Hash coding neural field based multi-class unified three-dimensional anomaly detection method

Abstract

This invention proposes a multi-class unified 3D anomaly detection method based on hash-coded neural fields, belonging to the fields of computer vision and industrial quality inspection technology. It solves the problem that existing 3D anomaly detection methods struggle to simultaneously handle multi-class unified modeling, pose change adaptation, and anomaly repair. The method includes: 1. Acquiring multi-class normal point cloud samples and performing pose normalization on the samples; 2. Extracting multi-scale geometric features from the normalized point clouds and training a class prediction model; 3. Constructing a class-conditionalized multi-resolution hash-coded neural field and training a signed distance field model using normal samples; 4. Performing pose normalization, class prediction, and prototype registration on the point cloud to be tested; 5. Querying the trained neural field to complete point-by-point anomaly localization and target-level anomaly determination; 6. Reconstructing the surface of the anomaly region based on the zero isosurface of the signed distance field to obtain the repaired point cloud result. This invention can achieve multi-class point cloud anomaly detection, localization, and repair under a single model, and has advantages such as good pose robustness, strong fine-grained geometric expression capability, and high detection efficiency.

Description

Technical Field

[0001] This invention belongs to the fields of computer vision and industrial quality inspection technology, specifically relating to a multi-category unified three-dimensional anomaly detection method based on hash-coded neural fields. Background Technology

[0002] In recent years, 3D anomaly detection has received widespread attention in industrial quality inspection. Its purpose is to establish the normal geometric distribution of a target using normal samples, thereby identifying abnormal regions in the sample under test. Compared with 2D anomaly detection, 3D point clouds can provide more direct geometric structural information, thus having a significant advantage in detecting structural defects such as dents, defects, and deformations.

[0003] Existing 3D anomaly detection methods mainly include feature embedding methods and shape reconstruction methods. Feature embedding methods detect anomalies by comparing the feature differences between the test sample and normal samples. Although they have good detection capabilities, they usually rely on discrete feature libraries and are difficult to support continuous surface reconstruction, thus making it difficult to further repair anomaly regions.

[0004] In recent years, some studies have attempted to apply implicit representations and pre-trained point cloud encoders to 3D anomaly detection tasks. For example, reconstruction methods based on signed distance fields achieve anomaly localization and repair by learning the continuous geometric distribution of normal objects. Although they can achieve good reconstruction results, in multi-class scenarios, when a single network models multiple geometries simultaneously, it is easily affected by the network's spectral bias, resulting in insufficient representation of fine-grained surface details.

[0005] In addition, some studies have extracted geometric features using pre-trained point cloud encoders and combined them with registration strategies to improve the comparability between different samples. While these methods improve detection accuracy to some extent, in real-world detection scenarios, the objects to be tested may appear in arbitrary poses, and residual pose errors can still lead to feature misalignment, thus affecting anomaly scoring results. This problem is even more pronounced in multi-class unified modeling scenarios.

[0006] In summary, existing methods still have shortcomings in terms of unified modeling across multiple categories, pose robustness, and fine-grained geometric representation, especially in simultaneously handling anomaly detection and anomaly repair. Therefore, ensuring pose consistency, improving fine-grained geometric representation capabilities, and simultaneously achieving anomaly detection and anomaly repair in multi-category scenarios remains a key challenge. Summary of the Invention

[0007] To address the aforementioned issues, this invention proposes a multi-category unified 3D anomaly detection method based on hash-encoded neural fields, which can achieve anomaly detection, anomaly localization, and anomaly repair for multi-category point clouds under a single model.

[0008] The method includes the following steps:

[0009] S1. The input point cloud is normalized by constructing a normalized coordinate system through centroid translation, covariance matrix decomposition and principal axis direction unification. The target point cloud is then mapped to the normalized coordinate system to reduce the impact of random object placement on subsequent feature extraction.

[0010] During the attitude normalization process, sign disambiguation is performed on the principal axis direction to ensure the stability of the normalized coordinate system.

[0011] The formula for calculating the homogeneous transformation matrix is ​​shown in equation (1):

[0012]

[0013] The input point cloud is subjected to coordinate transformation based on the homogeneous transformation matrix to obtain a normalized point cloud.

[0014] S2 inputs the normalized point cloud into the pre-trained point cloud encoder to extract multi-scale geometric features and outputs the category prediction results through a lightweight classification head;

[0015] The encoder is kept frozen during the training phase to preserve the general geometric priors obtained from pre-training.

[0016] S3: Based on the category prediction results, call the prototype set of the corresponding category and perform coarse-to-fine registration between the point cloud to be tested and the category prototype to obtain the aligned query point cloud.

[0017] The coarse-to-fine registration includes coarse registration based on feature correspondence and fine registration based on iterative optimization, in order to improve alignment accuracy under residual pose error.

[0018] The calculation formula for the coarse registration rigid body transformation is shown in Equation (2), and the calculation formula for the RANSAC interior point threshold is shown in Equation (3):

[0019]

[0020]

[0021] The coarse registration rigid body transformation is estimated using the RANSAC algorithm from the correspondence, and inlier points are selected based on an adaptive threshold. In equation (3), K is a subset of randomly sampled indices from the prototype downsampled point cloud. Let k be the average distance to the nearest neighbor. This is the multiplier coefficient.

[0022] The calculation formula for the fine registration incremental transformation is shown in equation (4):

[0023]

[0024] Using the coarse registration result as the initial value, fine registration is performed through the ICP algorithm to obtain the incremental transformation.

[0025] S4. The registered point cloud coordinates and category index are jointly input into the category-conditionalized multi-resolution hash-encoded neural field. Local geometric information at different scales is stored through a multi-resolution spatial hash grid, and the signed distance value of the corresponding coordinate position is output to establish an implicit signed distance field representation of normal objects.

[0026] The hash-encoded neural field uses category index and spatial coordinates as a four-dimensional input, enabling conditional modeling of objects of different categories in a unified model.

[0027] The formula for calculating the resolution of the multi-resolution grid is shown in equation (5):

[0028]

[0029] Set up L resolution levels, with the grid resolution of each level increasing geometrically.

[0030] The formula for calculating the hash index is shown in equation (6):

[0031]

[0032] Calculate the four-dimensional discrete grid coordinates at each level l. And the hash table is indexed by a hash function. In equation (6), This indicates a bitwise XOR operation. Let H be four large prime numbers, and H be the size of the hash table.

[0033] The formula for calculating the signed distance prediction value is shown in equation (7):

[0034]

[0035] A shallow MLP decoder is used to convert hash-coded features into signed distance predictions, where s represents the signed distance to the nearest surface; the decoder employs two hidden layers, each with 64 neurons.

[0036] S5, use the signed distance field to query the registered point cloud point by point, calculate the point-by-point anomaly score according to the degree of deviation of the point from the normal surface, and aggregate the anomaly scores in the local neighborhood to determine the target-level anomaly result.

[0037] The formula for calculating the point-by-point outlier score is shown in equation (8):

[0038]

[0039] After registration, the point cloud For each point p in the dataset, query the learned SDF field and calculate the point-by-point anomaly score accordingly.

[0040] To suppress the influence of isolated noise points, point-by-point anomaly scores within the local neighborhood are aggregated to obtain stable target-level judgment results.

[0041] The formula for calculating the target-level anomaly score is shown in equation (9):

[0042]

[0043] In equation (9), Let k be the k nearest neighbors of point p in the registration point cloud.

[0044] S6. For regions identified as abnormal, surface reconstruction is performed using the zero isosurface of the signed distance field to obtain the repaired point cloud result.

[0045] During the training phase, query points are sampled from the vicinity of the normal sample surface and within the bounding box space to constrain the expressive power of the signed distance field in the surface region and the blank space.

[0046] The calculation formula for the MAPE loss function is shown in equation (10):

[0047]

[0048] The mean absolute percentage error (MAPE) loss function is used to supervise SDF learning. In equation (10), B is the training batch, and d(x) is the true signed distance value. To prevent division by zero by a small constant, the objective function minimizes the relative error at non-surface points and constrains the SDF prediction to approach zero at the surface.

[0049] Through the above steps, this invention combines pose normalization, category prediction, category prototype registration, and hash implicit geometric modeling into the same detection framework, enabling integrated processing of anomaly detection, anomaly localization, and anomaly repair in multi-category scenarios.

[0050] Compared with existing technologies, this invention provides a multi-class unified three-dimensional anomaly detection method based on hash-coded neural fields, which can achieve more efficient and stable multi-class anomaly detection and has the following beneficial effects:

[0051] 1. Highly efficient multi-class unified detection: This invention enables a single detection framework to cover multiple categories of objects through category-conditional modeling, eliminating the need to train a separate model for each category and reducing deployment and maintenance costs.

[0052] 2. More stable pose consistency: This invention effectively reduces feature misalignment caused by pose changes by combining pose normalization with prototype registration, thereby improving the stability of anomaly detection in complex scenes.

[0053] 3. More refined geometric representation capability: This invention utilizes a multi-resolution hash-encoded neural field to store local geometric information at different scales, which can better preserve fine-grained surface details, thereby improving the accuracy of anomaly localization.

[0054] 4. Combining anomaly detection and repair: While outputting anomaly location results, this invention can also recover the surface of the anomaly region based on the zero isosurface of the signed distance field, providing a reference for industrial quality inspection and defect repair. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is the overall flowchart of this application;

[0057] Figure 2 This is a schematic diagram of the training phase provided by the present invention;

[0058] Figure 3 A schematic diagram of the reasoning stage provided by the present invention;

[0059] Figure 4 An anomaly localization result diagram provided by the present invention;

[0060] Figure 5 Anomaly repair comparison diagram provided by the present invention;

[0061] Figure 6 This is a schematic diagram of the results of the present invention under noise interference conditions. Detailed Implementation

[0062] The present invention provides a multi-class unified 3D anomaly detection method based on hash-coded neural fields. By employing pose normalization, prototype registration, and category-conditional multi-resolution hash-coded neural fields, it achieves anomaly detection, anomaly localization, and anomaly repair for multi-class point clouds. This method can address the randomness of pose in multi-class scenes while ensuring high-fidelity surface reconstruction. The specific implementation process is as follows.

[0063] The invention will be further described below with reference to specific embodiments.

[0064] Example 1:

[0065] like Figure 1 As shown, given input multi-class normal point cloud samples and their corresponding class labels, the point cloud is first normalized in pose. A homogeneous transformation matrix is ​​calculated through centroid translation and principal axis unification, mapping the point cloud to a normalized coordinate system. Next, the normalized point cloud is input into a pre-trained point cloud encoder to extract multi-scale geometric features, training a lightweight classification head and outputting class predictions. Simultaneously, a class-conditionalized multi-resolution hash-encoded neural field is constructed, and an implicit signed distance field (SDF) model is trained using normal samples. In the inference and detection phase, the target is also normalized and its class predicted. The corresponding healthy prototype is called to perform coarse-to-fine RANSAC and ICP registration. The aligned result is used to query the trained SDF and calculate point-by-point anomaly scores. The scores are then aggregated to complete target-level anomaly assessment. If an anomaly is determined, the damaged region is extracted and embedded using the SDF zero isosurface, yielding the final anomaly-corrected output.

[0066] Figure 2 This is a schematic diagram of the training phase provided by the present invention. It details the steps and principles of the network using the input point cloud for encoding, feature fusion, and updating the neural field hash table.

[0067] Figure 3 This is a schematic diagram of the inference stage provided for this invention. It illustrates the overall signal flow from unknown point cloud input to category prediction, then to prototype registration and extraction of anomalous heat values.

[0068] Figure 4 The image shows the anomaly localization results provided by this invention. It demonstrates the high-precision visual effect of locating defects in different categories of components by analyzing hash neural fields.

[0069] Figure 5 This invention provides a comparison diagram of anomaly repair. While achieving anomaly localization, directly utilizing implicit zero isosurfaces ensures both the smoothness and consistency of the repaired structure boundaries.

[0070] Figure 6This is a schematic diagram showing the results of the present invention under noise interference conditions. It can be seen that the method presented in this paper has strong robustness and competitiveness in dealing with systematic interference such as outlier scanning noise and attitude deviation.

[0071] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0072] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A multi-class unified three-dimensional anomaly detection and repair method based on category-conditional hashing encoded neural fields, characterized in that, Includes the following steps: S1, normalize the pose of the input point cloud to obtain the query point cloud in the normalized coordinate system; S2, input the query point cloud into the pre-trained point cloud encoder, extract multi-scale geometric features, and output the category prediction result based on the multi-scale geometric features; S3, based on the category prediction result, call the prototype set of the corresponding category, and perform coarse-to-fine registration between the query point cloud and at least one prototype point cloud in the prototype set in the same standard coordinate system to obtain the registered point cloud; S4, the three-dimensional coordinates of the registered point cloud and the category index corresponding to the category prediction result are combined as category condition input and sent to the multi-resolution hash-encoded neural field. The signed distance value of the corresponding position is output by a single decoding network to establish a unified implicit surface representation of normal objects for multiple categories. S5, calculate the point-by-point anomaly score based on the signed distance value, and aggregate the point-by-point anomaly scores in the local neighborhood to obtain the target-level anomaly judgment result; S6, when the target-level anomaly determination result is anomaly, the normal surface is reconstructed using the zero isosurface of the signed distance field to obtain the repaired point cloud result.

2. The multi-category unified three-dimensional anomaly detection and repair method as described in claim 1, characterized in that, S1 includes: translating the centroid of the input point cloud so that the centroid of the point cloud is located at the origin; calculating the covariance matrix of the translated point cloud and performing eigenvalue decomposition, and determining the principal axis direction according to the eigenvalue magnitude rule; constructing a homogeneous transformation matrix based on the principal axis direction to transform the input point cloud to a normalized coordinate system; disambiguating the principal axis symbols, and further eliminating any remaining ambiguities in the principal axis symbols or principal axis order after attitude normalization through prototype registration in step S3.

3. The multi-category unified three-dimensional anomaly detection and repair method as described in claim 1, characterized in that, S2 includes: sampling the query point cloud according to multiple spatial resolutions to generate a multi-scale point set; inputting the multi-scale point set into a pre-trained point cloud encoder to extract point-level features; keeping the pre-trained point cloud encoder frozen while training the category prediction results, and only training the lightweight classification head corresponding to the category prediction; performing pooling and concatenation on the features at each scale to obtain a global descriptor for category prediction.

4. The multi-category unified three-dimensional anomaly detection and repair method as described in claim 1, characterized in that, S3 includes: performing the same pose normalization as step S1 on the prototype point clouds in each category of prototype set, so that the prototype point clouds and the query point clouds are in the same normalized coordinate system; performing coarse registration based on the point cloud feature correspondence, and estimating the coarse registration rigid body transformation using a random sampling consensus algorithm; performing fine registration using the coarse registration rigid body transformation as the initial value and the iterative nearest point algorithm; and determining the target prototype point cloud and the corresponding final registration result from the prototype set according to the registration error.

5. The multi-category unified three-dimensional anomaly detection and repair method as described in claim 1, characterized in that, S4 includes: combining the category index as the fourth dimension with the three-dimensional coordinates of the registered point cloud to form a four-dimensional input; discretizing the four-dimensional input at multiple resolution levels to obtain four-dimensional grid coordinates; indexing the four-dimensional grid coordinates at each level using a hash function to obtain multi-resolution hash features; concatenating the hash features at each level and inputting them into a decoding network shared by multiple categories, outputting the signed distance value at the corresponding position, thereby representing the normal surface geometry of multiple categories within a single model.

6. The multi-category unified three-dimensional anomaly detection and repair method as described in claim 1, characterized in that, S5 includes: using the absolute signed distance value of the registration point in the signed distance field as the point-by-point anomaly score; performing local averaging on the point-by-point anomaly score in the neighborhood of each registration point, and using the maximum value after local averaging as the target-level anomaly score to suppress the influence of isolated noise points on the anomaly determination result; S6 includes: querying the zero isosurface of the signed distance field for the point cloud region determined to be anomaly, and outputting the repaired point cloud according to the zero isosurface.

7. The multi-category unified three-dimensional anomaly detection and repair method as described in claim 1, characterized in that, It also includes a training step for the multi-resolution hash-coded neural field: for normal training samples in a normal coordinate system, surface points are sampled from the sample surface, Gaussian perturbation is added near the surface points to sample non-surface points, and uniform points are sampled within the sample bounding box; Based on the true signed distance values ​​of the sampling points and the corresponding category labels, the multi-resolution hash-coded neural field is trained under supervision using relative error loss.

Images