A high-precision point cloud completion method and device based on deep learning

Abstract

The application discloses a kind of high-precision point cloud completion method and device based on deep learning, comprising: in feature extraction module, introduce dynamic kernel convolution PAConv, can learn weight coefficient according to the position relationship of each point and its neighborhood point, and combine weight matrix to adaptively construct convolution kernel.In feature fusion module, spatial attention mechanism is added, which helps the decoder to better learn the mutual relationship between multiple features, so as to better represent these feature information.The discriminator module includes global and local attention discriminator module, which uses multi-layer full connection for classification, respectively from the whole and the local to judge whether the generated result conforms to the real point cloud distribution, and then optimize the generated result, thereby improving the accuracy of point cloud completion, obtaining complete and accurate point cloud completion result, which also provides guarantee for the smooth progress of point cloud segmentation, classification, object recognition and point cloud reconstruction and many other downstream tasks.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer three-dimensional point cloud completion and deep learning, and particularly relates to a high-precision point cloud completion method and device based on deep learning. BACKGROUND

[0002] In 3D computer vision applications, due to occlusion and limited sensor resolution, the original point cloud captured by 3D scanners and depth cameras is usually sparse and incomplete. The defects in the shape of the point cloud model greatly limit the perception ability of vision and AI. Point cloud completion is to complete a defective point cloud through an algorithm model, which is a basic technology in the field of 3D vision. At the same time, it is also a necessary step to obtain a complete point cloud model of a 3D object, and is the basis for subsequent related work.

[0003] Existing deep learning-based point cloud completion methods can infer relatively complete and reasonable point cloud models, but often have defects in completing local detail features.

[0004] Therefore, making up for the lack of deep learning-based point cloud completion methods in local feature extraction will be beneficial to representing the complex spatial change relationship of point clouds to improve the accuracy of point cloud completion. SUMMARY

[0005] In order to solve the problems in the prior art and solve the problem of insufficient local feature extraction in the current point cloud completion method, the present application provides a high-precision point cloud completion method based on deep learning. In the feature extraction module, dynamic kernel convolution PAConv is introduced, which can learn weight coefficients according to the positional relationship of each point and its neighborhood points, and combine the weight matrix to adaptively construct the convolution kernel. The spatial attention mechanism is added in the feature fusion module, which helps the decoder to better learn the mutual relationship between multiple features, so as to better represent these feature information. The discriminator module includes global and local attention discriminator modules, which use multiple layers of full connection for classification, respectively from the whole and the local to judge whether the generated result conforms to the real point cloud distribution, and then optimize the generated result. Thus, the accuracy of point cloud completion is improved, and complete and accurate point cloud completion results are obtained, which also provides a guarantee for the smooth progress of point cloud segmentation, classification, object recognition and point cloud reconstruction and many other downstream tasks.

[0006] Technical scheme: In order to solve the above technical problems, the technical scheme adopted by the present application is:

[0007] In the first aspect, a high-precision point cloud completion method based on deep learning is provided, comprising:

[0008] obtaining point cloud data to be processed;

[0009] Preprocess the point cloud data to obtain preprocessed point cloud data;

[0010] input the preprocessed point cloud data into the trained point cloud completion model, wherein the point cloud completion model comprises a multi-resolution encoder module, a pyramid decoder module and an attention discriminator module;

[0011] The multi-resolution encoder module is configured to extract and fuse features of the input point cloud data to obtain a feature vector;

[0012] The pyramid decoder module is configured to process the feature vector to obtain point cloud completion results of three scales;

[0013] The attention discriminator module is configured to use the idea of generative adversarial network to generate results of global and local feature consistency through mutual game learning of generative model and discriminative model;

[0014] According to the output of the point cloud completion model, a high-precision point cloud completion result is determined.

[0015] In some embodiments, the multi-resolution encoder module comprises a feature extraction module and a feature fusion module,

[0016] In the feature extraction module, a dynamic convolution layer PAConv is embedded in a multi-layer perception MLP with shared weights, the weight coefficients are learned according to the positional relationship of each point and its neighborhood points, and the weight matrix is combined to adaptively construct a convolution kernel, thereby improving the local detail feature extraction capability;

[0017] A spatial attention mechanism is added in the feature fusion module to realize feature focusing in the spatial dimension;

[0018] The farthest point sampling generates three different scales of missing point clouds input into the multi-resolution encoder module;

[0019] The feature extraction module of the multi-layer perception MLP embedded with the dynamic kernel convolution PAConv extracts features of the three different scales of missing point clouds to generate multi-dimensional feature vectors V1, V2 and V3; the output multi-dimensional feature vectors V1, V2 and V3 are input into the feature fusion module composed of spatial attention mechanisms, the spatial attention mechanisms learn 1024-dimensional abstract features that integrate local features and global information, and output the features of each position after weighting; then, the three 1024-dimensional abstract features are spliced by a splicing array, and finally the latent features are mapped and integrated into the final feature vector V with a dimension of 1024 by using MLP.

[0020] Further, the construction method of the dynamic kernel convolution PAConv comprises:

[0021] Initialize a K-size Cin X C out weight matrix W = {W k | k = 1, 2, …, K}, where C in represents the input dimension of the network at the current layer, C out represents the output dimension of the network at the current layer.

[0022] Calculate the relative position relationship of each point p i in the input point cloud and the neighborhood point p j , learn the weight coefficient of different positions, and represent as:

[0023] E ij = Softmax (θ (p i , p j ))

[0024] Where θ is a nonlinear function implemented by a 1x1 convolution kernel; the normalization operation using the Softmax function ensures that the output score is in the range (0, 1); a higher score means that the corresponding position has more important local information.

[0025] The kernel of PAConv is composed of the weight matrix W k and the weight coefficient learned from the point position,

[0026]

[0027] So far, the dynamic kernel convolution PAConv has completed the work of adaptively constructing the convolution kernel, which is used to capture the information of the local region of the input feature and output the feature with local correlation.

[0028] Preferably, K is 16.

[0029] In some embodiments, the feature vector is processed to obtain a point cloud completion result of three scales, including: the feature vector V passes through a fully connected layer to obtain three sub-feature vectors U1, U2 and U3 of different resolutions, each sub-feature vector is responsible for completing a point cloud of different resolution; the relative coordinates of the primary point cloud P3 from the center point of P3 are predicted using U3, the relative coordinates of the secondary point cloud P2 from the center point of P3 are predicted using U2, and the secondary point cloud P2 is generated according to P3 using reorganization and full connection operations; the relative coordinates of the final point cloud P1 from the center point of P2 are predicted using U1 and P2, and the final point cloud P1 is supplemented in this way.

[0030] In some embodiments, the attention discriminator module includes a global attention discriminator and a local attention discriminator; the global discriminator is used to view the entire point cloud completion result to evaluate its overall consistency, and the local discriminator module views a small area centered on the completed area to ensure the local consistency of the generated point cloud.

[0031] In some embodiments, the processing process of the attention discriminator module includes: inputting the overall or local generated point cloud and the real point cloud into the attention discriminator, obtaining a feature vector with a dimension of 512 through the autoencoder therein, and then reducing the dimension through a continuous fully connected layer [512-256-128-16-1] to output a final binary result of fake or real.

[0032] In some embodiments, the training method of the point cloud completion model comprises:

[0033] The loss function includes two parts: a generation loss and an adversarial loss;

[0034] The average shortest point distance between the generated point cloud and the ground real point cloud is calculated by using the chamfer distance CD, and the calculation formula is:

[0035]

[0036] Wherein, x, y represent a certain point in the generated point cloud or the real point cloud; ||*|| represents the distance; CD calculates the average nearest square distance between the generated point cloud S1 and the real point cloud S2, and the final generated result is three generated point clouds P1, P2 and P3 of different scales, and the generation loss is also composed of three parts, respectively corresponding to the CD values of the three generated point clouds of different scales, wherein a represents the summation weight in the generation loss;

[0037] The generation loss L com is:

[0038]

[0039] Wherein, P 1gt , P 2gt , P 3gt are the real point clouds corresponding to the three generated point clouds of different scales, respectively.

[0040] The adversarial loss is inspired by the generative adversarial network (GAN), and the adversarial loss L adv is:

[0041] L adv =∑ 1≤i≤S log 10 (G(y i ))+∑ 1≤j≤S log 10(1-G(E(D(x i ))

[0042] where y i and x i are the original incomplete point cloud and the real point cloud respectively, S represents the size of the dataset; E, D and G represent the multi-resolution encoder, the pyramid decoder and the attention discriminator respectively;

[0043] The total loss function L is composed of the generation loss and the adversarial loss:

[0044] L=βL com +λL adv

[0045] β and λ are the weights of the generation loss L com and the adversarial loss L adv , and satisfy the following conditions: β+λ=1; the chamfer distance CD is also used as an evaluation index for testing the completion performance.

[0046] In the second aspect, the application provides a high-precision point cloud completion device based on deep learning, comprising a processor and a storage medium;

[0047] The storage medium is used for storing instructions;

[0048] The processor is used for operating according to the instructions to execute the steps of the method according to the first aspect.

[0049] In the third aspect, the application provides a storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to realize the steps of the method according to the first aspect.

[0050] The application has the following technical effects compared with the prior art: (1) the application introduces a convolution layer PAConv in the feature extraction module, which can learn weight coefficients according to the positional relationship between each point and its neighborhood points, and combine the weight matrix to adaptively construct a convolution kernel, so that the information of the local region can be flexibly captured.

[0051] (2) the application adds a spatial attention mechanism in the feature fusion module, so that the decoder can better learn the mutual relationship between various features and improve the accuracy of point cloud completion.

[0052] (3) the application includes global and local discriminator modules in the discriminator module, which uses multi-layer full connection for classification, respectively judges whether the generated result conforms to the real point cloud distribution from the whole and the local, and then optimizes the generated result.

[0053] The advantages of the present application are that the deficiencies of the point cloud completion method based on deep learning in local feature extraction are made up, the accuracy of point cloud completion is improved by introducing the PAConv convolution layer and the spatial attention mechanism, and more complete and accurate point cloud completion results are obtained, and the global and local attention discriminator modules are introduced to ensure the overall and local consistency between the generated point cloud and the real point cloud. At the same time, the point cloud completion also provides guarantee for the smooth progress of many downstream tasks such as point cloud segmentation, classification, object recognition and point cloud reconstruction. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 It is a schematic diagram of the overall network framework in the embodiment of the present application.

[0055] Figure 2 It is a schematic diagram of the spatial attention mechanism in the embodiment of the present application.

[0056] Figure 3 It is a schematic diagram of the PAConv structure in the embodiment of the present application.

[0057] Figure 4 It is a schematic diagram of the missing point cloud completion process in the embodiment of the present application. DETAILED DESCRIPTION

[0058] In order to make the technical means, creative features, purposes and effects realized by the present application easy to understand, the present application is further described below in combination with specific embodiments.

[0059] In the description of the present application, the meaning of several is more than one, the meaning of multiple is more than two, greater than, less than, more than, etc. are understood as not including the number, above, below, etc. are understood as including the number. If it is described as first, second, it is only used for the purpose of distinguishing technical features, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features or the sequence of indicated technical features.

[0060] In the description of the present application, the description of the terms "one embodiment", "some embodiments", "exemplary embodiment", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in combination with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the exemplary description of the above terms does not necessarily mean the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0061] Embodiment 1

[0062] A high-precision point cloud completion method based on deep learning, comprising:

[0063] obtain point cloud data to be processed;

[0064] pre-process the point cloud data to obtain pre-processed point cloud data;

[0065] input the pre-processed point cloud data into a trained point cloud completion model, wherein the point cloud completion model comprises a multi-resolution encoder module, a pyramid decoder module and an attention discriminator module;

[0066] The multi-resolution encoder module is configured to perform feature extraction and fusion on the input point cloud data to obtain a feature vector;

[0067] The pyramid decoder module is configured to process the feature vector to obtain point cloud completion results of three scales;

[0068] The attention discriminator module is configured to use the idea of a generative adversarial network to generate results with global and local feature consistency through mutual game learning between a generative model and a discriminative model;

[0069] According to the output of the point cloud completion model, a high-precision point cloud completion result is determined.

[0070] In some embodiments, a high-precision point cloud completion method based on deep learning comprises Figure 1 As shown in the figure, the overall framework of the point cloud completion method comprises a multi-resolution encoder, a pyramid decoder and an attention discriminator. The multi-resolution encoder extracts the features of the input point cloud; the pyramid decoder processes the fused feature vector to obtain point cloud completion results of three scales; the attention discriminator calculates the adversarial loss, and through mutual game learning, it produces good output, ensuring the overall and local consistency between the generated point cloud and the real point cloud. The implementation method and function of each module are described in detail as follows:

[0071] First, the farthest point sampling is used to select the farthest point from the existing sampling point set iteratively to obtain a set of skeleton points. This can more uniformly represent the distribution of the point set and will not damage the structure of the point cloud model. The farthest point sampling generates three different scale missing point clouds, which are input into the multi-resolution encoder for feature extraction. A multi-layer perceptron using embedded dynamic kernel convolution PAConv is used to generate multi-dimensional feature vectors V1, V2, V3. The output multi-dimensional feature vectors V1, V2, V3 are input into a feature fusion module composed of a spatial attention mechanism, and the structure is as shown in Figure 2As shown, the spatial attention mechanism learns 1024-dimensional abstract features that integrate local features and global information, realizes feature focusing in the spatial dimension, and outputs the weighted features of each position. Then, the three 1024-dimensional abstract features are spliced together, and finally, the latent features are mapped to the final feature vector V with a dimension of 1024 using MLP.

[0072] wherein the PAConv part first initializes a weight library W = {W in | k = 1, 2,..., K} composed of K weight matrices with a size of C out × C k , wherein C in represents the input dimension of the network at the current layer, and C out represents the output dimension of the network at the current layer. A larger K can ensure the diversity of the convolution kernel, but it will also increase the burden of the model. Therefore, in our network model, we take K = 16. Next, the relative position relationship of each point p i in the input point cloud and its neighborhood points p j is calculated, and the weight coefficients of different positions are learned.

[0073] E ij = Softmax(θ(p i , p j )) # (1)

[0074] wherein θ is a nonlinear function realized by a 1 × 1 convolution. Normalization operation using Softmax can ensure that the output score is in the range (0, 1). A higher score means that the corresponding position has more important local information. The kernel of PAConv is composed of the weight matrix W k and the weight coefficient learned from the position of the point,

[0075]

[0076] At this point, PAConv has completed the work of adaptively constructing the convolution kernel. It can flexibly capture the information of the local region of the input features and output features with local correlation.

[0077] The pyramid decoder module is composed of a full connection layer and a recombination layer, adopts the feature pyramid network thought, gradually completes the missing point cloud from coarse to fine, the input is the output feature vector V of the multi-resolution encoder, three sub-feature vectors U1, U2 and U3 with dimensions of 1024, 512 and 256 are obtained through the full connection layer, and each sub-feature vector is responsible for completing the point cloud of different resolutions. Firstly, U3 is used to predict the primary point cloud P3, then U2 is used to predict the relative coordinates of the secondary point cloud P2 from the center point of P3, and then the recombination and full connection operation are used to generate the secondary point cloud P2 according to P3. In the same way, U1 and P2 can be used to predict the relative coordinates of the final point cloud P1 from the center point of P2, so that the final point cloud P1 can be supplemented.

[0078] The attention discriminator module uses the idea of a generative adversarial network, and the generated model and the discriminant model in the framework learn from each other to produce good output. The module is composed of a global attention discriminator and a global attention discriminator, the global discriminator views the entire point cloud completion result to evaluate its overall consistency, and the local discriminator module only views a small area centered on the completed area to ensure the local consistency of the generated point cloud. The specific method is as follows: the generated point cloud and the real point cloud are sent into the discriminator, the feature vector with a dimension of 512 is obtained through the autoencoder therein, and then the dimension is reduced through a continuous full connection layer [512-256-128-16-1], and the final binary result of fake or real is output.

[0079] The loss function of the algorithm includes two parts: generation loss and adversarial loss.

[0080] The average shortest point distance between the generated point cloud and the ground real point cloud is calculated by using the chamfer distance CD, and the calculation formula is:

[0081]

[0082] In formula (3), CD calculates the average nearest square distance between the generated point cloud S1 and the real point cloud S2, since the final generated result is three generated point clouds P1, P2 and P3 of different scales, the total loss is also composed of three parts, respectively corresponding to the CD values of the three generated point clouds of different scales, wherein alpha represents the summation weight in the generation loss. The total loss expression is:

[0083]

[0084] In formula (4), P 1gt , P 2gt and P 3gt are the real point clouds corresponding to the three generated point clouds of different scales. The adversarial loss in this paper is inspired by the adversarial network GAN, and the calculation formula is as follows:

[0085] L adv =∑ 1≤i≤S log 10 (G(y i ))+∑ 1≤j≤S log 10 (1-G(E(D(x i ))))#(5)

[0086] In formula (5), y i and x i belong to the original incomplete point cloud and the real point cloud respectively, E, D and G represent the multi-resolution encoder, the pyramid decoder and the attention discriminator respectively, the total loss is composed of the generation loss and the adversarial loss, and the calculation formula is shown in formula (6):

[0087] L=βL com +λL adv #(6)

[0088] β and λ are the weights of the generation loss L com and the adversarial loss L adv , and satisfy the following conditions: β+λ=1, and the chamfer distance CD is also used as an evaluation index for testing the completion performance.

[0089] The system provided by the application has the following advantages:

[0090] (1) A feasible scheme is provided to make up for the deficiency of the point cloud completion method based on deep learning in local feature extraction.

[0091] (2) A point cloud model with high completion accuracy can be obtained, which provides a guarantee for the smooth progress of many downstream tasks such as point cloud segmentation, classification, object recognition and point cloud reconstruction.

[0092] The point cloud completion method based on deep learning disclosed in the application can extract global and local features of point cloud and integrate local correlation and global information of key points, make up for the deficiency of the point cloud completion method based on deep learning in local feature extraction, improve the accuracy of point cloud completion, and also provide a guarantee for the smooth progress of many downstream tasks such as point cloud segmentation, classification, object recognition and point cloud reconstruction.

[0093] Embodiment 2

[0094] In a second aspect, the embodiment provides a high-precision point cloud completion device based on deep learning, comprising a processor and a storage medium.

[0095] The storage medium is used to store instructions.

[0096] The processor is used to operate according to the instructions to execute the steps of the method according to embodiment 1.

[0097] Embodiment 3

[0098] In a third aspect, the present embodiment provides a storage medium having stored thereon a computer program, wherein the computer program is executed by a processor to implement the steps of the method in the first embodiment.

[0099] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) containing computer usable program code.

[0100] The present application is described with reference to the flowchart and / or block diagram of the method, apparatus (system) and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as a combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions, which are executed via the processor of the computer or other programmable data processing apparatus, generate a means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more flows and / or blocks Figure 1 an apparatus for implementing functions specified in one or more flows and / or blocks.

[0101] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction apparatus, which implements the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more flows and / or blocks Figure 1 an apparatus for implementing functions specified in one or more flows and / or blocks.

[0102] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are performed on the computer or other programmable data processing apparatus to produce a computer-implemented process, so that the instructions executed on the computer or other programmable data processing apparatus provide a process for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more flows and / or blocks Figure 1 an apparatus for implementing functions specified in one or more flows and / or blocks.

[0103] It is apparent that the application can be carried out by other embodiments that do not depart from the spirit or essential characteristics thereof. Consequently, the above disclosure is considered as illustrative only and not restrictive on the application. As various changes could be made in the above compositions without departing from the scope of the application, it is intended that all changes fall within the scope of the application as claimed in the appended claims.

Claims

1. A high-precision point cloud completion method based on deep learning, characterized in that, The method comprises the following steps: acquiring point cloud data to be processed; preprocessing the point cloud data to obtain preprocessed point cloud data; inputting the preprocessed point cloud data into a trained point cloud completion model, wherein the point cloud completion model comprises a multi-resolution encoder module, a pyramid decoder module and an attention discriminator module; the multi-resolution encoder module is configured to perform feature extraction and fusion on the input point cloud data to obtain a feature vector; the pyramid decoder module is configured to process the feature vector to obtain point cloud completion results of three scales; the attention discriminator module is configured to use the idea of a generative adversarial network to generate results with global and local feature consistency through mutual game learning of a generative model and a discriminative model; determining a high-precision point cloud completion result according to the output of the point cloud completion model; wherein the multi-resolution encoder module comprises a feature extraction module and a feature fusion module; in the feature extraction module, a dynamic convolution layer PAConv is embedded in a multi-layer perception machine MLP with shared weights; a weight coefficient is learned according to the positional relationship of each point and its neighborhood points, and a convolution kernel is adaptively constructed by combining a weight matrix to improve the local detail feature extraction capability; a spatial attention mechanism is added in the feature fusion module to realize feature focusing in the spatial dimension; the farthest point sampling generates missing point clouds of three different scales which are input into the multi-resolution encoder module; the feature extraction module of the multi-layer perception machine embedded with the dynamic kernel convolution PAConv extracts features of the missing point clouds of the three different scales to generate multi-dimensional feature vectors V1, V2 and V3; the output multi-dimensional feature vectors V1, V2 and V3 are input into the feature fusion module composed of the spatial attention mechanism, the spatial attention mechanism learns 1024-dimensional abstract features that comprehensively integrate local features and global information, and outputs the features of each position after weighting; the three 1024-dimensional abstract features are spliced into an array, and finally the latent features are mapped and integrated into the final feature vector V with a dimension of 1024 by using the MLP; the construction method of the dynamic kernel convolution PAConv comprises: Initialize a weight library W = {W in |k = 1, 2, …, K} composed of K weight matrices of size C out × C k , where C in represents the input dimension of the network at the current layer, and C out represents the output dimension of the network at the current layer; calculate the relative position relationship of each point p i in the input point cloud and the neighborhood point p j , and learn the weight coefficients of different positions , which is represented as: E ij = Softmax(θ(p i , p j )) where θ is a nonlinear function implemented by a convolution with a kernel size of 1x1; the normalization operation using the Softmax function ensures that the output scores are in the range (0, 1); a higher score means that the corresponding position has more important local information; the kernel of PAConv is composed of a weight matrix W k and the weight coefficients learned from the point positions , At this point, the dynamic kernel convolution PAConv has completed the work of adaptively constructing a convolution kernel for capturing information in the local region of the input features and outputting features with local correlation; the attention discriminator module comprises a global attention discriminator and a local attention discriminator; the global attention discriminator is used to view the entire point cloud completion result to evaluate its overall consistency, and the local attention discriminator is used to view a small area centered on a completed area to ensure the local consistency of the generated point cloud; the processing process of the attention discriminator module comprises: inputting the overall or local generated point cloud and the real point cloud into the attention discriminator, obtaining a feature vector with a dimension of 512 through the autoencoder therein, and then reducing the dimension through a continuous fully connected layer [512-256-128-16-1] to output a final binary result of fake or real.

2. The deep learning-based high-precision point cloud completion method according to claim 1, characterized in that, The value of K is 16.

3. The deep learning-based high-precision point cloud completion method according to claim 1, characterized in that, The feature vector is processed to obtain a point cloud completion result of three scales, including: the feature vector V passes through a full connection layer to obtain three sub-feature vectors U1, U2 and U3 of different resolutions, each of which is responsible for completing a point cloud of different resolution; a primary point cloud P3 is predicted by using U3, a secondary point cloud P2 is predicted by using U2 to obtain the relative coordinates of the secondary point cloud P2 from the center point of P3, and the secondary point cloud P2 is generated according to P3 by using recombination and full connection operation; and a final point cloud P1 is predicted by using U1 and P2 to obtain the relative coordinates of the final point cloud P1 from the center point of P2, so as to supplement the final point cloud P1.

4. The deep learning-based high-precision point cloud completion method according to claim 1, characterized in that, The training method of the point cloud completion model comprises: The loss function comprises two parts: a generation loss and an adversarial loss; The average shortest point distance between the generated point cloud and the ground truth point cloud is calculated by using the chamfer distance CD, and the calculation formula is: Wherein, x, y represent a certain point in the generated point cloud or the real point cloud; ||*|| represents the distance; CD calculates the average nearest square distance between the generated point cloud S1 and the real point cloud S2, and the final generated result is three generated point clouds P1, P2 and P3 of different scales, and the generation loss is also composed of three parts, d CD3 The CD values of the three generated point clouds of different scales respectively, wherein a represents the summation weight in the generation loss; Generation loss L com The expression is: wherein P 1gt , P 2gt , and P 3gt are respectively the real point clouds corresponding to the generated point clouds of three different scales; the adversarial loss L adv is: L adv =∑ 1≤i≤S log 10 (G(y i ))+∑ 1≤j≤S log 10 (1-G(E(D(x i )))) where y i and x i belong to the original incomplete point cloud and the real point cloud, respectively, S represents the dataset size; E, D, and G represent the multi-resolution encoder, the pyramid decoder, and the attention discriminator, respectively; The total loss function L is composed of the generation loss and the adversarial loss: L = βL com + λL adv , β and λ are weights for the generation loss L com and the adversarial loss L adv respectively, satisfying the following conditions: β + λ = 1; the chamfer distance CD is also used as an evaluation metric for testing the completion performance.

5. A high-precision point cloud completion device based on deep learning, characterized in that, It comprises a processor and a storage medium; The storage medium is used for storing instructions; The processor is used for operating according to the instructions to execute the steps of the method according to any one of claims 1 to 4.

6. A storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to realize the steps of the method according to any one of claims 1 to 4.

Images