A Point Cloud Completion Method Based on Dynamic Graph Convolution and Attention Mechanism

Abstract

The present invention introduces a point cloud completion method based on dynamic graph convolution and attention mechanism, which includes: Step S1, using dynamic graph convolution technology for feature extraction; Step S2, combining attention pooling method and max pooling method for feature aggregation; Step S3, local missing space target feature completion and reconstruction. This application defines the point cloud neighborhood and updates the dynamic neighborhood graph, and then combines the dynamic neighborhood graph to update and densely connect the point cloud; and uses a multi-channel method combining attention pooling and max pooling for feature aggregation of the point cloud, efficiently completing the completion of the missing point cloud, and maximizing the retention and restoration of the details and geometric structure of the input point cloud; moreover, this application proposes a one-stage mode network model, which fuses the point-by-point features and global geometric features of the point cloud, completes the point cloud in the feature space, expands and refines the features, so as to reconstruct the complete point cloud, and then directly outputs the complete point cloud.

Description

Technical Field

[0001] The present invention belongs to the field of model construction design, and particularly relates to a point cloud completion method based on dynamic graph convolution and attention mechanism. Background Art

[0002] In real scenarios, the point clouds obtained by 3D scanning devices such as lidar have problems such as sparsity and missing due to the limitations of sensor resolution, viewing angle, and the influence of occlusion between object structures or between objects, which brings certain difficulties to the deeper application of subsequent point clouds. Therefore, point cloud shape completion is to solve the problem of how to restore a complete point cloud from local observed point clouds. For any given point cloud with structural missing, a corresponding complete point cloud can be obtained. This task is the basis for many downstream tasks, such as shape classification, segmentation, etc., and is also an indispensable operation link in subsequent applications.

[0003] Chinese Patent CN112614071A discloses a diverse point cloud completion method and device based on self-attention, which relates to the fields of computer three-dimensional point cloud completion and deep learning technology. Among them, the method includes: obtaining point cloud data, processing the point cloud data to obtain an input point proxy sequence; encoding the point proxy sequence to obtain a point encoding vector, decoding the point encoding vector to obtain a predicted point proxy; inputting the predicted point proxy into a multi-layer perceptron to obtain a predicted point center, and restoring the complete point cloud data based on the predicted point center. Thus, the point cloud is processed into a point proxy sequence, and an encoder and a decoder are used to construct the long-range relationship between different points of the point cloud to achieve point cloud reconstruction. However, it uses a point-by-point shared multi-layer perceptron to extract the features of the point cloud, and does not fully utilize the local structure information between points. In the completion task, the missing and sparse nature of the input point cloud makes it difficult to obtain useful neighborhood information.

[0004] Chinese Patent CN114004871A discloses a point cloud registration method and system based on point cloud completion, which performs sampling on the source point cloud and the target point cloud, and extracts features respectively; uses an attention mechanism to fuse the features of the two point clouds to complement the semantic information of the two point clouds; extracts the high-dimensional features of the complemented point cloud, learns the position information of the other point cloud according to the high-dimensional features, and determines the corresponding points of each point in the source point cloud in the target point cloud; according to the corresponding points, uses singular value decomposition to obtain the current rigid transformation parameters, and uses the current rigid transformation parameters to achieve the registration of the source point cloud to the target point cloud. It does not require a large amount of deletion of the original point cloud, and can complement the missing point cloud information to achieve efficient and accurate registration. However, there are problems such as blurred details and distorted shapes in the completion results of the missing parts. Moreover, in order to improve the completion effect, the multi-stage network structure is becoming more and more complex. And point cloud completion, as an upstream task, should be committed to designing a more efficient network structure.

[0005] Most existing completion networks use a pointwise shared multi-layer perception mechanism to extract the features of point clouds, without fully utilizing the local structural information between points. Although some subsequent studies have used methods such as sphere queries to extract multi-scale and multi-resolution point neighborhood information, this method of selecting neighborhood points based on Euclidean space is very sensitive to the distribution and density of points. In the completion task, the missingness and sparsity of the input point cloud make it difficult to obtain useful neighborhood information. Moreover, in the feature extraction encoder part, there is also a risk of information loss due to only using the max pooling operation. Summary of the Invention

[0006] To solve the above problems, while fully utilizing the prior information of the observable part of the point cloud and learning relevant structural attributes, it is also possible to generate a complete and fine shape structure of the spatial object while retaining the detailed features of the input point cloud, and avoid information loss caused by only using the max pooling operation.

[0007] To achieve the above effects, the present invention designs a point cloud completion method based on dynamic graph convolution and attention mechanism.

[0008] A point cloud completion method based on dynamic graph convolution and attention mechanism, which includes:

[0009] Step S1, using dynamic graph convolution technology for feature extraction;

[0010] Step S2, combining the attention pooling method and the max pooling method for feature aggregation;

[0011] Step S3, local missing space target feature completion and reconstruction.

[0012] Preferably, the method of using dynamic graph convolution technology for feature extraction in step S1 includes:

[0013] Step S11, defining a neighborhood and updating the dynamic neighborhood graph;

[0014] Step S12, dense connection.

[0015] Preferably, the specific method of defining a neighborhood and updating the dynamic neighborhood graph in step S11 is to use the k-nearest neighbor algorithm k-NN to construct a local graph for each point in the feature space to aggregate local and structure-aware context features; for each point x in the point cloud i , according to their proximity in the feature space, select k neighbor nodes x j = {x1, x2,..., x k}, combine the feature embedding x i and x j - x i , to obtain the edge feature eij , can be expressed as:

[0016]

[0017] where N i represents the neighborhood of point i, represents the connection operation.

[0018] Preferably, the specific method of dense connection in step S12 is to use a dynamic graph convolutional module to convert the input features into new feature embeddings with a smaller feature dimension, and dynamically construct a neighborhood graph through the converted features; then transfer the edge features of the constructed neighborhood graph to a shared multi-layer perceptron layer; in order to retain the information of the low-dimensional layer, the output of each layer is used as the input of all subsequent layers, which is expressed as:

[0019]

[0020] where is the output feature of point x at the l-th layer i , is the output feature of point x at the (l - 1)-th layer i , h θ (·) represents the shared multi-layer perceptron layer, represents the connection operation; finally, we use max pooling to obtain the aggregated local features with permutation invariance in each local graph;

[0021] Outside the sub-units, the output of each sub-unit is also passed to each subsequent sub-unit as its input through a skip connection operation, which can be denoted as:

[0022]

[0023] where F i represents the output feature of the i-th sub-unit, E i (·) represents the i-th sub-unit, F in represents the input feature, represents the connection operation.

[0024] Preferably, the feature aggregation method in step S2 is: by combining attention pooling and max pooling, and aggregating with the per-point local feature F p to generate F a , which is used to extract global features.

[0025] Preferably, the aggregated feature F a is used as the input of the next stage to generate a fine-grained shape while retaining the original features of the input point cloud.

[0026] Preferably, the attention pooling method is to calculate the attention score a of each element of the input featurei :

[0027]

[0028] Among them, FC(·) represents the fully connected layer, represents the input feature of the i-th point;

[0029] Then multiply each element by the corresponding score and sum them up to obtain the output feature f out It is expressed as:

[0030]

[0031] where f out represents the output feature, and h θ (·) represents the shared multi-layer perceptron layer.

[0032] Preferably, the feature completion and reconstruction method in step S3 is: use the feature completion module to expand and refine the features, and reconstruct the coordinates of the complete point cloud accordingly.

[0033] Preferably, the feature completion module is a hierarchical residual structure.

[0034] Preferably, the feature completion and reconstruction method in step S3:

[0035] S301. First, obtain a feature matrix L0 of size N×C through the shared multi-layer perceptron, where C represents the feature dimension and N represents the number;

[0036] S302. Then, copy the transformed feature matrix r times to obtain a new feature map;

[0037] S303. Generate r different two-dimensional grids, each containing m grid points, and then expand them to the size of the repeated features, and concatenate the coordinates of the grid points with the repeated features;

[0038] S304. Obtain an expanded feature matrix H0 of size rN×C through the self-attention mechanism unit and the shared multi-layer perceptron;

[0039] S305. Reshape the feature H0 to restore it to a feature matrix L1 with the same dimension as L0, subtract L1 from L0 to obtain the residual △L;

[0040] S306. Pass △L through the steps of S302 - S304 again to obtain the expanded residual feature H1, and finally add H1 to H0 to obtain the expanded feature H out ;

[0041] S307. Finally, obtain a complete point cloud of size rN×3 through the shared multi-layer perceptron with parameters [C, 128, 64, 3].

[0042] The advantages and effects of this application are as follows:

[0043] 1. This application defines the neighborhood of each point in the feature space, that is, it finds neighboring points according to the similarity of feature maps, and dynamically updates the neighborhood graph after each layer, thereby obtaining the context information of the point cloud structure and content perception.

[0044] 2. While using max pooling, this application combines attention pooling to focus on some specific significant structural information, and proposes a multi-channel method that jointly performs max pooling and attention pooling operations, thereby extracting more abundant global geometric features of the point cloud.

[0045] 3. This application proposes a lightweight point cloud completion network that combines graph convolution and attention mechanism, efficiently generates the completion results of locally missing point clouds, and maximally preserves and restores the details and geometric structure of the input point cloud.

[0046] 4. This application proposes a network model in a one-stage mode, rather than using a multi-step completion strategy, which can directly output a complete point cloud. Compared with previous methods, our method does not generate a complete point cloud by directly decoding the feature embedding compressed by the pooling operation, but fuses the pointwise features and global geometric features, completes the point cloud in the feature space, and expands and refines the features to reconstruct the complete point cloud.

[0047] The above description is only an overview of the technical solution of this application. In order to be able to understand the technical means of this application more clearly, so that it can be implemented in accordance with the content of the specification, and in order to make the above and other purposes, features and advantages of this application more obvious and understandable, the following will be described in detail with reference to the preferred embodiments of this application and the accompanying drawings.

[0048] According to the following detailed description of the specific embodiments of this application in conjunction with the accompanying drawings, those skilled in the art will be more clear about the above and other purposes, advantages and features of this application. Brief Description of the Drawings

[0049] In order to more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are some embodiments of this application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings. In all the drawings, similar elements or parts are generally denoted by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to actual scale.

[0050] Figure 1The combined graph convolutional and attention mechanism deep learning model architecture for a point cloud completion method based on dynamic graph convolution and attention mechanism provided by the present invention;

[0051] Figure 2 The schematic diagram of dynamic graph convolution for a point cloud completion method based on dynamic graph convolution and attention mechanism provided by the present invention;

[0052] Figure 3 The schematic diagram of feature completion and coordinate reconstruction for a point cloud completion method based on dynamic graph convolution and attention mechanism provided by the present invention;

[0053] Figure 4 The effect comparison diagram for a point cloud completion method based on dynamic graph convolution and attention mechanism provided by the present invention. Detailed implementation manners

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are some, but not all, of the embodiments of the present application. In the following description, specific details such as specific configurations and components are provided only to assist in a comprehensive understanding of the embodiments of the present application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. In addition, descriptions of known functions and structures are omitted for clarity and conciseness.

[0055] It should be understood that the "one embodiment" or "the present embodiment" mentioned throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of the present application. Therefore, the "one embodiment" or "the present embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. In addition, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0056] In addition, the present application may repeat reference numerals and / or letters in different examples. This repetition is for the purpose of simplicity and clarity and does not in itself indicate the relationship between the various embodiments and / or settings discussed.

[0057] In this text, the term "and / or" is merely a description of the relationship between associated objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, B exists alone, and both A and B exist simultaneously. In this text, the term " / and" describes another relationship between associated objects, indicating that there can be two relationships. For example, A / and B can represent two situations: A exists alone, and both A and B exist. Additionally, in this text, the character " / " generally indicates that the associated objects before and after are in an "or" relationship.

[0058] In this text, the term "at least one" is merely a description of the relationship between associated objects, indicating that there can be three relationships. For example, at least one of A and B can represent three situations: A exists alone, both A and B exist simultaneously, and B exists alone.

[0059] It should also be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion.

[0060] Embodiment 1

[0061] This embodiment mainly introduces a point cloud completion method based on dynamic graph convolution and attention mechanism. Please refer to the overall model architecture diagram Figure 1 , Figure 1 which is a deep learning model architecture combining graph convolution and attention mechanism for the point cloud completion method provided by the present invention.

[0062] A point cloud completion method based on dynamic graph convolution and attention mechanism, which includes:

[0063] S1. Feature extraction;

[0064] S2. Feature aggregation;

[0065] S3. Feature completion and reconstruction.

[0066] Furthermore, the feature extraction method in the S1 step includes:

[0067] S21. Define the neighborhood;

[0068] S22. Update the dynamic neighborhood graph;

[0069] S23. Dense connection.

[0070] Furthermore, the method for defining the neighborhood and updating the dynamic neighborhood graph is specifically to use the k-nearest neighbor algorithm (k-NN) to construct a local graph for each point in the feature space to aggregate non-local and structure-aware context features; for each point x in the point cloud i , k neighbor nodes x j ={x1, x2, …, x k} are selected according to their proximity in the feature space, and by combining the feature embedding x i and x j -x i the edge feature e ij is obtained, which can be expressed as:

[0071]

[0072] where N i represents the neighborhood of point i, represents the concatenation operation.

[0073] Furthermore, the dense connection method is specifically to use a dynamic graph convolutional module, for details please refer to Figure 2 , Figure 2 which is the schematic diagram of the dynamic graph convolution of a point cloud completion method based on dynamic graph convolution and attention mechanism provided by the present invention.

[0074] The input features are converted into new feature embeddings with a smaller feature dimension, and a neighborhood graph is dynamically constructed through the converted features; then the edge features of the constructed neighborhood graph are passed to a shared multi-layer perceptron layer; in order to retain the information of the low-dimensional layer, the output of each layer is used as the input of all subsequent layers, which is expressed as:

[0075]

[0076] where is the output feature of point x i at the l-th layer, is the output feature of point x i at the (l - 1)-th layer, h θ (·) represents the shared multi-layer perceptron layer, represents the concatenation operation; finally, we use max pooling to obtain the aggregated local features with permutation invariance in each local graph;

[0077] Outside the sub-units, the output of each sub-unit is also passed to each subsequent sub-unit as its input through a skip connection operation, which can be denoted as:

[0078]

[0079] where F i represents the output feature of the i-th sub-unit, E i(·) represents the i-th sub-unit, F in represents the input feature, represents the connection operation.

[0080] Furthermore, the feature aggregation method in the S2 step is: through the combination of attention pooling and max pooling, the point-wise local feature F p is aggregated to generate the aggregated feature F a , which is used to extract the global feature.

[0081] Furthermore, the aggregated feature F a is used as the input for the next stage, generating a fine-grained shape while retaining the original features of the input point cloud.

[0082] Furthermore, the attention pooling method is to calculate the attention score a of each element of the input feature i :

[0083]

[0084] where FC(·) represents the fully connected layer, represents the input feature of the i-th point;

[0085] Then each element is multiplied by the corresponding score and summed to obtain the output feature f out which is expressed as:

[0086]

[0087] where f out represents the output feature, h θ (·) represents the shared multi-layer perceptron layer.

[0088] Furthermore, the feature completion and reconstruction method in the S3 step is: using the feature completion module to expand and refine the features, and reconstruct the coordinates of the complete point cloud based on this. For details, please refer to Figure 3 , Figure 3 which is the schematic diagram of feature completion and coordinate reconstruction of a point cloud completion method based on dynamic graph convolution and attention mechanism provided by the present invention.

[0089] Furthermore, the feature completion module is a hierarchical residual structure.

[0090] Furthermore, the feature completion and reconstruction method in the S3 step:

[0091] S301. First, a feature matrix L0 of size N×C is obtained through the shared multi-layer perceptron, where C represents the feature dimension and N represents the number;

[0092] S302. Then, the transformed feature matrix is copied r times to obtain a new feature map;

[0093] S303. Generate r different two-dimensional grids, each containing m grid points, then expand them to the size of the repeated feature, and concatenate the coordinates of the grid points with the repeated feature;

[0094] S304. Obtain the dilated feature matrix H0 of size rN×C through the self-attention mechanism unit and the shared multi-layer perceptron;

[0095] S305. Reshape the feature H0 to restore the feature matrix L1 with the same dimension as L0, subtract L1 from L0 to obtain the residual △L;

[0096] S306. Pass △L through the steps of S302 - S304 to obtain the dilated residual feature H1, and finally add H1 to H0 to obtain the dilated feature H out ;

[0097] S307. Finally, obtain the complete point cloud of size rN×3 through the shared multi-layer perceptron with parameters [C, 128, 64, 3].

[0098] This application defines the neighborhood of each point in the feature space, that is, searches for neighboring points according to the similarity of the feature map, and dynamically updates the neighborhood map after each layer to obtain the structure and content-aware context information.

[0099] This application combines attention pooling while using max pooling to focus on some specific significant structure information, and proposes a multi-channel method that combines max pooling and attention pooling operations, thereby extracting more rich global geometric features.

[0100] This application proposes a lightweight point cloud completion network that combines graph convolution and attention mechanism, efficiently generates the completion result of the locally missing point cloud, and maximally retains and restores the details and geometric structure of the input point cloud.

[0101] This application proposes a one-stage mode network model, which does not use a multi-step completion strategy and can directly output the complete point cloud. Compared with previous methods, our method does not generate the complete point cloud by directly decoding the feature embedding compressed by the pooling operation, but fuses the point-wise features and global geometric features, completes the point cloud in the feature space, and expands and refines the features to reconstruct the complete point cloud.

[0102] Embodiment 2

[0103] Based on the above Embodiment 1, this embodiment mainly introduces the network training model in the verification process of a point cloud completion method based on dynamic graph convolution and attention mechanism.

[0104] First, establish the hyperparameters of the network model: The model is implemented using Pytorch. All network models are trained using the Adam optimizer, with β1 = 0.9, β2 = 0.999, the initial learning rate of the generator is 5e-4, the initial learning rate of the discriminator is 1e-5, and it decays by 0.7 every 40 epochs. The batch size during training is set to 32, and our network converges at around 150 epochs of training.

[0105] Secondly, establish the loss function: The overall loss function of this paper consists of two parts. Among them, the completion loss (L c ) is used to ensure that the output point cloud approaches the ground truth point cloud, and the uniformity loss (L uni ) constrains the network to output a point cloud with a uniform distribution. The specific definitions are as follows: This paper selects the Chamfer Distance (Equation 6) as the completion loss function, and calculates the corresponding CD value for the complete point cloud Q output by the network and the corresponding ground truth point cloud Q gt .

[0106]

[0107] To output a point cloud with a uniform distribution, this paper introduces the uniformity loss (Equation 7). Among them, U imbalance constrains the number of points in the local neighborhood, and U clutter constrains the geometric distribution of points in the local neighborhood. Among them, S j (j = 1,…, M) represents a subset of points, and each subset is sampled by spherical query with a radius of r d ; is the expected number of points in the S j subset; represents the expected distance between a point and its neighboring points; this formula is derived based on the assumption that S j is planar and adjacent points are hexagonal.

[0108]

[0109] The overall loss function is the weighted sum of the above two loss functions.

[0110] Among them, β and γ are the weight values of the corresponding loss functions:

[0111] L = βL c + γL uni . (8)

[0112] Finally, a training dataset was established: We conducted experiments on the MVP dataset (Multi-View Partial Point Cloud Dataset). The MVP dataset is from the research of Pan et al. (2021) and has high-quality multi-view partial missing point clouds. It contains objects of 16 categories, with a total of 62,400 groups of data in the training set and 41,600 groups of data in the test set. The MVP dataset provides input point clouds of 2048 points and complete point clouds of different resolutions, including complete point clouds of 2048, 4096, 8192, and 16384 points, for evaluating the quality of completion at different resolutions.

[0113] Example 3

[0114] Based on the above Examples 1-2, this example mainly introduces the testing of the network training model in the verification process of a point cloud completion method based on dynamic graph convolution and attention mechanism.

[0115] Establish model evaluation parameters: The accuracy of the model completion is evaluated by calculating the Chamfer Distance (CD) between the predicted point cloud Q and the ground truth point cloud Q gt . The smaller the CD value, the better the completion effect of the model.

[0116] Model test results: The completion effect of our model was tested on the test set of the MVP dataset and compared with the following methods in the same environment:

[0117] 1. PCN (Yuan et al., 2018) generates a complete point cloud in a coarse-to-fine manner, uses two stacked shared multi-layer perceptron layers as the encoder to extract global features, and combines a fully connected-based decoder and a folding operation-based decoder to generate a dense complete point cloud;

[0118] 2. MSN (Liu et al., 2019) also completes the completion task with a two-stage model. In the first stage, a set of object surface patches is generated. In the second stage, the input point cloud is fused with the roughly predicted point cloud, and the lowest density sampling is proposed to sample the fused point cloud to obtain a uniformly distributed point cloud, and then a residual network is used to refine the point cloud to obtain the final complete point cloud;

[0119] 3. CRN (Wang et al., 2020) uses a cascaded refinement strategy to refine the positions of points locally and globally in a coarse-to-fine manner and designs a patch discriminator to further ensure the authenticity of each local area by using adversarial training;

[0120] 4. ECG (Pan, 2020) uses graph convolution to propagate multi-scale edge feature information in the refinement stage to achieve the purpose of retaining local geometric details;

[0121] 5. The variational relation network proposed by VRCNet (Pan et al., 2021) adopts a double parallel path mode in the feature extraction stage. The defective point cloud and the corresponding complete point cloud are input respectively, and through certain constraint conditions, the feature information extracted from the defective point cloud is made as similar as possible to the complete point cloud to obtain richer feature information.

[0122] The quantitative and qualitative results tested on the MVP dataset are shown in Table 1 and Figure 4 as follows. Figure 4 This is a comparison chart of the effects of a point cloud completion method based on dynamic graph convolution and attention mechanism provided by the present invention. Through Figure 4 it can be directly seen that the point cloud completion result of Ours is the best and closest to the original point cloud structure.

[0123] Table 1 Point cloud completion results on the MVP test dataset, expressed as the CD values (×10 4 )

[0124]

[0125] It can be seen from the comparison results that our method obtains the minimum CD value, and the advantage is that it can maximize the restoration of the original structure of the input point cloud, and use the geometric features of the known point cloud to learn similar structure information to complete the missing part, and can obtain a relatively realistic completion result.

[0126] The above are only the preferred embodiments of the present invention, and it does not limit the protection scope of the present invention. For those skilled in the art, the present invention can have various changes and modifications. All changes, modifications, substitutions, integrations and parameter changes made within the spirit and principle of the present invention by means of conventional substitutions or capable of achieving the same functions without departing from the principle and spirit of the present invention fall within the protection scope of the present invention.

Claims

1. A point cloud completion method based on dynamic graph convolution and attention mechanism, characterized in that, It includes: Step S1: Extract features using the dynamic graph convolution technique; Step S2: Aggregate features by combining the attention pooling method and the max pooling method; Step S3: Complement and reconstruct the local missing spatial target features; In the S3 step, the feature complementation and reconstruction method is to use a feature complementation module to expand and refine the features, and reconstruct the coordinates of the complete point cloud with this. The feature complementation module is a hierarchical residual structure, which specifically includes the following steps: S301: First, obtain a feature matrix L0 of size N×C through a shared multi-layer perceptron, where C represents the feature dimension and N represents the number; S302: Then, copy the transformed feature matrix r times to obtain a new feature map; S303: Generate r different two-dimensional grids, each containing m grid points, and then expand them to the size of the repeated features, and concatenate the coordinates of the grid points with the repeated features; S304: Obtain an expanded feature matrix H0 of size rN×C through a self-attention mechanism unit and a shared multi-layer perceptron; S305: Reshape the feature H0 to restore it to a feature matrix L1 with the same dimension as L0, and subtract L0 from L1 to obtain the residual △L; S306. Obtain the dilated residual feature H1 from △L through the steps of S302 - S304, and finally add H1 to H0 to obtain the dilated feature H out ; S307: Finally, obtain a complete point cloud of size rN×3 through a shared multi-layer perceptron with parameters [C, 128, 64, 3].

2. The point cloud completion method based on dynamic graph convolution and attention mechanism according to claim 1, wherein, The method of using the dynamic graph convolution technique for feature extraction in the step S1 includes: Step S11: Define the neighborhood and update the dynamic neighborhood graph; Step S12: Dense connection.

3. The point cloud completion method based on dynamic graph convolution and attention mechanism according to claim 2, characterized in that The specific method of defining the neighborhood and updating the dynamic neighborhood graph in step S11 is to use the k-nearest neighbor algorithm (k-NN) to construct a local graph for each point in the feature space to aggregate local and structure-aware context features; for each point x in the point cloud i , k neighbor nodes x j ={x1, x2, …, x k} are selected according to their proximity in the feature space, and the edge feature e i is obtained by combining the feature embedding x j and x i -x ij , which can be expressed as: where N i represents the neighborhood of point i, represents the connection operation.

4. A point cloud completion method based on dynamic graph convolution and attention mechanism according to claim 2, characterized in that The specific method of dense connection in the step S12 is to use a dynamic graph convolution module to convert the input features into new feature embeddings with a smaller feature dimension, and dynamically construct a neighborhood graph through the transformed features; then transfer the edge features of the constructed neighborhood graph to the shared multi-layer perceptron layer; in order to retain the information of the low-dimensional layer, the output of each layer is used as the input of all subsequent layers, which can be expressed as: where f i l is the output feature of the point x i at the l-th layer, and f i l-1 is the output feature of the point x i at the (l - 1)-th layer, h θ (·) represents a shared multi-layer perceptron layer, and represents a concatenation operation; Outside the sub-unit, the output of each sub-unit is also passed to each subsequent sub-unit as its input through a skip connection operation, which can be denoted as: Among them, F i represents the output feature of the i-th sub-unit, E i (·) represents the i-th sub-unit, F in represents the input feature, represents the connection operation.

5. A point cloud completion method based on dynamic graph convolution and attention mechanism according to claim 1, characterized in that In the S2 step, the feature aggregation method is as follows: By combining attention pooling and max pooling, the per-point local feature F p is aggregated to generate an aggregated feature F a , which is used to extract global features.

6. The point cloud completion method based on dynamic graph convolution and attention mechanism according to claim 5, wherein The aggregated feature F a As the input for the next stage, it generates a fine-grained shape while preserving the original features of the input point cloud.

7. A point cloud completion method based on dynamic graph convolution and attention mechanism according to claim 5, characterized in that The attention pooling method is to calculate the attention score a of each element of the input feature i : where FC(·) represents the fully connected layer, represents the input feature of the i-th point; Then each element is multiplied by the corresponding score and summed to obtain the output feature f out Expressed as: where f out represents the output feature, h θ (·) represents a shared multi-layer perceptron layer, and N represents the number of them.

Images