A point cloud anomaly detection method and device considering prototype score correction

Abstract

The present application belongs to the technical field of three-dimensional detection, and discloses a point cloud anomaly detection method and equipment considering prototype score correction, which comprises the following steps: (1) synthesizing abnormal data based on normal point cloud data; (2) training a lightweight point cloud feature extractor based on the synthesized abnormal data; (3) extracting features of normal point cloud data and features of synthesized abnormal data based on the lightweight point cloud feature extractor to form a normal feature set and an abnormal feature set, respectively, which are denoted as a normal prototype set and an abnormal prototype set; (4) calculating a traditional prototype anomaly score by using the minimum distance of test data relative to the normal prototype, and then correcting the traditional prototype anomaly score by using the minimum distance of test data relative to the abnormal prototype to obtain a calibrated anomaly score, thereby realizing point cloud anomaly detection. The present application improves the accuracy of anomaly recognition.

Description

Technical Field

[0001] This invention belongs to the field of 3D detection technology, and more specifically, relates to a point cloud anomaly detection method and device that considers prototype fraction correction. Background Technology

[0002] In today's industrial production, ensuring product quality is paramount, making the application of industrial vision inspection technology particularly important. Industrial vision inspection technology uses automated vision systems to inspect products, identifying defects, flaws, or non-conforming items to ensure product quality meets standards. In recent years, anomaly detection technology has been widely used in the surface defect quality inspection of industrial products. In industrial vision inspection, by analyzing product surface images, anomaly detection can identify various defects such as scratches, dents, and impurities, thus eliminating non-conforming products early in the production process. However, while image anomaly detection technology performs well in many aspects, it has limitations when handling certain complex scenarios. For example, for objects with indistinct surface textures, images struggle to accurately capture their surface details, causing image anomaly detection to fail. In contrast, point cloud data can provide three-dimensional geometric information of objects, making detection more comprehensive and accurate.

[0003] Current point cloud anomaly detection methods mainly rely on prototype-based unsupervised detection methods, where feature extractors map test points into a distinguishable feature space and calculate anomaly scores based on the minimum distance between a test point and its nearest normal prototype (called the normal prototype distance). However, relying solely on the normal prototype distance is insufficient; ideally, test points with similar normal prototype distances but different distances from the nearest anomalous prototype should receive different anomaly scores. Summary of the Invention

[0004] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a point cloud anomaly detection method and device that considers prototype fraction correction, which aims to solve the problem of low accuracy of existing point cloud anomaly identification.

[0005] To achieve the above objectives, according to one aspect of the present invention, a point cloud anomaly detection method considering prototype anomaly score correction is provided, the method comprising the following steps:

[0006] (1) Synthetic anomalous data is obtained by simulation based on normal point cloud data;

[0007] (2) Training a lightweight point cloud feature extractor based on synthetic anomalous data;

[0008] (3) Based on the lightweight point cloud feature extractor, extract the features of normal point cloud data and synthesize the features of abnormal data to form a set of normal features and a set of abnormal features, respectively denoted as normal prototype set and abnormal prototype set.

[0009] (4) The abnormal score of the traditional prototype is calculated by using the minimum distance of the test data relative to the normal prototype. Then, the abnormal score of the traditional prototype is corrected by using the minimum distance of the test data relative to the abnormal prototype to obtain the calibration abnormal score and realize point cloud anomaly detection.

[0010] Furthermore, there are three ways to generate synthetic anomalous data: local stretching, local rotation, and local randomization.

[0011] Furthermore, the anomaly generation process includes three steps: reference point selection, neighborhood point search, and local point transformation. For inclusion A normal point cloud of points, from Randomly select a reference point Then, the nearest neighbor search algorithm is used to find the distance. Recent k Each point forms a neighborhood. Next, regarding Perform a transformation to generate a synthetic anomaly. ,make Represented as a reference point The i Neighboring points, This represents the transformed point.

[0012] Furthermore, local stretching: stretching the neighborhood Points within the reference point direction of the normal vector Stretch, the total length of the stretch is D dir represents the abnormal properties formed. When dir=1, local stretching forms a bulge, and when dir=-1, local stretching forms a depression.

[0013] .

[0014] Furthermore, local rotation: using a rotation matrix. For the neighboring area The points within the range are transformed by selecting an angle range along three rotation angles of the three coordinate values, with the rotation point around which the transformation revolves serving as the reference point. , .

[0015] Furthermore, local randomization: [representing the neighborhood / neighborhood / etc.] The points within are randomly replaced with points distributed in a Gaussian distribution.

[0016] Furthermore, the lightweight feature extractor contains two point transformer layers, followed by a segmentation head consisting of a multilayer perceptron; the combined model is trained end-to-end using IoU loss and Focal loss, with the loss function... for:

[0017]

[0018] in Indicates feature extractor, This indicates the split header.

[0019] Furthermore, the formula for calculating the calibration anomaly score is as follows:

[0020]

[0021] in Represents test data The probability of an anomaly. The calibration prototype score represents the distance of the test data from the nearest anomalous prototype. It consists of two parts: the difference term. Among them, the balance factor Used for weighted averaging of two uncertainties; The ratio representing uncertainty is called relative uncertainty.

[0022] The present invention also provides a point cloud anomaly detection system that considers prototype fraction correction. The system includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it performs the point cloud anomaly detection method that considers prototype fraction correction as described above.

[0023] The present invention also provides a computer-readable storage medium storing machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the point cloud anomaly detection method considering prototype fraction correction as described above.

[0024] In summary, compared with the prior art, the point cloud anomaly detection method and device considering prototype fraction correction provided by the present invention have the following beneficial effects:

[0025] 1. Based on the lightweight point cloud feature extractor, features of normal point cloud data and features of synthetic anomalous data are extracted to form sets of normal features and sets of anomalous features, respectively denoted as normal prototype set and anomalous prototype set. The anomalous score of the traditional prototype is calculated using the minimum distance of the test data relative to the normal prototype. Then, the anomalous score of the traditional prototype is corrected using the minimum distance of the test data relative to the anomalous prototype. At the same time, the absolute distance and relative distance between the test data and the normal prototype and the anomalous prototype are considered to provide a more accurate anomalous score.

[0026] 2. Synthetic anomalous data is obtained by simulating normal point cloud data, which reduces the cost of collecting anomalous points.

[0027] 3. There are three ways to generate synthetic anomalous data: local stretching, local rotation, and local randomization. Three anomalous generation strategies are introduced to generate diverse and realistic synthetic anomalous data to support unsupervised anomaly detection methods. These synthetic anomalous data are further used to train a lightweight point cloud feature extractor, which is specifically designed to extract local geometric information for anomaly detection.

[0028] 4. By training on synthetic anomalous data, the feature extractor is encouraged to gather local geometric information, enabling the segmentation head to effectively identify normal and anomalous points. Attached Figure Description

[0029] Figure 1 This is a flowchart of a point cloud anomaly detection method considering prototype fraction correction provided by the present invention;

[0030] Figure 2 This is a flowchart of the abnormal data synthesis process;

[0031] Figure 3 It is a composite visualization of abnormal data;

[0032] Figure 4 These are comparison images of the prototype distance before and after correction, where (1) is for the case of considering only the normal prototype distance, and (2) is for the case of considering both normal and abnormal prototype distances. Detailed Implementation

[0033] 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.

[0034] Please see Figure 1This invention provides a point cloud anomaly detection method that considers prototype anomaly score correction. The method proposes a traditional prototype anomaly score correction approach, improving the score by combining the distances between normal and anomalous prototypes. It considers both the absolute and relative distances between the test data and both normal and anomalous prototypes, providing a more accurate anomaly score. Furthermore, since collecting real anomaly points is too costly, three anomaly generation strategies are introduced to generate diverse and realistic synthetic anomaly data to support unsupervised anomaly detection methods. These synthetic anomalies are further used to train a lightweight point cloud feature extractor specifically designed to extract local geometric information for anomaly detection.

[0035] In one implementation, firstly, an anomaly generation strategy is proposed to generate diverse and realistic synthetic anomaly data; then, a lightweight point cloud feature extractor is trained using the synthetic anomalies; next, features of normal training data and synthetic anomaly data are extracted using the point cloud feature extractor, and normal prototype sets and anomaly prototype sets are established; finally, prototype anomaly score correction is proposed, which considers both the absolute and relative distances between the test data and the normal and anomaly prototypes, providing more accurate anomaly scores.

[0036] The detection method mainly includes the following steps:

[0037] Step 1: Simulate and obtain synthetic anomalous data based on normal point cloud data.

[0038] Point cloud anomalies typically manifest as localized changes; therefore, this implementation devises three anomaly generation methods that introduce point cloud anomalies through local geometric perturbations: local stretching, local rotation, and local randomization. For example... Figure 2 As shown, the anomaly generation process includes three steps: reference point selection, neighborhood point search, and local point transformation. Let... For inclusion A normal point cloud of points. First, from Randomly select a reference point Then, the nearest neighbor search algorithm (KNN) is used to find the distance. Recent k Each point forms a neighborhood. Next, regarding Perform a transformation to generate a synthetic anomaly. ,make Represented as a reference point The i Neighboring points, This represents the transformed point. The transformation is defined as follows:

[0039] Local stretching: stretching the neighborhood Points within the reference point direction of the normal vector Stretch, the total length of the stretch is D dir represents the abnormal property formed. When dir=1, local stretching forms a bulge, and when dir=-1, local stretching forms a depression.

[0040]

[0041] Local rotation: using a rotation matrix For the neighboring area The transformation is performed on points within the coordinate system, selecting a small range (e.g., 0-30°) around which the rotation is performed. The point around which the rotation is performed is used as the reference point. .

[0042]

[0043] Local randomization: This involves the neighborhood... Points within the range are randomly replaced with points distributed by a Gaussian distribution. Gauss is the generating function of the Gaussian distribution, where Gauss is the distribution's inequality. The mean and standard deviation are given. Generally selected as the neighborhood 1 / 3 of the maximum distance between midpoints:

[0044]

[0045] The resulting abnormal data like Figure 3 As shown.

[0046] Step 2: Train a lightweight point cloud feature extractor based on synthetic anomalous data.

[0047] The above synthetic anomalies introduce regions with unique shapes and point distributions compared to normal point clouds, and can obtain point-level anomaly labels. y If the first point cloud i These points are outliers. =1, otherwise =0. By flexibly selecting reference points, this method can generate various realistic anomalies that approximate real anomaly distributions. Existing point cloud extractors rely on pre-training with publicly available real-world data, but these models are not optimized for anomaly detection and are often inefficient. Therefore, this implementation utilizes various generated realistic synthetic anomalies and their corresponding point-level anomaly labels to train a lightweight point cloud feature extractor specifically tailored for anomaly detection from scratch.

[0048] Local geometric features are particularly valuable for anomaly detection in point clouds. Therefore, this implementation introduces a lightweight feature extractor that is simpler and shallower than traditional backbones. This lightweight feature extractor consists of only two pointtransformer layers, focusing on capturing core geometric information. A segmentation head composed of a multilayer perceptron (MLP) is added after the feature extractor, which directly predicts the anomaly probability. The combined model is trained end-to-end using IoU loss and Focal loss. Loss function for:

[0049]

[0050] in Indicates feature extractor, This represents the segmentation head. By training on synthetic anomalous data, the feature extractor is encouraged to aggregate local geometric information, enabling the segmentation head to effectively identify normal and anomalous points.

[0051] Step 3: Based on the lightweight point cloud feature extractor, extract features from normal point cloud data and synthesize features from abnormal data to form sets of normal features and sets of abnormal features, respectively denoted as normal prototype set and abnormal prototype set.

[0052] After obtaining the feature extractor, it is necessary to extract normal training data. Features and synthetic anomalous data Features ,in i A sequence representing normal or abnormal point clouds. Normal features. for:

[0053]

[0054] Abnormal characteristics It is obtained from synthetic anomalous data using a feature extractor:

[0055]

[0056] Next, the set of normal features is constructed by taking the union of all extracted normal features, denoted as the normal prototype set. :

[0057]

[0058] Take the union of all extracted anomalous features and construct a set of anomalous features, denoted as the anomalous prototype set. :

[0059]

[0060] Step 4: Calculate the anomaly score of the traditional prototype using the minimum distance between the test data and the normal prototype. Then, use the minimum distance between the test data and the abnormal prototype to correct the anomaly score of the traditional prototype to obtain the calibration anomaly score, thus realizing point cloud anomaly detection.

[0061] First, anomaly scores are calculated using the minimum distance between test data and normal prototypes. This can be reinterpreted from the perspective of probability and information theory. Specifically, the set of normal prototypes can be viewed as a Gaussian isotropic mixture distribution. :

[0062]

[0063] in For the characteristics of the test data, This represents the total number of normal prototypes. Representing the i The Gaussian distribution containing the normal prototypes, which is based on... As the mean, with For covariance. For the first i The weights of each prototype distribution in contributing to the outlier score. Approximately considering only the probability distribution of the most recent prototype in the set:

[0064]

[0065] definition The mean of the normal prototype closest to the test data, when satisfying the independent Gaussian distribution assumption, is... for Its covariance can be defined as ,but It can be further written as:

[0066]

[0067] outlier scores in test data Uncertainty can be calculated using self-information:

[0068]

[0069] in and and Consistently related, The anomaly score represents the distance between the test data feature and the nearest normal prototype; therefore, the anomaly score for different test data only varies with... Change, therefore =0, If the value is 1, then:

[0070]

[0071] However, this measure of normal uncertainty cannot fully capture the possibility of anomalies. For example... Figure 4 As shown, and They have the same minimum normal prototype distance, but because of... compared to, The distance to the anomalous prototype is small, i.e. It is closer to the outlier. Therefore Theoretically there should be more A larger anomaly score. Therefore, the traditional prototype anomaly score is calibrated by introducing anomaly prototype distance. Intuitively, when a test point is closer to the anomaly prototype, it is more likely to be an anomaly. This leads to a decrease in uncertainty in the anomaly distribution, and vice versa. Conversely, the higher the uncertainty in the normal distribution, the higher the anomaly score. Therefore, the formula for calibrating the anomaly score is:

[0072]

[0073] in Represents test data The probability of an anomaly. The calibration prototype score represents the distance of the test data from the nearest anomalous prototype. It consists of two parts: the difference term. Among them, the balance factor Used for weighted averaging of two uncertainties; The ratio representing uncertainty is called relative uncertainty.

[0074] The present invention also provides a point cloud anomaly detection system that considers prototype fraction correction. The system includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it performs the point cloud anomaly detection method that considers prototype fraction correction as described above.

[0075] The present invention also provides a computer-readable storage medium storing machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the point cloud anomaly detection method considering prototype fraction correction as described above.

[0076] 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 detection method considering prototype anomaly score correction, characterized in that, The method includes the following steps: (1) Synthetic anomalous data is obtained by simulation based on normal point cloud data; (2) Training a lightweight point cloud feature extractor based on synthetic anomalous data; (3) Based on the lightweight point cloud feature extractor, extract the features of normal point cloud data and synthesize the features of abnormal data to form a set of normal features and a set of abnormal features, respectively denoted as normal prototype set and abnormal prototype set. (4) The abnormal score of the traditional prototype is calculated by using the minimum distance of the test data relative to the normal prototype, and then the abnormal score of the traditional prototype is corrected by using the minimum distance of the test data relative to the abnormal prototype to obtain the calibration abnormal score and realize point cloud anomaly detection. The formula for calculating the calibration anomaly score is: in Represents test data The probability of an anomaly. The calibration prototype score represents the distance of the test data from the nearest anomalous prototype. It consists of two parts: the difference term. Among them, the balance factor Used for weighted averaging of two uncertainties; The ratio representing uncertainty is called relative uncertainty.

2. The point cloud anomaly detection method considering prototype anomaly score correction as described in claim 1, characterized in that: There are three ways to generate synthetic anomalous data: local stretching, local rotation, and local randomization.

3. The point cloud anomaly detection method considering prototype anomaly score correction as described in claim 2, characterized in that: The anomaly generation process includes three steps: reference point selection, neighborhood point search, and local point transformation. For inclusion A normal point cloud of points, from Randomly select a reference point Then, the nearest neighbor search algorithm is used to find the distance. Recent k Each point forms a neighborhood. Next, regarding Perform a transformation to generate a synthetic anomaly. ,make Represented as a reference point The i Neighboring points, This represents the transformed point.

4. The point cloud anomaly detection method considering prototype anomaly score correction as described in claim 3, characterized in that: Local stretching: stretching the neighborhood Points within the reference point direction of the normal vector Stretch, the total length of the stretch is D dir represents the abnormal properties formed. When dir=1, local stretching forms a bulge, and when dir=-1, local stretching forms a depression. 。 5. The point cloud anomaly detection method considering prototype anomaly score correction as described in claim 3, characterized in that: Local rotation: using a rotation matrix For the neighboring area The points within the range are transformed by selecting an angle range along three rotation angles of the three coordinate values, with the rotation point around which the transformation revolves serving as the reference point. , .

6. The point cloud anomaly detection method considering prototype anomaly score correction as described in claim 3, characterized in that: Local randomization: This involves the neighborhood... The points within are randomly replaced with points distributed in a Gaussian distribution.

7. The point cloud anomaly detection method considering prototype anomaly score correction as described in any one of claims 1-6, characterized in that: The lightweight feature extractor consists of two point transformer layers, followed by a segmentation head composed of multilayer perceptrons. The combined model is trained end-to-end using IoU loss and Focal loss, with the loss function... for: in Indicates feature extractor, This indicates the split header.

8. A point cloud anomaly detection system considering prototype fraction correction, characterized in that: The system includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it performs the point cloud anomaly detection method considering prototype fraction correction as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores machine-executable instructions that, when invoked and executed by a processor, cause the processor to implement the point cloud anomaly detection method considering prototype fraction correction as described in any one of claims 1-7.

Images