A local patch based edge detection method for 3D point clouds

Abstract

The application relates to a local patch-based three-dimensional point cloud edge detection method, which fully utilizes local neighborhood feature information of a point cloud, constructs a local patch feature descriptor, and can effectively detect a three-dimensional point cloud edge; solves the problem that an existing method cannot accurately distinguish edge points and non-edge points, and overcomes the shortcomings of some methods in generalization performance and processing of dense edge feature points; in addition, the method has the characteristics of high recall rate.

Description

Technical Field

[0001] This application relates to the field of computer vision technology, and more specifically, to a method for edge detection of 3D point clouds based on local patch data. Background Technology

[0002] Currently, point clouds, as virtual representations of real-world objects, have attracted significant attention due to their ability to create highly detailed object models. Point cloud edge detection has become an important research area in computer vision, finding widespread application in tasks such as point cloud simplification, non-realistic rendering, and 3D reconstruction. The focus of point cloud edge detection is to compute features that distinguish edge points from non-edge points, identifying and classifying edge points in the point cloud to achieve a precise understanding of the object's geometric structure.

[0003] In recent years, numerous methods have been proposed, achieving excellent performance in 3D point cloud edge detection. These methods can be categorized into traditional methods and deep learning-based methods. Traditional methods primarily utilize geometric invariants, local neighborhood information of the point cloud, or surface fitting. Deep learning-based methods include point classification, parameter inference, and edge fitting.

[0004] Existing deep learning-based methods offer superior performance in 3D point cloud edge detection, but they still face several limitations. While point classification-based methods utilize point clouds with complete information as input, the low-density distribution of edge points and noise significantly hinder further development, preventing accurate differentiation between minute edge points and non-feature points. Parametric inference-based methods exhibit poor generalization, achieving only good results for specific features such as lines, circles, ellipses, and B-splines. Edge fitting methods, during upsampling, fit edges, leading to overly dense feature points. In summary, 3D point cloud edge feature extraction methods suffer from numerous limitations, preventing further performance improvements.

[0005] Based on the above analysis, the problems and shortcomings of existing technologies are as follows: First, due to the low-density distribution of edge points in point clouds and factors such as noise, existing technologies cannot accurately distinguish between edge points and non-edge points in point clouds. Second, some methods have poor generalization performance when dealing with complex or unknown shapes. Finally, some methods generate dense feature points, leading to information redundancy and a lack of representativeness in the extracted features. Summary of the Invention

[0006] To overcome at least one deficiency in the prior art, this application provides a three-dimensional point cloud edge detection method based on local patch.

[0007] Firstly, a method for constructing an edge detection network is provided, including:

[0008] Obtain the network training dataset, where the samples in the network training dataset are point clouds;

[0009] For each point cloud, calculate the mutual projection distance between each point in the point cloud and the points in its corresponding local patch to obtain two sets of local patch feature codes;

[0010] The two sets of local patch feature codes corresponding to each point cloud are input into the edge detection network for network training, resulting in the trained edge detection network.

[0011] The edge detection network includes a feature embedding module, a feature enhancement module, a feature fusion module, and a feature decoding and classification module. The feature embedding module performs dimensionality transformation, dimensionality increase, and dimensionality mapping on two sets of local patch feature codes to obtain two sets of local features. The feature enhancement module performs correlation calculation, dimensionality transformation, and regularization on the two sets of local features to obtain two sets of local patch shape descriptors. The feature fusion module fuses the two sets of local patch shape descriptors to obtain a fused local patch shape descriptor. The feature decoding and classification module outputs the probability that each point in the point cloud belongs to an edge point and a non-edge point based on the fused local patch shape descriptor.

[0012] In one embodiment, the mutual projected distance between each point in the point cloud and points in its corresponding local patch is calculated using the following formula:

[0013]

[0014]

[0015] in, p is the nth point in the point cloud n The k-th point p in its corresponding local patch k The projection distance between them, where N is the number of points in the point cloud and K is the number of points in the local patch. For point p k The normalized normal vector, p is the nth point in the point cloud n The k-th point p in the corresponding local patch k With point p n The projected distance between them For point p n The normalized normal vector.

[0016] In one embodiment, the feature embedding module includes a dimension transformation unit and a multilayer perceptron; the dimension transformation unit is used to perform dimension transformation and dimension increase on two sets of local patch feature codes to obtain two sets of dimension-transformed local patch feature codes; the multilayer perceptron is used to perform dimension mapping on the two sets of dimension-transformed local patch feature codes to obtain two sets of local features.

[0017] In one embodiment, the feature enhancement module includes an Encoder module, a dimension transformation unit, and a regularization unit. The Encoder module is used to perform correlation calculation on two sets of local features to obtain two sets of calculated features. The dimension transformation unit is used to perform dimension transformation on the two sets of calculated features to obtain two sets of dimension-transformed features. The regularization unit is used to regularize the two sets of dimension-transformed features to obtain two sets of local patch shape descriptors.

[0018] In one embodiment, the feature decoding and classification module includes a first multilayer perceptron, a second multilayer perceptron, and a third multilayer perceptron. The first and second multilayer perceptrons employ ReLU activation functions and a BatchNorm module, while the third multilayer perceptron employs a BatchNorm module. A dropout layer is provided between the first and second multilayer perceptrons.

[0019] In one embodiment, the loss function Loss used during network training is:

[0020]

[0021] Among them, y m,c Let y be a sign function. If the true class of the m-th point in the sample is equal to c, then y m,c =1, otherwise y m,c =0, p m,c Let M be the probability that the m-th point in the network output belongs to type c, where C is the total number of categories and M is the number of points.

[0022] Secondly, an edge detection network construction apparatus is provided, comprising:

[0023] The dataset acquisition module is used to acquire the network training dataset, in which the samples are point clouds;

[0024] The feature encoding module is used to calculate the mutual projection distance between each point in the point cloud and the points in its corresponding local patch for each point cloud, and to obtain two sets of local patch feature codes.

[0025] The network training module is used to input the two sets of local patch feature codes corresponding to each point cloud into the edge detection network for network training, and obtain the trained edge detection network.

[0026] The edge detection network includes a feature embedding module, a feature enhancement module, a feature fusion module, and a feature decoding and classification module. The feature embedding module performs dimensionality transformation, dimensionality increase, and dimensionality mapping on two sets of local patch feature codes to obtain two sets of local features. The feature enhancement module performs correlation calculation, dimensionality transformation, and regularization on the two sets of local features to obtain two sets of local patch shape descriptors. The feature fusion module fuses the two sets of local patch shape descriptors to obtain a fused local patch shape descriptor. The feature decoding and classification module outputs the probability that each point in the point cloud belongs to an edge point and a non-edge point based on the fused local patch shape descriptor.

[0027] Thirdly, a method for edge detection of 3D point clouds based on local patches is provided, including:

[0028] For the point cloud to be detected, calculate the mutual projection distance between each point in the point cloud and the points in its corresponding local patch to obtain two sets of local patch feature codes.

[0029] Two sets of local patch feature codes are input into the trained edge detection network to obtain the probability that each point in the point cloud to be detected belongs to an edge point and a non-edge point.

[0030] Each point is determined as either an edge point or a non-edge point based on the probability that each point belongs to an edge point or a non-edge point.

[0031] The trained edge detection network is obtained according to the edge detection network construction method described above.

[0032] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the aforementioned edge detection network construction method or the aforementioned three-dimensional point cloud edge detection method based on local patches.

[0033] Fifthly, a computer program product is provided, including a computer program / instructions, which, when executed by a processor, implement the aforementioned edge detection network construction method or the aforementioned three-dimensional point cloud edge detection method based on local patches.

[0034] Compared with the prior art, this application has the following advantages: The 3D point cloud edge detection method based on local patches in this application makes full use of the local neighborhood feature information of the point cloud to construct a local patch feature descriptor, which can effectively detect the edge of the 3D point cloud; it solves the problem that the existing methods cannot accurately distinguish between edge points and non-edge points, and overcomes the shortcomings of some methods in generalization performance and handling the dense edge feature points; in addition, the method of this application has the characteristic of high recall. Attached Figure Description

[0035] This application can be better understood by referring to the description given below in conjunction with the accompanying drawings, which, together with the detailed description below, are incorporated in and form part of this specification. In the drawings:

[0036] Figure 1 A flowchart illustrating an edge detection network construction method according to an embodiment of this application is shown;

[0037] Figure 2 A schematic diagram of the edge detection network is shown.

[0038] Figure 3 A structural block diagram of an edge detection network construction apparatus according to an embodiment of this application is shown;

[0039] Figure 4 A flowchart of a three-dimensional point cloud edge detection method based on local patches according to an embodiment of this application is shown;

[0040] Figure 5 The results of point cloud edge detection using the method of this application are shown, where (a) is the result of edge detection for triangular targets, (b) is the result of edge detection for elliptical targets, and (c) is the result of edge detection for hexagonal targets. Detailed Implementation

[0041] Exemplary embodiments of the present application will be described below with reference to the accompanying drawings. For clarity and brevity, not all features of the actual embodiments are described in the specification. However, it should be understood that many embodiment-specific decisions can be made in the development of any such actual embodiment to achieve the developer’s specific objectives, and these decisions may vary as the embodiments differ.

[0042] It should also be noted that, in order to avoid obscuring this application with unnecessary details, only the device structure closely related to the solution according to this application is shown in the accompanying drawings, while other details that are not closely related to this application are omitted.

[0043] It should be understood that this application is not limited to the described embodiments by virtue of the following description with reference to the accompanying drawings. In this document, embodiments may be combined with each other, features may be substituted or borrowed between different embodiments, and one or more features may be omitted in one embodiment, where feasible.

[0044] Due to the low-density distribution of edge points in point clouds and factors such as noise, most methods cannot accurately identify minute edge features. Analysis shows that spatially adjacent points often exhibit high correlation; considering the relationship between a point and its neighborhood in point cloud edge detection can help us effectively identify minute edge features. Improving the accuracy of minute edge recognition in 3D point cloud edge feature extraction can provide a more reliable data processing foundation for various application scenarios; it can improve the generalization performance when handling complex or unknown shapes, making the algorithm more versatile and adaptable; and by classifying only one point per local patch, it avoids generating dense feature points, helping to reduce information redundancy and making the extracted features more representative, providing more valuable information for subsequent tasks.

[0045] This application provides a method for constructing an edge detection network. Figure 1 A flowchart of an edge detection network construction method according to an embodiment of this application is shown. See also: Figure 1 The methods include:

[0046] Step S11: Obtain the network training dataset. The samples in the network training dataset are point clouds.

[0047] Here, the samples are 3D point cloud models based on the existing ABC dataset.

[0048] Step S12: For each point cloud, calculate the mutual projection distance between each point in the point cloud and the points in its corresponding local patch to obtain two sets of local patch feature codes.

[0049] Specifically, for each point p in the point cloud n The KNN algorithm is used to obtain its nearest point set. For local patches, the projection distance is calculated using the following formula:

[0050]

[0051]

[0052] in, p is the nth point in the point cloud n The k-th point p in its corresponding local patch k The projection distance between them, where N is the number of points in the point cloud and K is the number of points in the local patch, can be set to 20. For point p k The normalized normal vector, p is the nth point in the point cloud n The k-th point p in the corresponding local patch k With point p n The projected distance between them For point pn The normalized normal vector.

[0053] Finally, two sets of local patch feature codes were obtained as follows: and

[0054] Step S13: Input the two sets of local patch feature codes corresponding to each point cloud into the edge detection network for network training to obtain the trained edge detection network.

[0055] Figure 2 A schematic diagram of the edge detection network is shown. (See attached diagram) Figure 2 The edge detection network includes a feature embedding module, a feature enhancement module, a feature fusion module, and a feature decoder and point classification module. The feature embedding module performs dimensionality transformation, dimensionality increase, and dimensionality mapping on two sets of local patch feature codes to obtain two sets of local features. The feature enhancement module performs correlation calculation, dimensionality transformation, and regularization on the two sets of local features to obtain two sets of local patch shape descriptors. The feature fusion module fuses the two sets of local patch shape descriptors to obtain a fused local patch shape descriptor. The feature decoder and point classification module outputs the probability that each point in the point cloud belongs to an edge point and a non-edge point based on the fused local patch shape descriptor.

[0056] In this embodiment, local neighborhood feature information of point cloud is fully utilized to construct local patch feature descriptors, which can effectively detect the edges of 3D point cloud; it solves the problem that existing methods cannot accurately distinguish between edge points and non-edge points, and overcomes the shortcomings of some methods in generalization performance and handling of dense edge feature points; in addition, the method of this application has the characteristic of high recall rate.

[0057] In one embodiment, see Figure 2 The feature embedding module includes a dimension transformation unit (Transpose) and a multilayer perceptron (MLP). The dimension transformation unit is used to transform and increase the dimensions of the two sets of local patch feature codes to obtain two sets of dimension-transformed local patch feature codes. The multilayer perceptron is used to perform dimension mapping on the two sets of dimension-transformed local patch feature codes to obtain two sets of local features.

[0058] Here, the two sets of local patch features are encoded as two-dimensional features, each with a dimension of N×K. A dimension transformation unit is used to convert the dimension of the local patch feature encoding to K×N and add a third dimension. The initial value of the third dimension is 1. Then, after passing through a multilayer perceptron, the dimension of the third dimension is mapped to 128, and finally the dimension of the two sets of local features is K×N×128.

[0059] In one embodiment, the feature enhancement module includes an Encoder module, a Transpose unit, and a LayerNorm (LN) regularization unit. The Encoder module is used to calculate the correlation between two sets of local features to obtain two sets of calculated features. Here, the number of layers (num_layers) of the Encoder module can be set to 4, and the number of attention heads (nhead) can be set to 8. The Transpose unit is used to perform dimension transformation on the two sets of calculated features to obtain two sets of dimension-transformed features with a dimension size of N×K×128. The Regularization unit is used to regularize the two sets of dimension-transformed features to obtain two sets of local patch shape descriptors with a dimension size of N×K×128.

[0060] In one embodiment, the feature fusion module includes a reshape unit and a concatenate unit. Here, the reshape unit is used to reshape the last two dimensions of the local patch shape descriptor to obtain a local patch shape descriptor with two dimensions, the size of which is N×(K×128). The concatenate unit is used to concatenate the two sets of local patch shape descriptors with two dimensions to obtain a fused local patch shape descriptor with a size of N×(2×K×128).

[0061] In one embodiment, the feature decoding and classification module includes a first multilayer perceptron (MLP), a second multilayer perceptron (MLP), and a third multilayer perceptron (MLP). The first and second MLPs employ ReLU activation functions and a BatchNorm module, while the third MLP employs a BatchNorm module. A dropout layer is provided between the first and second MLPs. It should be noted that during network testing, the dropout layer is disabled and does not process data.

[0062] Here, the size of the first multilayer perceptron is 512, the size of the second multilayer perceptron is 128, the dropout layer value is set to 0.5, and the size of the third multilayer perceptron is 2. The output of the third multilayer perceptron is the network output, which is the probability that each point in the point cloud belongs to an edge point and a non-edge point, and is used to calculate the loss function.

[0063] In one embodiment, the loss function Loss used during network training is:

[0064]

[0065] Among them, y m,c Let y be a sign function. If the true class of the m-th point in the sample is equal to c, then y m,c =1, otherwise y m,c =0, p m,c Let C be the probability that the m-th point in the network output belongs to type c, where C is the total number of categories and M is the number of points. Here, C = 2, including both edge points and non-edge points.

[0066] Here, CrossEntropyLoss is used as the loss function, and the Adam optimizer is used to optimize the network. The initial learning rate is set to 1e-6, and milestones and gamma are used to adjust the learning rate. At the 75th and 150th training sessions, the learning rate is reduced to 0.1 times the original value. The network is trained for 200 epochs and edge detection is performed.

[0067] Employing the same inventive concept as the edge detection network construction method, this embodiment also provides a corresponding edge detection network construction apparatus. Figure 3 A structural block diagram of an edge detection network construction apparatus according to an embodiment of this application is shown. See also: Figure 3 ,include:

[0068] The dataset acquisition module 31 is used to acquire the network training dataset, in which the samples are point clouds.

[0069] The feature encoding module 32 is used to calculate the mutual projection distance between each point in the point cloud and the points in its corresponding local patch for each point cloud, and to obtain two sets of local patch feature codes.

[0070] Network training module 33 is used to input the two sets of local patch feature codes corresponding to each point cloud into the edge detection network for network training, so as to obtain the trained edge detection network.

[0071] The edge detection network includes a feature embedding module, a feature enhancement module, a feature fusion module, and a feature decoding and classification module. The feature embedding module performs dimensionality transformation, dimensionality increase, and dimensionality mapping on two sets of local patch feature codes to obtain two sets of local features. The feature enhancement module performs correlation calculation, dimensionality transformation, and regularization on the two sets of local features to obtain two sets of local patch shape descriptors. The feature fusion module fuses the two sets of local patch shape descriptors to obtain a fused local patch shape descriptor. The feature decoding and classification module outputs the probability that each point in the point cloud belongs to an edge point and a non-edge point based on the fused local patch shape descriptor.

[0072] The edge detection network construction device of this embodiment has the same inventive concept as the edge detection network construction method described above. Therefore, the specific implementation of the device can be found in the embodiment section of the edge detection network construction method described above, and its technical effects correspond to the technical effects of the above method. It will not be repeated here.

[0073] This application also provides a method for edge detection of 3D point clouds based on local patch data. Figure 4 A flowchart of a 3D point cloud edge detection method based on local patches according to an embodiment of this application is shown. See also... Figure 4 The methods include:

[0074] Step S41: For the point cloud to be detected, calculate the mutual projection distance between each point in the point cloud and the points in its corresponding local patch to obtain two sets of local patch feature codes; for specific implementation methods, please refer to the aforementioned embodiments.

[0075] Step S42: Input the two sets of local patch feature codes into the trained edge detection network to obtain the probability that each point in the point cloud to be detected belongs to an edge point and a non-edge point.

[0076] Step S43: Determine whether each point is an edge point or a non-edge point based on the probability that each point belongs to an edge point or a non-edge point. Here, the larger of the probabilities of a point belonging to an edge point and a non-edge point is determined. If the probability of belonging to an edge point is larger, then the point belongs to an edge point; otherwise, the point belongs to a non-edge point.

[0077] Here, the trained edge detection network is obtained according to the edge detection network construction method of the above embodiments.

[0078] To further verify the effectiveness of the 3D point cloud edge detection method based on local patches proposed in this application, the following experiments were conducted.

[0079] The experiments in this application were conducted on the ABC dataset, which contains CAD models and their corresponding feature annotation files. Feature annotations are categorized into five types: lines, circles, ellipses, B-splines, and other shapes. For each point cloud, the coordinates of feature points are extracted from the feature annotation file and saved as a ground truth point cloud model to obtain ground truth labels. Ground truth labels are generated by calculating the closest points on the feature point cloud to the original point cloud and recording their indices; these indices correspond to feature points (label 1); other points are non-feature points (label 0). This application uses 50 models for training, 200 models for validation, and 6000 models for testing. To address the imbalance between the number of feature points and non-feature points in the point cloud data, a specific training set sampling strategy is adopted: all feature points are selected in each chosen model, and twice the number of non-feature points are randomly selected. This strategy aims to balance the ratio of feature points to non-feature points in the training set, thereby enhancing the model's learning efficiency and adaptability. These 50 point cloud training models include all feature types from the ABC dataset. A comparison was made with six baseline methods: BE, PBRG, SGLBP, EC-Net, PIE-Net, and NerVE, and the results are shown in Table 1. Since the outputs of EC-Net and PIE-Net are fitted points rather than the original points on the point cloud, the outputs of the baseline methods and the ground truth were normalized. Furthermore, corresponding point pairs were calculated using the Iterative Closest Point (ICP) algorithm and treated as TPs. It can be seen that the method in this application significantly outperforms all baseline methods. Figure 5 The results of point cloud edge detection using the method of this application are shown, where (a) is the result of edge detection for triangular-like targets, (b) is the result of edge detection for elliptical-like targets, and (c) is the result of edge detection for hexagonal-like targets. These results verify that this application can detect minute edges and various types of edge points without redundant features.

[0080] Table 1

[0081] BE 0.486 0.165 0.182 0.271 0.469 PBRG 0.316 0.206 0.261 0.349 0.408 SGLBP 0.252 0.255 0.275 0.350 0.543 EC-Net 0.248 0.066 -0.084 0.081 0.421 PIE-Net 0.354 0.070 0.161 0.309 0.082 NerVE 0.252 0.388 0.541 0.894 0.401 EdgeFormer (this application) 0.115 0.839 0.885 0.89 0.922

[0082] This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the above-described edge detection network construction method or the above-described three-dimensional point cloud edge detection method based on local patches.

[0083] This application provides a computer program product, including a computer program / instruction. When the computer program / instruction is executed by a processor, it implements the edge detection network construction method described above, or the three-dimensional point cloud edge detection method based on local patches described above.

[0084] The above descriptions are merely various embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for constructing an edge detection network, characterized in that, include: Obtain the network training dataset, wherein the samples in the network training dataset are point clouds; For each point cloud, the mutual projection distance between each point in the point cloud and the points in its corresponding local patch is calculated to obtain two sets of local patch feature codes. The two sets of local patch feature codes corresponding to each point cloud are input into the edge detection network for network training to obtain the trained edge detection network. The edge detection network includes a feature embedding module, a feature enhancement module, a feature fusion module, and a feature decoding and classification module. The feature embedding module performs dimensionality transformation, dimensionality increase, and dimensionality mapping on the two sets of local patch feature codes to obtain two sets of local features. The feature enhancement module performs correlation calculation, dimensionality transformation, and regularization on the two sets of local features to obtain two sets of local patch shape descriptors. The feature fusion module fuses the two sets of local patch shape descriptors to obtain a fused local patch shape descriptor. The feature decoding and classification module outputs the probability that each point in the point cloud belongs to an edge point and a non-edge point based on the fused local patch shape descriptor. For each point in the point cloud The KNN algorithm is used to obtain its nearest point set. Local patches, or local faces, are used to calculate the mutual projection distance between each point in the point cloud and its corresponding point in the local patch, using the following formula: in, For the first point cloud n Points The first in its corresponding local patch k Points The projected distance between them N The number of points in the point cloud. K The number of points in the local patch. For point The normalized normal vector, For the first point cloud n Points The first in the corresponding local patch k Points With point The projected distance between them For point The normalized normal vector; The feature embedding module includes a dimension transformation unit and a multilayer perceptron. The dimension transformation unit is used to perform dimension transformation and dimension increase on two sets of local patch feature codes to obtain two sets of dimension-transformed local patch feature codes. The multilayer perceptron is used to perform dimension mapping on the two sets of dimension-transformed local patch feature codes to obtain two sets of local features. The two sets of local patch features are encoded as two-dimensional features, both having a dimension of [missing value]. The dimension transformation unit is used to convert the dimension size of the local patch feature encoding to... A third dimension is added, initially set to 1. Then, after passing through a multilayer perceptron, the size of the third dimension is mapped to 128, resulting in both sets of local features having a dimension of 128. ; The feature enhancement module includes an Encoder module, a dimension transformation unit, and a regularization unit. The Encoder module is used to calculate the correlation between two sets of local features to obtain two sets of calculated features. The correlation calculation is performed by the Encoder module to perform attention calculation on the two sets of local features. The dimension transformation unit is used to perform dimension transformation on the two sets of calculated features to obtain two sets of dimension-transformed features. The regularization unit is used to regularize the two sets of dimension-transformed features to obtain two sets of local patch shape descriptors.

2. The method as described in claim 1, characterized in that, The feature decoding and classification module includes a first multilayer perceptron, a second multilayer perceptron, and a third multilayer perceptron. The first and second multilayer perceptrons use the ReLU activation function and the BatchNorm module, while the third multilayer perceptron uses the BatchNorm module. A dropout layer is provided between the first and second multilayer perceptrons.

3. The method as described in claim 1, characterized in that, The loss function used during network training for: in, Let be the sign function, if the th in the sample m The true category of each point is equal to c ,but ,otherwise , The first output of the network m Each point belongs to type c The probability, The total number of categories, M The number of points.

4. An edge detection network construction device, characterized in that, include: The dataset acquisition module is used to acquire the network training dataset, wherein the samples in the network training dataset are point clouds; The feature encoding module is used to calculate the mutual projection distance between each point in the point cloud and the points in its corresponding local patch for each point cloud, and to obtain two sets of local patch feature codes. The network training module is used to input the two sets of local patch feature codes corresponding to each point cloud into the edge detection network for network training, so as to obtain the trained edge detection network. The edge detection network includes a feature embedding module, a feature enhancement module, a feature fusion module, and a feature decoding and classification module. The feature embedding module performs dimensionality transformation, dimensionality increase, and dimensionality mapping on the two sets of local patch feature codes to obtain two sets of local features. The feature enhancement module performs correlation calculation, dimensionality transformation, and regularization on the two sets of local features to obtain two sets of local patch shape descriptors. The feature fusion module fuses the two sets of local patch shape descriptors to obtain a fused local patch shape descriptor. The feature decoding and classification module outputs the probability that each point in the point cloud belongs to an edge point and a non-edge point based on the fused local patch shape descriptor. For each point in the point cloud The KNN algorithm is used to obtain its nearest point set. Local patches, or local faces, are used to calculate the mutual projection distance between each point in the point cloud and its corresponding point in the local patch, using the following formula: in, For the first point cloud n Points The first in its corresponding local patch k Points The projected distance between them N The number of points in the point cloud. K The number of points in the local patch. For point The normalized normal vector, For the first point cloud n Points The first in the corresponding local patch k Points With point The projected distance between them For point The normalized normal vector; The feature embedding module includes a dimension transformation unit and a multilayer perceptron. The dimension transformation unit is used to perform dimension transformation and dimension increase on two sets of local patch feature codes to obtain two sets of dimension-transformed local patch feature codes. The multilayer perceptron is used to perform dimension mapping on the two sets of dimension-transformed local patch feature codes to obtain two sets of local features. The two sets of local patch features are encoded as two-dimensional features, both having a dimension of [missing value]. The dimension transformation unit is used to convert the dimension size of the local patch feature encoding to... A third dimension is added, initially set to 1. Then, after passing through a multilayer perceptron, the size of the third dimension is mapped to 128, resulting in both sets of local features having a dimension of 128. ; The feature enhancement module includes an Encoder module, a dimension transformation unit, and a regularization unit. The Encoder module is used to calculate the correlation between two sets of local features to obtain two sets of calculated features. The correlation calculation is performed by the Encoder module to perform attention calculation on the two sets of local features. The dimension transformation unit is used to perform dimension transformation on the two sets of calculated features to obtain two sets of dimension-transformed features. The regularization unit is used to regularize the two sets of dimension-transformed features to obtain two sets of local patch shape descriptors.

5. A method for edge detection of 3D point clouds based on local patches, characterized in that, include: For the point cloud to be detected, the mutual projection distance between each point in the point cloud and the points in its corresponding local patch is calculated to obtain two sets of local patch feature codes. The two sets of local patch feature codes are input into the trained edge detection network to obtain the probability that each point in the point cloud to be detected belongs to an edge point and a non-edge point. Each point is determined to be an edge point or a non-edge point based on the probability that each point belongs to an edge point or a non-edge point. The trained edge detection network is obtained by the edge detection network construction method according to any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the edge detection network construction method according to any one of claims 1-3, or the three-dimensional point cloud edge detection method based on local patches according to claim 5.

7. A computer program product, characterized in that, Includes a computer program / instruction, which, when executed by a processor, implements the edge detection network construction method according to any one of claims 1-3, or the three-dimensional point cloud edge detection method based on local patches according to claim 5.

Images