Point cloud classification / segmentation method based on local geometric contour and global structure preservation

Abstract

The invention discloses a point cloud classification / segmentation method based on local geometric contour and global structure maintenance, which comprises the following steps: firstly, extracting local and global neighborhood relation graphs of a point cloud from the angles of spatial distance and feature space, and accurately capturing complex geometric and semantic information in point cloud data by spatial distance calculation and a k-nearest neighbor algorithm; meanwhile, a neighborhood relation graph self-diffusion mechanism further strengthens the feature diffusion of the surface contour of the point cloud, so that local geometric contour information is fully utilized, the expression ability of a local structure is improved, and then, the accuracy and robustness of feature extraction are improved; local features and global features are respectively extracted from the local neighborhood relation graph and the global neighborhood relation graph and are sent to a feature fusion module, the local features and the global features are effectively fused by using a cross attention mechanism, and key information is adaptively enhanced, so that the classification and segmentation precision is improved, and the classification and segmentation efficiency is improved. According to the fusion method, the classification / segmentation output head can capture features of different scales more accurately, the processing capacity of complex point cloud data is improved, an accurate classification / segmentation result is generated according to the input fusion features, and the classification / segmentation precision is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of dialogue recommendation, and more specifically, relates to a point cloud classification / segmentation method based on local geometric contours and global structure preservation. Background Art

[0002] With the continuous development of 3D scanning technology, point cloud data is increasingly used in computer vision, robotics, autonomous driving, urban modeling and other fields. Point cloud data usually consists of a large number of 3D coordinate points, containing rich geometric shapes and structural information, so point cloud classification and segmentation tasks play an important role in these applications. However, existing point cloud classification and segmentation methods generally have some technical bottlenecks, especially when processing point cloud surface contour information.

[0003] Traditional point cloud data processing methods, especially deep learning-based graph convolutional networks (GCN) and convolutional neural networks (CNN) methods, often focus on the extraction of global features or local features of point clouds. For example, in the invention patent application published on October 20, 2023 with publication number CN116912561A, a point cloud data classification and segmentation method based on a multi-view adaptive graph convolutional network is disclosed. By constructing a graph network in the offline stage, the collected point cloud data is input into the adaptive rotation matrix generator to obtain the adaptively rotated point cloud data and generate three multi-view projection images; the point cloud data and the three multi-view projection images are respectively constructed to obtain global information graph data and local information graph data, and then respectively input into the global feature extraction network and the local feature extraction network, and after the global features and local features are extracted, the two are fused and input into the functional neural network output head, and the loss function is calculated according to the obtained results to realize the training of the graph network; in the online stage, the point cloud data collected in real time is input into the trained graph network to obtain the point cloud classification or segmentation results. The present invention obtains an optimized viewing angle by adaptively rotating the point cloud data, flexibly adjusts the position of the viewing angle according to the geometric features and spatial distribution of the point cloud data, and introduces the connection between points of different depths on a specific projection surface, so as to better capture the local and global features of the point cloud data and make the composition more accurate and robust. The global feature extraction network of the invention patent application is a six-layer tandem neural network with a maximum pooling layer, and the local feature extraction network is a two-layer tandem neural network. It simply uses a convolutional neural network to extract global features and local features and splice them. Some other methods extract local features by constructing a neighborhood relationship graph and using spatial distance calculation. However, these methods ignore that the local neighborhood construction method based on spatial distance usually relies on the geometric distance between points to select neighbors. This method is easily affected by the density change of the point cloud distribution in the point cloud. In dense areas, neighbor points may be too concentrated, resulting in information redundancy and deviation in feature expression; while in sparse areas, neighbor points may be insufficient, resulting in incomplete local feature information. Due to the sparsity and complexity of point cloud data, especially in surface contour areas such as the edges of objects, existing methods often do not process these areas sufficiently, resulting in limited improvement in point cloud classification / segmentation accuracy.

[0004] In addition, most existing point cloud data processing simply splices local features with global features. Local features and global features exist independently, which may lead to inconsistent or redundant information. Especially when facing scenes with complex shapes and rich semantics, existing technologies often cannot fully integrate local geometric contour information with global structural semantic information, resulting in insufficient robustness and generalization ability of point cloud classification / segmentation models when processing diverse and unstructured point cloud data.

[0005] Therefore, how to overcome the limitations of existing technologies, make full use of the local geometric contour information and global structural semantic information in point cloud data, and improve the accuracy and robustness of point cloud classification and segmentation has become a key issue that needs to be urgently solved in the current field of point cloud analysis. Summary of the invention

[0006] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a point cloud classification / segmentation method based on local geometric contours and global structure preservation, so as to fully consider the local geometric contour information and the global structural semantic information, and utilize the cross-attention mechanism to perform multi-scale feature fusion of local features and global features, so as to improve the accuracy and robustness of point cloud classification / segmentation.

[0007] To achieve the above-mentioned object of the invention, the present invention is based on a point cloud classification / segmentation method that preserves local geometric contours and global structures, characterized in that it includes the following steps:

[0008] (1) Use multi-stage cascade feature extractors to extract fusion features from point cloud data

[0009] The first-level feature extractor takes the three-dimensional coordinates of the midpoints in the point cloud data as input features, and the subsequent feature extractors take the fusion features output by the previous-level feature extractor as input features. Each level of feature extractors includes a local neighborhood construction module, a global neighborhood construction module, a local feature extraction network, a global feature extraction network, and a feature fusion module. In each level of feature extractors, the input features are input into the local neighborhood construction module and the global neighborhood construction module respectively.

[0010] 1.1) Obtain local neighborhood relationship graph and global neighborhood relationship graph

[0011] In the local neighborhood construction module, the spatial distance between each point and other points is first calculated based on the input features. Then, based on the spatial distance, the neighborhood point set of each point is determined from the point cloud data, that is, the neighborhood relationship graph is constructed by selecting the first k minimum spatial distances. The obtained neighborhood relationship graph is then propagated and enhanced using the self-diffusion mechanism to quickly expand to the contour points on the point cloud surface. Finally, the first k maximum values ​​and indexes on the specified dimension in the diffused neighborhood relationship graph are determined as the local neighborhood relationship graph.

[0012] In the global neighborhood construction module, the spatial distance between each point and other points is first calculated based on the input features, and then each point finds the indexes of the nearest k neighbors as the global neighborhood relationship graph;

[0013] 1.2) Extract local features and global features

[0014] Both the local feature extraction network and the global feature extraction network are multi-layer cascade neural networks. Each layer of the neural network consists of a two-dimensional convolution layer and a maximum pooling layer. The feature update function of each layer is:

[0015]

[0016] Among them, h i (n) is the input feature of the ith point of the nth layer neural network, h i (n+1) is the output feature of the i-th point of the n-th layer neural network, Θ is the parameter of the two-dimensional convolutional layer, max() is the maximum pooling function, N(i) represents the neighborhood point set of the i-th point, and j represents the j-th point in the neighborhood point set N(i);

[0017] The local neighborhood relationship graph is used as the input feature to the first layer of the local feature extraction network. The output feature of the last layer of the local feature extraction network is the extracted local feature F. local The global neighborhood relationship graph is used as the input feature to the first layer of the global feature extraction network. The output feature of the last layer of the global feature extraction network is the extracted global feature F global ;

[0018] 1.3) Use the cross attention mechanism to fuse local features and global features to obtain fused features

[0019] In the feature fusion module, the cross attention mechanism is used to fuse local features and global features: First, a set of learnable global query values ​​Q is introduced, and for a given local feature F local Use the cross attention mechanism to aggregate it with the global query value Q to obtain the intermediate local feature CA(Q,F local ), then, the intermediate local feature CA(Q,F local ) and the global feature F global Further fusion is performed to obtain the fusion feature F fusion :

[0020] F fusion =CA(F global ,CA(Q,F local ))

[0021] Among them, CA(·,·) represents the cross attention operation;

[0022] (2) Point cloud classification / segmentation

[0023] Using the classification / segmentation output head, the fusion feature F obtained from each point in the point cloud data fusion Perform classification / segmentation to obtain point cloud classification / segmentation results.

[0024] The object of the present invention is achieved in this way.

[0025] The point cloud classification / segmentation method based on local geometric contour and global structure preservation of the present invention first extracts the local and global neighborhood relationship graphs of the point cloud from the perspectives of spatial distance and feature space respectively through the local neighborhood construction module and the global neighborhood construction module, wherein the spatial distance calculation and the k-nearest neighbor algorithm accurately capture the complex geometric and semantic information in the point cloud data, and enhance the accuracy and robustness of feature extraction. At the same time, the neighborhood relationship graph self-diffusion mechanism further enhances the feature diffusion of the point cloud surface contour, so as to make full use of the local geometric contour information and improve the expression ability of the local structure. Then, local features and global features are extracted from the local neighborhood relationship graph and the global neighborhood relationship graph respectively, and sent to the feature fusion module, and the cross-attention mechanism is used to effectively fuse the local features and the global features, and the key information is adaptively enhanced, thereby improving the accuracy of classification and segmentation. This fusion method enables the classification / segmentation output head to capture features of different scales more accurately, improves the processing capability of complex point cloud data, and generates accurate classification / segmentation results according to the input fusion features, significantly improves the classification / segmentation accuracy, has good online adaptability and real-time processing capabilities, and meets the efficiency and reliability requirements in practical applications.

[0026] The present invention shows significant advantages in task adaptability, expansion capability and integration of multi-view composition through the organic combination of modular design and innovative technology. The modular design enables each functional module to be independently optimized and flexibly combined, thereby improving the system's adaptability to different tasks. Whether it is the classification or segmentation task of point cloud data, it can be efficiently configured and adjusted according to specific needs. In addition, the feature fusion mechanism enhances the system's ability to process complex point cloud data, especially in the application scenario of multi-view composition. By effectively fusing information from different perspectives, the model's robustness and generalization ability for diversified data are improved. Therefore, the present invention not only performs well in traditional point cloud analysis tasks, but also has good scalability and can adapt to the needs of more complex tasks in the future. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a flowchart of a specific implementation of the point cloud classification / segmentation method based on local geometric contour and global structure preservation of the present invention;

[0028] Figure 2 It is the overall network structure diagram of the point cloud classification / segmentation method based on local geometric contour and global structure preservation of the present invention;

[0029] Figure 3 It is a self-diffusion effect diagram of the neighborhood relationship graph of any point in the point cloud data, where (a) is before self-diffusion and (b) is after self-diffusion;

[0030] Figure 4 This is the point cloud segmentation effect diagram of 'airplane', 'chair', 'table' and 'lamp' in the ShapNetPart dataset. DETAILED DESCRIPTION

[0031] The specific implementation of the present invention is described below in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when the detailed description of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.

[0032] Example 1

[0033] Figure 1 It is a flow chart of a specific implementation of the point cloud classification / segmentation method based on local geometric contours and global structure preservation of the present invention.

[0034] In this embodiment, if Figure 1 As shown, the point cloud classification / segmentation method based on local geometric contour and global structure preservation of the present invention includes the following steps:

[0035] Step S1: Use a multi-stage cascade feature extractor to extract fusion features from point cloud data

[0036] In this embodiment, if Figure 2 As shown, four-stage series feature extractors 1 to 4 are used to extract fusion features from point cloud data. The first-stage feature extractor uses the three-dimensional coordinates of the midpoint of the point cloud data as input features, and the subsequent feature extractors use the fusion features output by the previous-stage feature extractor as input features. Each stage of feature extractors includes a local neighborhood construction module, a global neighborhood construction module, a local feature extraction network, a global feature extraction network, and a feature fusion module. In each stage of feature extractors, the input features are input into the local neighborhood construction module and the global neighborhood construction module respectively.

[0037] In this embodiment, point cloud data with three-dimensional coordinates P=[x, y, z] is obtained through laser radar (LiDAR) scanning or a three-dimensional model generation tool, and the data size is N×3, where N is the number of point cloud points.

[0038] Step S1.1: Obtain local neighborhood relationship graph and global neighborhood relationship graph

[0039] In the local neighborhood construction module, the spatial distance between each point and other points is first calculated based on the input features. Then, based on the spatial distance, the neighborhood point set of each point is determined from the point cloud data, that is, a neighborhood relationship graph is constructed by selecting the first k minimum spatial distances. The obtained neighborhood relationship graph is then used to propagate and enhance features using the self-diffusion mechanism, thereby quickly expanding to the point cloud surface contour points. Finally, the first k maximum values ​​and indexes on the specified dimension in the diffused neighborhood relationship graph are determined as the local neighborhood relationship graph.

[0040] In this embodiment, the local neighborhood construction module is composed of a spatial distance calculation algorithm, a k-nearest neighbor algorithm, a neighborhood relationship graph symmetry method, a neighborhood relationship graph normalization algorithm, a neighborhood relationship graph self-diffusion algorithm, and a tensor maximum k element selection operation. Among them, the spatial distance calculation algorithm is used to calculate the spatial distance between each point and other points based on the input features, the k-nearest neighbor algorithm is used to determine the neighborhood point set of each point from the point cloud data according to the spatial distance, the algorithm constructs a neighborhood relationship graph by selecting the first k minimum values, the neighborhood relationship graph symmetry method is used to convert an asymmetric neighborhood relationship graph into a symmetric matrix to meet the positive definiteness of the matrix diffusion kernel, and the neighborhood relationship graph normalization algorithm is used to normalize the neighborhood relationship graph of the graph structure, thereby alleviating the numerical instability problem that may arise during the point diffusion process of the neighborhood relationship graph.

[0041] The normalization algorithm of the neighborhood relationship graph is:

[0042]

[0043] Among them, A represents the neighborhood relationship graph, and D represents the degree matrix of the neighborhood relationship graph.

[0044] The neighborhood relationship graph self-diffusion algorithm uses the self-diffusion mechanism to propagate and enhance the features of the acquired point cloud neighborhood relationship graph, thereby quickly expanding it to the surface contour points of the point cloud. The tensor maximum k element selection operation is used to determine the top k maximum values ​​and indexes on the specified dimension in the diffused neighborhood relationship graph.

[0045] The neighborhood relationship graph self-diffusion algorithm, i.e., the self-diffusion mechanism calculation formula is:

[0046]

[0047] in, n represents the number of diffusion times. In this embodiment, n=5, the neighborhood relationship graph self-diffusion effect diagram is as follows Figure 3As shown in the figure, the neighborhood before diffusion is the nearest neighborhood based on the center point. Due to ignoring the local geometric structure, some points belonging to different components are included in the neighborhood of the center point, resulting in noise in feature extraction. After using the self-diffusion method, the neighborhood of the center point changes adaptively according to the local geometric structure of the point cloud, avoiding noise while expanding the neighborhood range.

[0048] In the global neighborhood construction module, the spatial distance between each point and other points is first calculated based on the input features, and then each point finds the k nearest neighbor indexes as the global neighborhood relationship graph.

[0049] In this embodiment, the global neighborhood construction module is composed of a feature space distance calculation algorithm and a k-nearest neighbor algorithm, wherein the feature space distance calculation algorithm is used to calculate the distance between each point and other points in the feature space, and the k-nearest neighbor algorithm finds the k nearest neighbor indexes for each point.

[0050] Step S1.2: Extract local features and global features

[0051] Both the local feature extraction network and the global feature extraction network are multi-layer cascade neural networks. Each layer of the neural network consists of a two-dimensional convolution layer and a maximum pooling layer. The feature update function of each layer is:

[0052]

[0053] Among them, h i (n) is the input feature of the ith point of the nth layer neural network, h i (n+1) is the output feature of the i-th point of the n-th layer neural network, Θ is the parameter of the two-dimensional convolutional layer, max() is the maximum pooling function, N(i) represents the neighborhood point set of the i-th point, and j represents the j-th point in the neighborhood point set N(i).

[0054] The local neighborhood relationship graph is used as the input feature to the first layer of the local feature extraction network. The output feature of the last layer of the local feature extraction network is the extracted local feature F. local The global neighborhood relationship graph is used as the input feature to the first layer of the global feature extraction network. The output feature of the last layer of the global feature extraction network is the extracted global feature F global .

[0055] In this embodiment, the local feature extraction network and the global feature extraction network are both four-layer serial neural networks, wherein the two-dimensional convolution layer of the first layer is a graph convolution layer with an input dimension of 3 and an output dimension of 64, the two-dimensional convolution layer of the second layer is a graph convolution layer with an input dimension of 64 and an output dimension of 64, the two-dimensional convolution layer of the third layer is a graph convolution layer with an input dimension of 64 and an output dimension of 128, and the two-dimensional convolution layer of the fourth layer is a graph convolution layer with an input dimension of 128 and an output dimension of 256.

[0056] Step S1.3: Use the cross attention mechanism to fuse local features and global features to obtain fused features

[0057] In this embodiment, if Figure 2 As shown in the figure, in the feature fusion module, the cross attention mechanism is used to fuse local features and global features: First, a set of learnable global query values ​​Q are introduced, and for a given local feature F local Use the cross attention mechanism to aggregate it with the global query value Q to obtain the intermediate local feature CA(Q,F local ), then, the intermediate local feature CA(Q,F local ) and the global feature F global Further fusion is performed to obtain the fusion feature F fusion :

[0058] F fusion =CA(F global ,CA(Q,F local ))

[0059] Here, CA(·,·) represents the cross attention operation.

[0060] Step S2: Point cloud classification / segmentation

[0061] Using the classification / segmentation output head, the fusion feature F obtained from each point in the point cloud data fusion Perform classification / segmentation to obtain point cloud classification / segmentation results.

[0062] In this embodiment, the classification / segmentation output head is a functional neural network output head including several layers of multi-layer perceptrons, and the output dimension is the number of categories in the classification task or the number of components to be segmented in the segmentation task.

[0063] According to the fusion feature F obtained from each point in the point cloud data fusion Input the functional neural network output head, and update the network weights after calculating the loss function based on the obtained results. The functional neural network output head is a multilayer perceptron with a series structure of two linear layers and one maximum pooling layer, and its output result is a vector of the number of categories in the classification task, where the index corresponding to the maximum value is the final result of the network prediction.

[0064] The step of updating the network weights is: calculating the cross entropy error between the obtained vector and the data set label and passing it backward to update the network weights. The updating network weights uses the public data set ModelN et40, in which the training set has 9843 point clouds, the test set has 2468 point clouds, and the validation set is not included. 1024 points are randomly sampled from each point cloud. The optimizer used in the training process is the Adam optimizer with a learning rate of 0.001. The training generation number is 300 generations, and the final trained graph network is obtained, which includes a local neighborhood construction module, a global neighborhood construction module, a feature extraction network, a local feature extraction network, a global feature extraction network, and a classification / segmentation output head, i.e., a functional neural network output head.

[0065] After specific practical experiments, using PyTorch as the underlying deep learning framework and adopting the hardware environment of NVIDIA RT X4090GPU (24GB), the above method was run. On the official test set, the average instance accuracy reached 94.2%, which is 1.3% higher than the existing baseline technology, such as Wang et al. in "Dynamic gra ph cnn for learning onpoint clouds" ([J]. Acm Transactions On Graphics (tog), 2019, 38(5): 1-12).

[0066] Example 2

[0067] In this embodiment, the ShapeNetPart dataset is selected for training and testing, which includes 12137 training set samples, 1870 validation set samples, and 2874 test set samples. The number of points in each point cloud is about 2000, and the specific number of point clouds may vary. During the training process, the Adam optimizer with a learning rate of 0.001 is used, and the training generation number is 200 generations.

[0068] In the point cloud segmentation task, this embodiment first extracts local geometric features through the local neighborhood construction module, and then extracts global semantic information through the global neighborhood construction module. The local features and global features after feature extraction are fused through the cross attention mechanism to obtain a more accurate point cloud representation.

[0069] In the experiment, PyTorch was used as the underlying deep learning framework, and the hardware environment of NVIDIA RTX 4090GPU (24GB) was used for training and testing. After verification on the training set and the test set, the average instance accuracy on the test set reached 86.5%, an increase of 1.4% compared to the existing technology; the average category accuracy was 83.5%, an increase of 1.2% compared to the existing technology.

[0070] Figure 4 This is the point cloud segmentation effect diagram of 'airplane', 'chair', 'table' and 'table lamp' in the ShapNetPart dataset, where different colors represent different parts. Figure 4 It can be seen that in categories such as 'airplane', 'chair' and 'table lamp', each component is accurately segmented, and the point cloud segmentation performance is better than the existing methods.

[0071] Although the above describes the illustrative specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A point cloud classification / segmentation method based on local geometric contours and global structure preservation, characterized in that: The following steps are involved: (1) Use multi-stage cascade feature extractors to extract fusion features from point cloud data The first-level feature extractor takes the three-dimensional coordinates of the midpoints in the point cloud data as input features, and the subsequent feature extractors take the fusion features output by the previous-level feature extractor as input features. Each level of feature extractors includes a local neighborhood construction module, a global neighborhood construction module, a local feature extraction network, a global feature extraction network, and a feature fusion module. In each level of feature extractors, the input features are input into the local neighborhood construction module and the global neighborhood construction module respectively. 1.1) Obtain local neighborhood relationship graph and global neighborhood relationship graph In the local neighborhood construction module, the spatial distance between each point and other points is first calculated based on the input features. Then, based on the spatial distance, the neighborhood point set of each point is determined from the point cloud data, that is, the neighborhood relationship graph is constructed by selecting the first k minimum spatial distances. The obtained neighborhood relationship graph is then propagated and enhanced using the self-diffusion mechanism to quickly expand to the contour points on the point cloud surface. Finally, the first k maximum values ​​and indexes on the specified dimension in the diffused neighborhood relationship graph are determined as the local neighborhood relationship graph. In the global neighborhood construction module, the spatial distance between each point and other points is first calculated based on the input features, and then each point finds the indexes of the nearest k neighbors as the global neighborhood relationship graph; 1.2) Extract local features and global features Both the local feature extraction network and the global feature extraction network are multi-layer cascade neural networks. Each layer of the neural network consists of a two-dimensional convolution layer and a maximum pooling layer. The feature update function of each layer is: Among them, h i (n) is the input feature of the ith point of the nth layer neural network, h i (n+1) is the output feature of the i-th point of the n-th layer neural network, Θ is the parameter of the two-dimensional convolutional layer, max() is the maximum pooling function, N(i) represents the neighborhood point set of the i-th point, and j represents the j-th point in the neighborhood point set N(i); The local neighborhood relationship graph is used as the input feature to the first layer of the local feature extraction network. The output feature of the last layer of the local feature extraction network is the extracted local feature F. local The global neighborhood relationship graph is used as the input feature to the first layer of the global feature extraction network. The output feature of the last layer of the global feature extraction network is the extracted global feature F global ; 1.3) Use the cross attention mechanism to fuse local features and global features to obtain fused features In the feature fusion module, the cross attention mechanism is used to fuse local features and global features: First, a set of learnable global query values ​​Q is introduced, and for a given local feature F local Use the cross attention mechanism to aggregate it with the global query value Q to obtain the intermediate local feature CA(Q,F local ), then, the intermediate local feature CA(Q,F local ) and the global feature F global Further fusion is performed to obtain the fusion feature F fusion : F fusion =CA(F global ,CA(Q,F local )) Among them, CA(·,·) represents the cross attention operation; (2) Point cloud classification / segmentation Using the classification / segmentation output head, the fusion feature F obtained from each point in the point cloud data fusion Perform classification / segmentation to obtain point cloud classification / segmentation results.

2. The point cloud classification / segmentation method based on local geometric contours and global structure preservation according to claim 1, characterized in that: Before the neighborhood relationship graph described in step 1.1) uses the self-diffusion mechanism to propagate and strengthen the features, the asymmetric neighborhood relationship graph needs to be converted into a symmetric matrix and normalized. The normalization of the neighborhood relationship graph is: Among them, A represents the neighborhood relationship graph, and D represents the degree matrix of the neighborhood relationship graph.

3. The point cloud classification / segmentation method based on local geometric contours and global structure preservation according to claim 1, characterized in that: The calculation formula for the self-diffusion mechanism described in step 1.1) is: in, n represents the number of diffusions.

4. The point cloud classification / segmentation method based on local geometric contours and global structure preservation according to claim 1, characterized in that: The local feature extraction network and the global feature extraction network described in step 1.2) are both four-layer cascaded neural networks, wherein the two-dimensional convolutional layer of the first layer is a graph convolutional layer with an input dimension of 3 and an output dimension of 64, the two-dimensional convolutional layer of the second layer is a graph convolutional layer with an input dimension of 64 and an output dimension of 64, the two-dimensional convolutional layer of the third layer is a graph convolutional layer with an input dimension of 64 and an output dimension of 128, and the two-dimensional convolutional layer of the fourth layer is a graph convolutional layer with an input dimension of 128 and an output dimension of 256.

5. The point cloud classification / segmentation method based on local geometric contours and global structure preservation according to claim 1, characterized in that: The classification / segmentation output head described in step (2) is a functional neural network output head including several layers of multi-layer perceptrons, and the output dimension is the number of categories in the classification task or the number of components to be segmented in the segmentation task.

Images