CN114693873A - A point cloud completion method based on dynamic graph convolution and attention mechanism - Google Patents

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 maximum pooling The feature aggregation method is performed by the method; step S3, the local missing spatial target feature is completed and reconstructed. This application defines the point cloud neighborhood and updates the dynamic neighborhood map, and then combines the dynamic neighborhood map to update and densely connect the point cloud; and uses a multi-channel method combining attention pooling and max pooling to characterize the point cloud aggregation, efficiently completes the completion of the missing point cloud, and preserves and restores the details and geometric structure of the input point cloud to the greatest extent; and this application proposes a network model of one-stage mode, which integrates the point-by-point features and The global geometric feature complements the point cloud in the feature space, expands and refines the feature, and reconstructs the complete point cloud, and then directly outputs the complete point cloud.

Description

A point cloud completion method based on dynamic graph convolution and attention mechanism

technical field

The invention belongs to the field of model construction and design, in particular to a point cloud completion method based on dynamic graph convolution and attention mechanism.

Background technique

In the real scene, the point cloud obtained by 3D scanning equipment such as lidar, due to the limitation of sensor resolution, viewing angle, and the influence of the occlusion between object structures or objects, the collected space target point cloud has problems such as sparseness and lack. Subsequent application of point clouds at a deeper level brings certain difficulties. Therefore, point cloud shape completion is to solve the problem of how to recover the complete point cloud from the locally observed point cloud. For any given point cloud with missing structure, the corresponding point cloud with complete shape can be obtained. This task is a lot of downstream The basis of tasks, such as shape classification, segmentation, etc., is also an indispensable operation link in subsequent applications.

Chinese patent CN112614071A discloses a self-attention-based multiple point cloud completion method and device, which relates to the technical field of computer three-dimensional point cloud completion and deep learning, wherein the method includes: acquiring point cloud data, processing the point cloud data , obtain the input point proxy sequence; encode the point proxy sequence, obtain the point encoding vector, decode the point encoding vector, and obtain the predicted point proxy; input the predicted point proxy into the multilayer perceptron to obtain the predicted point center, and at the predicted point The complete point cloud data is recovered on the basis of the center. Therefore, the point cloud is processed into a point proxy sequence, and the encoder and decoder are used to construct the long-range relationship between different points of the point cloud to achieve point cloud reconstruction. However, it uses point-by-point shared multi-layer perceptron to extract the features of point clouds, and does not make full use of the local structural information between points. Domain information becomes difficult.

Chinese patent CN114004871A discloses a point cloud registration method and system based on point cloud completion. The source point cloud and the target point cloud are sampled to extract features respectively; The semantic information of the point cloud complements each other; extract the high-dimensional features of the completed point cloud, learn the position information of the opposite point cloud according to the high-dimensional features, and determine the corresponding point of each point in the source point cloud in the target point cloud ; According to the corresponding points, the current rigid transformation parameters are obtained by singular value decomposition, and the registration of the source point cloud to the target point cloud is realized by using the current rigid change parameters. It does not need to delete a lot of the original point cloud, and can complete 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 network structure based on multi-stage becomes more and more complex. As an upstream task, point cloud completion should be devoted to designing a more efficient network structure.

Most of the existing completion networks use the point-by-point shared multi-layer perception mechanism to extract the features of point clouds, and do not fully utilize the local structural information between points. Later, although some researches used methods such as ball query to extract multi-scale and multi-resolution. However, this Euclidean space-based neighborhood point selection method is very sensitive to the distribution and density of points. In the completion task, the lack and sparseness of the input point cloud make it possible to obtain useful neighbors. Domain information becomes difficult. Moreover, in the feature extraction encoder part, there is a certain risk of information loss due to only using the maximum pooling operation.

SUMMARY OF THE INVENTION

In order to solve the above problems, it is possible to make full use of the prior information of the observable part of the point cloud, while learning the relevant structural attributes, while retaining the detailed characteristics of the input point cloud, it can also generate a complete and fine shape structure of the spatial object, avoiding the use of only the maximum pooling. The operation results in the loss of information.

In order to achieve the above effects, the present invention designs a point cloud completion method based on dynamic graph convolution and attention mechanism.

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 the attention pooling method and the maximum pooling method to perform feature aggregation;

Step S3: Completion and reconstruction of local missing spatial target features.

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

Step S11, define the neighborhood and update the dynamic neighborhood graph;

Step S12, dense connection.

Preferably, the specific method of defining the neighborhood and updating the dynamic neighborhood graph in the step S11 is to use the nearest neighbor algorithm k-NN to construct a local graph of each point in the feature space to aggregate local and structure-aware contextual features; For each point x _i in the point cloud, select k neighbor nodes x _j ={x ₁ ,x ₂ ,...,x _k } according to their proximity in the feature space, combining the feature embeddings x _i and x _j -x _i , get the edge feature e _ij , which can be expressed as:

表示连接操作。where Ni represents the neighborhood of point _i ,

Represents a connect operation.

Preferably, the specific method for dense connection in the step S12 is to use a dynamic graph convolution module to convert the input features into new feature embeddings with smaller feature dimensions, and dynamically construct a neighborhood graph through the converted features; then The edge features of the constructed neighborhood graph are passed 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, expressed as:

是第l层点x_i的输出特征，

是第l-1层点x_i的输出特征，h_θ(·)表示共享多层感知机层，

表示连接操作；最后，我们使用最大池化来获取在每个局部图中具有排列不变性的聚合局部特征；in

is the output feature of the point _xi of the lth layer,

is the output feature of the point x _i of the l-1th layer, h _θ ( ) represents the shared multi-layer perceptron layer,

represents the join operation; finally, we use max pooling to obtain aggregated local features with permutation invariance in each local graph;

In addition to the subunits, the output of each subunit is also passed to each subsequent subunit as its input through the skip connection operation, which can be written as:

表示连接操作。where F _i represents the output feature of the ith subunit, E _i ( ) represents the ith subunit, F _in represents the input feature,

Represents a connect operation.

Preferably, the feature aggregation method in the step S2 is as follows: attention pooling and max pooling are combined to extract global features, and aggregated with point-by-point local features F _p to generate _Fa .

Preferably, the feature aggregation is to use the point-by-point feature F _p and the global feature aggregation as the input of the next stage to generate fine-grained shapes while retaining the original features of the input point cloud.

Preferably, the attention pooling method is to calculate the attention score a _i of each element of the input feature:

表示第i个点的输入特征；where FC( ) represents the fully connected layer,

Represents the input feature of the i-th point;

Each element is then multiplied by the corresponding score and added to get the aggregated feature _fout expressed as:

where f _out represents the output features and h _θ ( ) represents the shared multilayer perceptron layer.

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

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

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

S101. First, a feature matrix L ₀ with a size of N×C is obtained by sharing a multi-layer perceptron, where C represents the feature dimension, and N represents the number;

S102, then copy r times to the transformed feature matrix to obtain a new feature map;

S103, generating r different two-dimensional grids, including m grid points, and then expanding them to the size of the repeated features, and splicing the coordinates of the grid points with the repeated features;

S104, obtaining an expanded feature matrix H ₀ with a size of rN ₃ ×C through the self-attention mechanism unit and the shared multilayer perceptron;

S105, reshape the feature H ₀ to restore the feature matrix L ₁ with the same dimension as L ₀ , and make a difference between L ₁ and L ₀ to obtain the residual ΔL;

S106, obtaining the expanded residual feature H ₁ through the steps of S102-S104 by ΔL, and finally adding H ₁ and H ₀ to obtain the expanded feature H _out ;

S107. Finally, a complete point cloud with a size of rN ₃ × 3 is obtained through a shared multilayer perceptron with parameters [C, 128, 64, 3].

The advantages and effects of the present application are as follows:

1. This application defines the neighborhood of each point in the feature space, that is, finds adjacent points according to the similarity of the feature map, and dynamically updates the neighborhood map after each layer, thereby obtaining the context of the point cloud structure and content perception information.

2. While using max pooling, this application combines attention pooling to focus on some specific salient structural information, and proposes a multi-channel method of combining max pooling and attention pooling operations to extract point clouds More informative global geometric features.

3. This application proposes a lightweight point cloud completion network combining graph convolution and attention mechanisms, which efficiently generates the completion results of locally missing point clouds, and preserves and restores the details and geometric structure of the input point cloud to the greatest extent. .

4. This application proposes a network model of one-stage mode, instead of using a multi-step completion strategy, it can directly output a complete point cloud. Compared with previous methods, our method does not compress through direct decoding and pooling operation. Instead, point-by-point features and global geometric features are fused to complete the point cloud in the feature space, and the features are expanded and refined to reconstruct the complete point cloud.

The above description is only an overview of the technical solution of the present application, in order to be able to understand the technical means of the present application more clearly, so that it can be implemented according to the contents of the description, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand , the preferred embodiments of the present application and the accompanying drawings are described in detail below.

The above and other objects, advantages and features of the present application will be more apparent to those skilled in the art from the following detailed description of the specific embodiments of the present application in conjunction with the accompanying drawings.

Description of drawings

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are For some embodiments of the present application, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without any creative effort. Similar elements or parts are generally identified by similar reference numerals throughout the drawings. In the drawings, each element or section is not necessarily drawn to actual scale.

1 is a deep learning model architecture combining graph convolution and attention mechanism of a point cloud completion method based on dynamic graph convolution and attention mechanism provided by the present invention;

2 is a schematic diagram of dynamic graph convolution of a point cloud completion method based on dynamic graph convolution and attention mechanism provided by the present invention;

3 is a 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;

FIG. 4 is a comparison diagram of the effect of a point cloud completion method based on dynamic graph convolution and attention mechanism provided by the present invention.

Detailed ways

In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments It is a part of the embodiments of the present application, but not all of the embodiments. In the following description, specific details such as specific configurations and components are provided merely to assist in a comprehensive understanding of embodiments of the present application. Accordingly, it should be apparent to those skilled in the art that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the present application. Also, descriptions of well-known functions and constructions are omitted in the embodiments for clarity and conciseness.

It should be understood that reference throughout the specification to "one embodiment" or "the present embodiment" means that a particular feature, structure or characteristic associated with the embodiment is included in at least one embodiment of the present application. Thus, appearances of "one embodiment" or "this embodiment" in various places throughout this specification are not necessarily necessarily referring to the same embodiment. Furthermore, the particular features, structures or characteristics may be combined in any suitable manner in one or more embodiments.

Furthermore, this application may repeat reference numerals and/or letters in different instances. This repetition is for the purpose of simplicity and clarity and does not in itself indicate a relationship between the various embodiments and/or arrangements discussed.

The term "and/or" in this article is only an association relationship to describe associated objects, indicating that there can be three kinds of relationships, for example, A and/or B, which can mean: A alone exists, B alone exists, and A and B exist simultaneously. There are three cases of B. In this article, the term "/and" is to describe another related object relationship, which means that there can be two relationships, for example, A/ and B, which can mean that A exists alone, and A and B exist alone. , In addition, the character "/" in this text generally indicates that the related objects are an "or" relationship.

The term "at least one" in this paper is only an association relationship to describe the associated objects, which means that there can be three kinds of relationships, for example, at least one of A and B, it can mean that A exists alone, A and B exist at the same time, There are three cases of B alone.

It should also be noted that in this document, relational terms such as first and second are used only to distinguish one entity or operation from another, and do not necessarily require or imply those entities or operations There is no such actual relationship or order between them. Moreover, the terms "comprising", "comprising" or any other variation thereof are intended to encompass non-exclusive inclusion.

Example 1

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

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

S1. Feature extraction;

S2, feature aggregation;

S3. Feature completion and reconstruction.

Further, the feature extraction method in the step S1 includes:

S21, define a neighborhood;

S22, update the dynamic neighborhood graph;

S23, dense connection.

Further, the method for defining neighborhoods and updating dynamic neighborhood graphs is specifically to use the nearest neighbor algorithm k-NN to construct a local graph of each point in the feature space to aggregate non-local and structure-aware contextual features; for point clouds For each point x _i in , select k neighbor nodes x _j ={x ₁ ,x ₂ ,...,x _k } according to their proximity in the feature space, and combine the feature embeddings x _i and x _{j -xi} to get The edge feature e _ij can be expressed as:

表示连接操作。where Ni represents the neighborhood of point _i ,

Represents a connect operation.

Further, the dense connection method specifically uses a dynamic graph convolution module, please refer to FIG. 2 for details. FIG. 2 is a dynamic graph of a point cloud completion method based on dynamic graph convolution and attention mechanism provided by the present invention. Schematic diagram of convolution.

Transform the input features into new feature embeddings with smaller feature dimensions, and dynamically construct a neighborhood graph from the transformed features; then pass the edge features of the constructed neighborhood graph to the shared multilayer perceptron layer; in order to preserve low The information of the dimension layer, the output of each layer is used as the input of all subsequent layers, which is expressed as:

是第l层点x_i的输出特征，

是第l-1层点x_i的输出特征，h_θ(·)表示共享多层感知机层，

表示连接操作；最后，我们使用最大池化来获取在每个局部图中具有排列不变性的聚合局部特征；in

is the output feature of the point _xi of the lth layer,

is the output feature of the point x _i of the l-1th layer, h _θ ( ) represents the shared multi-layer perceptron layer,

represents the join operation; finally, we use max pooling to obtain aggregated local features with permutation invariance in each local graph;

In addition to the subunits, the output of each subunit is also passed to each subsequent subunit as its input through the skip connection operation, which can be written as:

表示连接操作。where F _i represents the output feature of the ith subunit, E _i ( ) represents the ith subunit, F _in represents the input feature,

Represents a connect operation.

Further, the feature aggregation method in the step S2 is as follows: attention pooling and max pooling are combined to extract global features, and aggregated with point-by-point local features F _p to generate F _a .

Further, the feature aggregation is to use the point-by-point feature F _p and the global feature aggregation as the input of the next stage to generate fine-grained shapes while retaining the original features of the input point cloud.

Further, the attention pooling method is to calculate the attention score a _i of each element of the input feature:

表示第i个点的输入特征；where FC( ) represents the fully connected layer,

Represents the input feature of the i-th point;

Each element is then multiplied by the corresponding score and added to get the aggregated feature _fout expressed as:

where f _out represents the output features and h _θ ( ) represents the shared multilayer perceptron layer.

Further, 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. Please refer to FIG. 3 for details. FIG. 3 is a 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.

Further, the feature completion module is a hierarchical residual structure.

Further, the feature completion and reconstruction method in the S3 step:

S101. First, a feature matrix L ₀ with a size of N×C is obtained by sharing a multi-layer perceptron, where C represents the feature dimension, and N represents the number;

S102, then copy r times to the transformed feature matrix to obtain a new feature map;

S103, generating r different two-dimensional grids, including m grid points, and then expanding them to the size of the repeated features, and splicing the coordinates of the grid points with the repeated features;

S104, obtaining an expanded feature matrix H ₀ with a size of rN ₃ ×C through the self-attention mechanism unit and the shared multilayer perceptron;

S105, reshape the feature H ₀ to restore the feature matrix L ₁ with the same dimension as L ₀ , and make a difference between L ₁ and L ₀ to obtain the residual ΔL;

S106, obtaining the expanded residual feature H ₁ through the steps of S102-S104 by ΔL, and finally adding H ₁ and H ₀ to obtain the expanded feature H _out ;

S107. Finally, a complete point cloud with a size of rN ₃ × 3 is obtained through a shared multilayer perceptron with parameters [C, 128, 64, 3].

The present application defines the neighborhood of each point in the feature space, that is, finds adjacent points according to the similarity of the feature maps, and dynamically updates the neighborhood map after each layer to obtain the context information of structure and content awareness.

In this application, while using maximum pooling, it combines attention pooling to focus on some specific salient structural information, and proposes a multi-channel method of combining maximum pooling and attention pooling operations, so as to extract richer information. Global geometric features.

This application proposes a lightweight point cloud completion network combining graph convolution and attention mechanisms, which efficiently generates the completion results of locally missing point clouds, and preserves and restores the details and geometric structure of the input point cloud to the greatest extent.

This application proposes a network model of one-stage mode, instead of using a multi-step completion strategy, it can directly output a complete point cloud. Compared with the previous methods, our method does not directly decode the compressed result of the pooling operation. Feature embedding is used to generate a complete point cloud, but point-by-point features and global geometric features are fused, the point cloud is completed in the feature space, and the features are expanded and refined to reconstruct the complete point cloud.

Example 2

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

First establish the network model hyperparameters: The model is implemented using Pytorch. The network models are all trained with the Adam optimizer, and set β ₁ =0.9, β ₂ =0.999, the initial learning rate of the generator is 5e-4, the initial learning rate of the discriminator is 1e-5, and the decay is 0.7 every 40 epochs . With a batch size of 32 during training, our network converges around 150 epochs of training.

Secondly, the loss function is established: the overall loss function in this paper consists of two parts, in which the completion loss (L _c ) is used to ensure that the output point cloud is close to the true point cloud, and the uniform loss (L _u ) constrains the network to output points with uniform distribution cloud. The specific definition is as follows: In this paper, ChamferDistance (Formula 8) is selected as the completion loss function, and the corresponding CD value is calculated from the complete point cloud Q output by the network and the corresponding ground-truth point cloud _Qgt .

是S_j子集中点的期望数量；

表示点与邻域点之间期望的距离；该公式基于S_j是平面的并且相邻点是六边形的假设推导出来的。In order to output a uniformly distributed point cloud, this paper introduces a homogenization loss (Equation 9). Among them, U _imbalance is to constrain the number of points in the local neighborhood, and U _clutter is to constrain the geometric distribution of points in the local neighborhood. where S _j (j=1,...,M) represents a subset of points, each subset is obtained by sampling a spherical query with radius r _d ;

is the expected number of points in the subset of S _j ;

represents the desired distance between a point and a neighbor point; the formula is derived on the assumption that S _j is planar and the neighbor points are hexagonal.

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

where β and γ are the weight values of the corresponding loss function:

L=βL _c +γL _uni . (8)

It is best to build a training dataset: We conduct experiments on the MVP dataset (Multi-View Partial Point CloudDataset). The MVP dataset is derived from the work of Pan et al. (2021) with high-quality multi-view locally missing point clouds. It contains 16 categories of objects, with a total of 62,400 sets of data in the training set and 41,600 sets of data in the test set. The MVP dataset provides input point clouds of 2048 points and full point clouds of different resolutions, including full point clouds of 2048, 4096, 8192 and 16384 points, for evaluating the quality of completion at different resolutions.

Example 3

Based on the foregoing Embodiments 1-2, this embodiment mainly introduces a network training model test in the verification process of a point cloud completion method based on dynamic graph convolution and attention mechanism.

Establish model evaluation parameters: Evaluate the accuracy of model completion 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 model completion effect.

Model test results: The completion effect of our model is tested on the test set of the MVP dataset and compared with the following methods under the same environment:

1. PCN (Yuan et al., 2018), which generates a complete point cloud in a coarse-to-fine pattern, uses two stacked shared MLP layers as encoders to extract global features, combines a fully connected-based decoder and The decoder based on the folding operation generates dense and complete point clouds;

2. MSN (Liu et al., 2019) also completes the completion task with a two-stage model. The first stage generates a set of object surface blocks, and the second stage fuses the input point cloud with the roughly predicted point cloud, and proposes the lowest Density sampling samples the fused point cloud to obtain a uniformly distributed point cloud, and then refines the point cloud through a residual network to obtain the final complete point cloud;

3. CRN (Wang et al., 2020) utilizes a cascade refinement strategy to refine the location of points locally and globally in a coarse-to-fine manner, and designs a block discriminator that utilizes adversarial training to further ensure every part is real;

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

5. The variational relational network proposed by VRCNet (Pan et al., 2021) adopts a dual-parallel path mode in the feature extraction stage, inputs the residual defect cloud and the corresponding complete point cloud respectively, and makes the extracted data from the residual defect cloud through certain constraints. The feature information is as similar as possible to the full point cloud to get richer feature information.

The quantitative and qualitative results tested on the MVP data set are shown in Table 1 and Figure 4. Figure 4 is a comparison diagram of the effect of a point cloud completion method based on dynamic graph convolution and attention mechanism provided by the present invention. From Figure 4, it can be directly seen that the point cloud completion result of Ours is the best and the closest to the original point cloud structure.

Table 1. Point cloud completion results on the MVP test data set, expressed as the CD value (×10 ⁴ ) obtained by each model on each category of objects

From the comparison results, it can be seen that our method achieves the smallest 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 structural information to complete the missing parts. more realistic results can be obtained.

The above descriptions are only preferred embodiments of the present invention, which are not intended to limit the protection scope of the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any changes, modifications, substitutions, integrations and parameter changes to these embodiments without departing from the principles and spirit of the present invention, through conventional substitutions or capable of achieving the same function within the spirit and principles of the present invention, all fall within the scope of the present invention. into the protection scope of the present invention.

Claims (10)

1. A point cloud completion method based on a dynamic graph convolution and attention mechanism is characterized by comprising the following steps:

step S1, extracting features by using a dynamic graph convolution technology;

step S2, combining the attention pooling method and the maximum pooling method to perform feature aggregation;

and S3, completing and reconstructing the local missing space target characteristics.

2. The point cloud completion method based on dynamic graph convolution and attention mechanism as claimed in claim 1, wherein the method for feature extraction using dynamic graph convolution technique in step S1 includes:

step S11, defining neighborhood and updating a dynamic neighborhood map;

and step S12, dense connection.

3. The point cloud completion method based on dynamic graph convolution and attention mechanism as claimed in claim 2, wherein the specific method for defining neighborhood and updating dynamic neighborhood graph in step S11 is to use nearest neighbor algorithm k-NN to construct local graph of each point in feature space to aggregate local and structure-aware context features; for each point x in the point cloud_iSelecting k neighbor nodes x according to their proximity in feature space_j＝{x₁,x₂,…,x_k}, binding feature embedding x_iAnd x_j-x_iTo obtain an edge feature e_ijIt can be expressed as:

wherein N is_iA neighborhood of the point i is represented and,

indicating a connect operation.

4. The point cloud completion method based on dynamic graph convolution and attention mechanism as claimed in claim 2, wherein the specific method of dense connection in step S12 is to use a dynamic graph convolution module to convert the input features into new feature embedding with smaller feature dimension and dynamically construct a neighborhood graph from the converted features; then transmitting the edge characteristics of the constructed neighborhood graph to a shared multilayer perceptron layer; to preserve the information of the lower dimensional layers, the output of each layer is taken as the input of all subsequent layers, expressed as:

wherein

Is the l layer point x_iThe output characteristics of (a) to (b),

is the point x of layer l-1_iOutput characteristic of h_θ() represents the shared multi-layer perceptron layer,

representing a join operation; finally, we use maximal pooling to obtain aggregated local features with permutation invariance in each local graph;

outside the sub-units, the output of each sub-unit is also passed to each subsequent sub-unit as its input through a jump connection operation, which can be written as:

wherein F_iRepresenting the output characteristic of the ith sub-unit, E_i(. represents the ith subunit, F_inThe characteristics of the input are represented by,

indicating a connect operation.

5. The point cloud completion method based on dynamic graph volume and attention mechanism as claimed in claim 1, wherein the feature aggregation method in step S2 is: extracting global feature by combining attention pooling and maximum pooling, and combining with point-by-point local feature F_pPolymerization to form F_a。

6. The point cloud completion method based on dynamic graph convolution and attention mechanism as claimed in claim 5, wherein said features are aggregated into point-by-point features F_pAnd global feature aggregation is used as the input of the next stage, and the original features of the input point cloud are reserved while the shapes of fine granularity are generated.

7. The point cloud completion method based on dynamic graph convolution and attention mechanism as claimed in claim 5, wherein said 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,

representing the input features of the ith point;

then multiplying each element by the corresponding fraction and adding to obtain the aggregation characteristic f_outExpressed as:

wherein f is_outRepresents the output characteristic, h_θ(. -) represents the shared multi-layer perceptron layer.

8. The point cloud completion method based on dynamic graph convolution and attention mechanism as claimed in claim 1, wherein the feature completion and reconstruction method in step S3 is: and expanding and refining the features by using a feature completion module, and reconstructing the coordinates of the complete point cloud.

9. The point cloud completion method based on dynamic graph convolution and attention mechanism as claimed in claim 8, wherein said feature completion module is a hierarchical residual structure.

10. The point cloud completion method based on dynamic graph convolution and attention mechanism as claimed in claim 8, wherein the feature completion and reconstruction method in step S3 is as follows:

s101, firstly, obtaining a feature matrix L with the size of NxC through a shared multilayer perceptron₀C represents a characteristic dimension, and N represents the number;

s102, copying the transformed feature matrix by r times to obtain a new feature map;

s103, generating r different two-dimensional grids comprising m grid points, expanding the grid points into the size of the repeated features, and splicing the coordinates of the grid points with the repeated features;

s104, obtaining the size rN through the self-attention mechanism unit and the shared multilayer perceptron₃Expanded feature matrix H of xC₀；

S105, pair characteristic H₀Performing reshaping to recover to L₀Feature matrix L with same dimension₁To L for₁And L₀Making a difference to obtain a residual error delta L;

s106, obtaining the expansion residual error feature H through the steps S102-S104 for delta L₁Finally, H is finally converted to₁And H₀Adding to obtain the expansion characteristic H_out；

S107, the final pass parameter is [ C, 128, 64, 3 ]]Get the size of rN₃X 3 complete point cloud.

Images