an outlier and 0 represents a non-outlier. The student feature extraction module is trained while the parameters of the teacher feature extraction module are fixed. The module is trained using features from normal points in the output features, and the normal feature distribution from the teacher feature extraction module is distilled into the student feature extraction module. The loss function is:
[0048]
[0049] Where N is the number of test samples and C is the number of feature channels.
[0050] (3) Train the discriminant network using student output features and teacher output features. The specific steps include:
[0051] (3.1) Subtract student output features F from teacher output features t -F s The input to the one-dimensional convolutional layer has C output channels, which are then combined with the student output features F. s The features are concatenated and then fed into another one-dimensional convolutional layer to reduce the dimensionality to C, resulting in the fused product.
[0052] (3.2) The fused features F dThe input is a three-layer Point Transformer network with initialized parameters, followed by a two-layer multilayer perceptron. The first layer has C / 3 output channels, and the second layer has 2 output channels. The softmax function is used to transform the network into a system that distinguishes between abnormal and non-abnormal regions. Given a binary probability point cloud of 1 point, interpolate the remaining points in the original point cloud using inverse distance weighted interpolation, and output the binary probability point cloud of the original point cloud, pred∈R. N×2 Optimization is performed based on pseudo-anomaly labels, and the objective function is:
[0053] L d =λ*FL(pred,gt)+SmoothL1(pred,gt)
[0054] Where FL is focal loss, SmoothL1 is smooth L1 loss, gt is pseudo-anomaly label, and λ is the ratio balance coefficient of the two loss functions, which is 5 in this invention.
[0055] (4) During training, the network parameters in the teacher feature extraction module are fixed, and the student extraction module and the discriminant network are trained end-to-end. The total loss function is L = L ST +L d .
[0056] (5) During the testing phase, the test point cloud is input into the teacher feature extraction module and the student feature extraction module respectively to obtain the teacher output feature and the student output feature. The difference between the teacher output feature and the student output feature is then input together with the student output feature into the discriminant network to obtain the binary probability point cloud pred∈R. N×2 The last dimension is taken as the anomaly score point cloud. The larger the anomaly score of a point in the point cloud, the greater the probability that the point is an anomaly. Inverse distance weighting is used to interpolate into the original point cloud, and the anomaly segmentation result is obtained according to the preset threshold.
[0057] Furthermore, the present invention also provides a point cloud anomaly segmentation device based on a pseudo-anomaly distillation discriminant network, characterized in that it includes at least one processor and a memory, the at least one processor and the memory are connected through a data bus, the memory stores instructions that can be executed by the at least one processor, and the instructions, after being executed by the processor, are used to complete the point cloud anomaly segmentation method based on the pseudo-anomaly distillation discriminant network.
[0058] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A point cloud anomaly segmentation method based on a pseudo-anomaly distillation discriminant network, characterized in that, Includes the following steps: (1) Use the pseudo-anomaly generation module to simulate anomalies on the training point cloud; This includes (1.1) setting the pseudo-anomaly position parameters; (1.2) setting the pseudo-anomaly shape parameters; and (1.3) implementing pseudo-anomalies with protrusions or depressions. (2) Use the generated pseudo-anomaly point cloud for teacher-student network training; (3) Train the discriminant network using student output features and teacher output features; (4) During training, fix the network parameters in the teacher feature extraction module and train the student extraction module and discrimination network end-to-end; (5) During the testing phase, the test point cloud is input into the teacher feature extraction module and the student feature extraction module respectively to obtain the teacher output feature and the student output feature. The difference between the teacher output feature and the student output feature is calculated and then input into the discriminant network together with the student output feature to obtain a binary probability point cloud. The last dimension is taken as the anomaly score point cloud. The larger the anomaly score value of a point in the point cloud, the greater the probability that the point is an anomaly. The inverse distance weight is used to interpolate into the original point cloud, and the anomaly segmentation result is obtained according to the preset threshold.
2. The point cloud anomaly segmentation method based on a pseudo-anomaly distillation discriminant network as described in claim 1, characterized in that, Step (1.1) specifically includes: determining the center point for pseudo-anomaly generation so that it cannot be close to the edge region; mapping the point cloud onto a 2D image, filling the unmapped areas in the 2D image using a 3×3 morphological dilation operation, calculating the positions of object edge points, randomly selecting points in the foreground region of the object, and setting the center point c1 if the 4n neighborhood of the point does not contain any object edge points, defining the pseudo-anomaly region as the n-neighborhood of point c1, where n is the pseudo-anomaly size; marking the positions of these n points in the point cloud and creating pseudo-anomaly labels gt∈R. N×1 , where N is the original point cloud size, the n points in the pseudo-anomaly region have a value of 1, and the rest of the region has a value of 0.
3. The point cloud anomaly segmentation method based on a pseudo-anomaly distillation discriminant network as described in claim 1 or 2, characterized in that, Step (1.2) specifically includes: defining two types of generated pseudo-anomalies, depression and convexity, with each type having an equal probability of generation; the pseudo-anomaly size n is adaptively determined by the original point cloud size N, and follows the distribution as follows: The maximum value dm for depressions or convexities is determined based on the shape of the 3D bounding box of the object's point cloud. First, calculate the length, width, and height of the bounding box, and sort them from smallest to largest as s, m, and l. Then, calculate dm according to the following formula: k~U(0.08,0.15) This method avoids generating overly distorted pseudo-anomalies. The degree of concavity or convexity (d) at each point in the pseudo-anomaly is calculated using a Gaussian kernel function. First, the geometric center c2 of the n-neighborhood of point c1 is calculated. Then, a point is randomly selected within the n-neighborhood as c3. The coordinates of c2 and c3 are used to calculate c4 in a 1:3 ratio. c4 is then used as the mean μ of the Gaussian kernel function to achieve anomalies with more diverse shapes, σ∈U(1,5). After calculating the degree of convexity or concavity (d) at each point within the n-neighborhood, it is scaled to [0,dm].
4. The point cloud anomaly segmentation method based on a pseudo-anomaly distillation discriminant network as described in claim 1 or 2, characterized in that, The specific steps (1.3) include: defining the lifting distance as d for each point in the pseudo-anomaly region, and the lifting distance of the remaining points as 0; using Poisson image editing at the lifting distance level to make the pseudo-anomaly edge better fit the object surface; and performing forward or reverse translation in the normal vector direction of the pseudo-anomaly region points according to the anomaly type to achieve convex or concave pseudo-anomalies.
5. The point cloud anomaly segmentation method based on a pseudo-anomaly distillation discriminant network as described in claim 1 or 2, characterized in that, Step (2) specifically includes: (2.1) Prepare a Point Transformer network pre-trained on a point cloud dataset as the teacher feature extraction module. Any existing Point Transformer network architecture and pre-trained network parameters can be used. Use a Point Transformer pre-trained on the ShapeNet dataset using the Point-MAE method, fix the parameters, input pseudo-anomaly point clouds, and output the features of layers 3, 7, and 11. After layer standardization, the features are concatenated along the feature dimension to serve as the teacher output features. in C represents the number of points in the point cloud after sampling from the farthest point, and C represents the feature dimension. The student feature extraction module adopts the same network structure as the teacher feature extraction module, only performing initialization without loading pre-trained weights. It takes a pseudo-anomaly point cloud as input and outputs the features of layers 3, 7, and 11. After layer standardization, these features are concatenated along the feature dimension to form the student output features. (2.2) Obtain the anomaly label of the corresponding point by sampling the coordinates of the points after sampling from the farthest point in the pseudo-anomaly point cloud. Where 1 represents an outlier and 0 represents a non-outlier; the student feature extraction module is trained while the parameters of the teacher feature extraction module are fixed. The normal feature distribution from the teacher feature extraction module is then distilled into the student feature extraction module, and the loss function is: Where N is the number of test samples and C is the number of feature channels.
6. The point cloud anomaly segmentation method based on a pseudo-anomaly distillation discriminant network as described in claim 1 or 2, characterized in that, Step (3) specifically includes: (3.1) Subtract student output features F from teacher output features t -F s The input to the one-dimensional convolutional layer has C output channels, which are then combined with the student output features F. s The features are concatenated and then fed into another one-dimensional convolutional layer to reduce the dimensionality to C, resulting in the fused product. (3.2) The fused features F d The input is a three-layer Point Transformer network with initialized parameters, followed by a two-layer multilayer perceptron. The first layer has C / 3 output channels, and the second layer has 2 output channels. The softmax function is used to transform the network into a system that distinguishes between abnormal and non-abnormal regions. Given a binary probability point cloud of 1 point, interpolate the remaining points in the original point cloud using inverse distance weighted interpolation, and output the binary probability point cloud of the original point cloud, pred∈R. N×2 Optimization is performed based on pseudo-anomaly labels, and the objective function is: L d =λ*FL(pred,gt)+SmoothL1(pred,gt) Where FL is focal loss, SmoothL1 is smooth L1 loss, gt is pseudo-anomaly label, and λ is the ratio balance coefficient between the two loss functions.
7. The point cloud anomaly segmentation method based on a pseudo-anomaly distillation discriminant network as described in claim 6, characterized in that, The value of λ is 5.
8. The point cloud anomaly segmentation method based on a pseudo-anomaly distillation discriminant network as described in claim 1 or 2, characterized in that, In step (4), the total loss function L = L ST +L d .
9. The point cloud anomaly segmentation method based on a pseudo-anomaly distillation discriminant network as described in claim 5, characterized in that, In step (2.1), C = 1152.
10. A point cloud anomaly segmentation device based on a pseudo-anomaly distillation discrimination network, characterized in that, The method includes at least one processor and a memory, which are connected via a data bus. The memory stores instructions that can be executed by the at least one processor. After being executed by the processor, the instructions are used to complete the point cloud anomaly segmentation method based on the pseudo-anomaly distillation discrimination network as described in any one of claims 1-9.