Generative adversarial network-based point cloud reconstruction method

Abstract

The invention discloses a generative adversarial network-based point cloud reconstruction method, and belongs to the technical fields of computer three-dimensional reconstruction, computer vision, big data cloud computing and the like. Aiming at the problems of how to improve the point cloud reconstruction learning capability of a network architecture and how to reduce the training complexity of the network architecture in the existing point cloud reconstruction method, the method adopts a generative adversarial network as a main body framework, and designs a generator to map a point cloud from a feature space to a signed distance field; and constructing a discriminator to optimize a global gradient and a spatial point signed distance prediction value so as to improve the detail precision of reconstruction. The generator directly generates an initial shape from the point cloud data, and the discriminator optimizes the implicit field and gradient predicted by the generator to improve the quality of the final shape. According to the method, the generative adversarial network framework is improved, the point cloud reconstruction capability of the neural network model is enhanced, and the neural network model can more effectively reconstruct a correct and fine topological curved surface from the point cloud data.

Description

Technical Field

[0001] The present invention belongs to the technical fields of computer three-dimensional reconstruction, computer vision and big data cloud computing, and specifically relates to a point cloud reconstruction method based on a generative adversarial network. Background Art

[0002] Point cloud is a set of discrete 3D points representing the geometric shape of an object surface. It can describe the geometric surface structure of an object and flexibly describe complex geometric structures, especially those irregular objects without clear topological relationships. However, since it is a discrete set, point cloud has certain limitations in representing details and topological structures. In the fields of computer vision, big data cloud computing, etc., high-quality 3D shape generation is an important task.

[0003] At present, the implicit function representation of three-dimensional shapes has received more and more attention, especially the implicit representation based on neural networks has become a very important method. Point cloud reconstruction refers to combining discrete points in the point cloud into a continuous three-dimensional surface, usually using meshing or implicit surface fitting to generate a three-dimensional model. Through point cloud reconstruction, discrete points can be converted into a continuous triangular mesh surface or voxel representation, which enables the point cloud to be used in subsequent rendering, analysis, physical simulation and manufacturing tasks. Point cloud reconstruction methods usually include the following: Poisson surface reconstruction, AlphaShapes, Delaunay partitioning algorithm and implicit function reconstruction method. Poisson surface reconstruction is a point cloud reconstruction method based on the mathematical Poisson equation, which is widely used to generate smooth and complete three-dimensional surfaces from three-dimensional point clouds. Poisson surface reconstruction can handle noisy point cloud data well and has high computational efficiency; Alpha Shapes is a geometric method for generating 3D surface models from point cloud data. After determining a parameter, the Alpha Shapes method selects simple shapes with specific properties from the Delaunay triangulation to form the surface of the object; Delaunay triangulation algorithm, Delaunay triangulation is based on the empty circle rule, that is, by maximizing the minimum angle of the triangle to avoid generating thin and long triangles, so that the reconstructed surface has better quality and stability; Implicit function reconstruction method, by constructing an implicit function, the point cloud is represented as the zero isosurface of the function. This method can handle complex shapes and topological structures and is suitable for the reconstruction of smooth and continuous surfaces.

[0004] Point cloud reconstruction has important application value in many fields, including computer graphics, reverse engineering, robot navigation, autonomous driving, and cultural relics protection. However, among the many methods of point cloud reconstruction, the reconstructed surface accuracy is insufficient or the training complexity is too high. Therefore, how to improve the ability of network architecture to learn point cloud reconstruction and reduce the training complexity of network architecture is an important problem to be solved in this field. Summary of the invention

[0005] In order to solve the problem of how to improve the ability of network architecture to learn point cloud reconstruction and reduce the training complexity of network architecture in current point cloud reconstruction methods, the present invention provides a point cloud reconstruction method based on generative adversarial network.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A point cloud reconstruction method based on a generative adversarial network comprises the following steps:

[0008] Step 1, data processing of ShapeNet dataset;

[0009] Step 1.1, import the os module, torch module, and numpy module for data processing operations;

[0010] Step 1.2, initialize the point cloud scaling factor and set the total number of sampling points;

[0011] Step 1.3, set the number of random sampling points, set the number of signed distance sampling points, extract the position and normal vector of the random sampling points, and calculate the signed distance value, gradient and normal vector of the signed distance sampling points;

[0012] Step 1.4, create a new point set from the data of the randomly sampled points, and convert the new point set into an octree structure;

[0013] Step 1.5, saving the original point cloud data, the octree point cloud data, and the signed distance sampling point data as a data set;

[0014] Step 2: construct a generative adversarial network framework, adjust the hyperparameters and weights of the neural network model, and obtain a trained neural network model;

[0015] Step 2.1, set the voxel size of the generator and discriminator, as well as the latent vector dimension;

[0016] Step 2.2, construct the generator, design the loss function of the generator, build the vector module of the generator, the convolutional network module of the generator and the generation module of the generator, wherein the vector module of the generator is composed of a linear layer, the convolutional network module of the generator is composed of four generation groups and an initial convolutional layer, and the generation module of the generator is composed of a nine-layer fully connected network with a periodic activation function;

[0017] The vector module of the generator outputs a style vector, the convolutional network module of the generator outputs voxel features, and the generation module of the generator outputs a signed distance field;

[0018] Step 2.3, build the discriminator and design the loss function of the discriminator. The discriminator consists of five discriminant modules and a four-layer fully connected network. First, build the discriminant module. Each discriminant module includes a residual convolution layer and two convolution layers with a convolution kernel size of 3. The residual convolution layer is used as a shortcut connection, and the convolution kernel size is 1. Then build a fully connected network consisting of four fully connected layers. The discriminator outputs a scalar value to judge whether the voxel points generated by the generator are correct or wrong;

[0019] Step 2.4, input the training set obtained in step 1.5 into the generative adversarial network framework constructed in steps 2.2 and 2.3, the input latent vector is obtained by sampling from the standard Gaussian distribution, the hyperparameters and weights of the neural network model are adjusted, and the trained neural network model is obtained;

[0020] Step 3, input the test set obtained in step 1.5 into the neural network model trained in step 2.4, and use the marching cube algorithm to reconstruct the shape of the point cloud data;

[0021] Step 3.1, set the grid file save path and import the trained neural network model;

[0022] Step 3.2, set 150,000 random latent vectors that follow a standard Gaussian distribution and calculate the mean of the latent space;

[0023] Step 3.3, set the truncation parameter to the mean of the latent space, so that the mean of the latent space where the style vector generated by the generator is located gradually approaches the truncation parameter;

[0024] Step 3.4, using the marching cubes algorithm to extract the signed distance values ​​output by the trained neural network model as isosurfaces, and reconstruct the three-dimensional shape;

[0025] In step 3.5, the reconstructed three-dimensional shape is subjected to a spherical normalization operation and saved in the form of an .obj file in the mesh file saving path set in step 3.1.

[0026] Furthermore, the sampling range of the signed distance sampling points in step 1.3 includes the surface of the original point cloud and the interior of the original point cloud. The surface of the original point cloud adopts a random sampling method, and the interior of the original point cloud adopts a uniform sampling method. Finally, points are randomly selected from the signed distance sampling points until the number of points reaches 5000.

[0027] Furthermore, in step 1.2, the total number of sampling points is 9000; in step 1.3, the number of random sampling points is 3500, and the number of signed distance sampling points is 5500.

[0028] Furthermore, in step 1.5, the data set includes a training set and a test set, and the ratio of the training set to the test set is set to 7:3.

[0029] Furthermore, the dimension of the potential vector in step 2.1 is 512 dimensions.

[0030] Furthermore, in step 2.2, the loss function of the generator is designed specifically as follows: the training of the generator depends on the feedback of the discriminator, and the loss function of the generator is expressed as follows:

[0031] L Z =γB gh -E a～A [log(X F (Q F (M(a))))];

[0032] Among them, L Z represents the loss function of the generator; γ represents the weight of the control regularization term; B gh represents the regularization term of the generator; E represents the constructed discriminator; a~A represents the sampling of latent vectors from the latent space A; X F represents the fully connected network of the discriminator, M represents the generated SDF; M(a) represents the SDF generated by the latent vector a; Q F Represents the discriminant module of the discriminator, which acts on the generated SDF, i.e., M(a).

[0033] Furthermore, the vector module of the generator in step 2.2 is composed of linear layers, specifically: first, nine linear layers are built, and the input feature vector is transformed through a linear layer with an equal learning rate to generate a style vector for modulating the generator.

[0034] Furthermore, the convolutional network module of the generator in the step 2.2 consists of four generation groups and an initial convolutional layer, specifically: four generation groups and an initial convolutional layer with a channel number of 512 and a convolution kernel size of 3 are constructed, each generation group is composed of three convolutional layers, the input and output channel numbers of the first two convolutional layers are both 512, the convolution kernel size is 3, and the input channel number of the last convolutional layer is 512, the output channel number is 64, and the convolution kernel size is 1.

[0035] Furthermore, the generation module of the generator in step 2.2 is composed of a nine-layer fully connected network with a periodic activation function, specifically: it is composed of a nine-layer fully connected network, the activation function of each hidden layer is a periodic function, the dimensions of the first four hidden layers are 512 dimensions, the dimensions of the fifth and sixth hidden layers are 256 dimensions, and the dimensions of the last three hidden layers are 512 dimensions.

[0036] Furthermore, the loss function of the discriminator designed in step 2.3 is expressed as follows:

[0037]

[0038] in, represents the loss of the discriminator; E represents the constructed discriminator; a~A represents the sampling of latent vectors from the latent space A; X F represents the fully connected network of the discriminator; M represents the generated SDF; M(a) represents the SDF generated by the latent vector a; Q F represents the discriminant module of the discriminator, acting on the generated SDF, i.e., M(a); M~preal represents the real SDF sampled from the real data distribution; η F represents the weight of the regularization term controlling the discriminator; W(M; φ F ) represents the regularization term of the discriminator, and the parameter φ F Represents the SDF generated by the constraints.

[0039] Compared with the prior art, the present invention has the following advantages:

[0040] The point cloud reconstruction method designed by the present invention can effectively reconstruct point cloud data into high-quality three-dimensional shapes. The present invention adopts a generative adversarial network as the main framework. Unlike the traditional point cloud reconstruction technology based on implicit function representation, which requires relatively complex data preprocessing steps for grid files and is different from the point cloud reconstruction technology based on geometric inference and topological connection, which lacks the flexibility of topological structure, the advantage is that the generator directly generates the initial shape from the point cloud data, and the discriminator optimizes the implicit field and gradient predicted by the generator to improve the quality of the final shape. The generative adversarial network framework adopted by the present invention designs the generator to map the point cloud from the feature space to the signed distance field, and constructs the discriminator to optimize the global gradient and the predicted value of the signed distance of the spatial point to improve the accuracy of the reconstructed details. The method of the present invention improves the generative adversarial network framework, strengthens the point cloud reconstruction capability of the neural network model, and enables the neural network model to more effectively reconstruct correct and fine topological surfaces from point cloud data. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0042] Figure 1 It is a network architecture diagram of a point cloud reconstruction method based on a generative adversarial network described in an embodiment of the present invention;

[0043] Figure 2 is a flow chart of a point cloud reconstruction method based on a generative adversarial network described in an embodiment of the present invention;

[0044] Figure 3 is a comparison diagram of a point cloud reconstruction method based on a generative adversarial network described in an embodiment of the present invention and different methods (IMGAN, ShapeGAN);

[0045] Figure 4 This is a rendering of a point cloud reconstruction method based on a generative adversarial network described in an embodiment of the present invention. DETAILED DESCRIPTION

[0046] In order to gain a deeper understanding of the present invention, we will provide a comprehensive and detailed description of the present invention. However, the present invention has multiple implementations and is not limited to the specific examples listed herein. The presentation of these examples is intended to deepen the comprehensive understanding of the disclosure of the present invention.

[0047] Example 1

[0048] like Figure 1 As shown in FIG. 1 , it is a network architecture diagram of a point cloud reconstruction method based on a generative adversarial network described in an embodiment of the present invention; Figure 2As shown, it is a flow chart of a point cloud reconstruction method based on a generative adversarial network described in an embodiment of the present invention; Figure 3 As shown in FIG. 1 , it is a comparison diagram of a point cloud reconstruction method based on a generative adversarial network described in an embodiment of the present invention and different methods (IMGAN, ShapeGAN), where GT represents the true value; Figure 4 , which is a rendering of a point cloud reconstruction method based on a generative adversarial network in an embodiment of the present invention; comprising the following steps:

[0049] Step 1: Perform data processing operations on the ShapeNet dataset, including the following sub-steps:

[0050] Step 1.1, import the os module, torch module and numpy module for data processing operations;

[0051] Step 1.2, initialize the point cloud scaling factor and set the total number of sampling points to 9000;

[0052] Step 1.3, set the number of random sampling points to 3500, set the number of signed distance sampling points to 5500, extract the position and normal vector of the random sampling points, calculate the signed distance value, gradient and normal vector of the signed distance sampling points, where the sampling range of the signed distance sampling points includes the surface of the original point cloud and the interior of the original point cloud, the surface of the original point cloud adopts the random sampling method, and the interior of the original point cloud adopts the uniform sampling method, and finally randomly select points from the signed distance sampling points until the number of points reaches 5000;

[0053] Step 1.4, create a new point set from the data of the randomly sampled points, and convert the created point set into an octree structure;

[0054] Step 1.5, save the original point cloud data, the octree point cloud data, and the signed distance sampling point data as a data set, the data set includes a training set and a test set, and the division ratio of the training set and the test set is set to 7:3;

[0055] Step 2: Build a generative adversarial network framework, adjust the hyperparameters and weights of the neural network model, and obtain a trained neural network model, including the following sub-steps:

[0056] Step 2.1, set the generator and discriminator voxel sizes, and set the potential vector dimension to 512 dimensions;

[0057] Step 2.2, construct the generator, design the loss function of the generator, build the vector module of the generator, the convolutional network module of the generator and the generation module of the generator, wherein the vector module of the generator is composed of a linear layer, the convolutional network module of the generator is composed of four generation groups and an initial convolutional layer, and the generation module of the generator is composed of a nine-layer fully connected network with a periodic activation function;

[0058] The vector module of the generator outputs a style vector, the convolutional network module of the generator outputs voxel features, and the generation module of the generator outputs a signed distance field;

[0059] The vector module of the generator is composed of linear layers. Specifically, nine linear layers are first built, and the input feature vector is transformed through linear layers with equal learning rates to generate a style vector for modulating the generator.

[0060] The convolutional network module of the generator consists of four generation groups and an initial convolutional layer. Specifically, four generation groups and an initial convolutional layer with a channel number of 512 and a convolution kernel size of 3 are constructed. Each generation group consists of three convolutional layers. The input and output channel numbers of the first two convolutional layers are both 512, and the convolution kernel size is 3. The input channel number of the last convolutional layer is 512, the output channel number is 64, and the convolution kernel size is 1.

[0061] The generation module of the generator consists of a nine-layer fully connected network with a periodic activation function. Specifically, it consists of a nine-layer fully connected network, the activation function of each hidden layer is a periodic function, the dimensions of the first four hidden layers are 512 dimensions, the dimensions of the fifth and sixth hidden layers are 256 dimensions, and the dimensions of the last three hidden layers are 512 dimensions.

[0062] The loss function of the generator is designed as follows: the training of the generator depends on the feedback of the discriminator. The loss function of the generator is expressed as follows:

[0063] L Z =γB gh -E a～A [log(X F (Q F (M(a))))];

[0064] Among them, L Z represents the loss function of the generator; γ represents the weight of the control regularization term; B gh represents the regularization term of the generator; E represents the constructed discriminator; a~A represents the sampling of latent vectors from the latent space A; X F represents the fully connected network of the discriminator, M represents the generated SDF; M(a) represents the SDF generated by the latent vector a; Q F Represents the discriminant module of the discriminator, which acts on the generated SDF, i.e., M(a).

[0065] Step 2.3, build the discriminator and design the loss function of the discriminator. The discriminator consists of five discriminant modules and a four-layer fully connected network. First, build the discriminant module. Each discriminant module includes a residual convolution layer and two convolution layers with a convolution kernel size of 3. The residual convolution layer is used as a shortcut connection, and the convolution kernel size is 1. Then build a fully connected network consisting of four fully connected layers. The discriminator outputs a scalar value to judge whether the voxel points generated by the generator are correct or wrong;

[0066] The loss function of the designed discriminator is expressed as follows:

[0067]

[0068] in, represents the loss of the discriminator; E represents the constructed discriminator; a~A represents the sampling of latent vectors from the latent space A; X F represents the fully connected network of the discriminator; M represents the generated SDF; M(a) represents the SDF generated by the latent vector a; Q F represents the discriminant module of the discriminator, acting on the generated SDF, i.e., M(a); M~preal represents the real SDF sampled from the real data distribution; η F represents the weight of the regularization term controlling the discriminator; W(M; φ F ) represents the regularization term of the discriminator, and the parameter φ F The SDF representing the constraint generation;

[0069] Step 2.4, input the training set obtained in step 1.5 into the generative adversarial network framework constructed in steps 2.2 and 2.3, the input latent vector is obtained by sampling from the standard Gaussian distribution, the hyperparameters and weights of the neural network model are adjusted, and the trained neural network model is obtained;

[0070] Step 3, input the test set obtained in step 1.5 into the neural network model trained in step 2.4, and use the marching cube algorithm to reconstruct the shape of the point cloud data;

[0071] Step 3.1, set the grid file save path and import the trained neural network model;

[0072] Step 3.2, set 150,000 random latent vectors that follow a standard Gaussian distribution and calculate the mean of the latent space;

[0073] Step 3.3, set the truncation parameter to the mean of the latent space, so that the mean of the latent space where the style vector generated by the generator is located gradually approaches the truncation parameter;

[0074] Step 3.4, using the marching cubes algorithm to extract the signed distance values ​​output by the trained neural network model as isosurfaces, and reconstruct the three-dimensional shape;

[0075] In step 3.5, the reconstructed three-dimensional shape is subjected to a spherical normalization operation and saved in the form of an .obj file in the mesh file saving path set in step 3.1.

[0076] The loss function for the generative adversarial network framework is expressed as follows:

[0077]

[0078] in, represents the loss function of the generative adversarial network framework, the goal is to minimize the difference between the generated shape and the true SDF value; U represents a set of points sampled from the volume space containing the input shape; q represents a sampled point in the volume space; G λ Represents the constructed generator, predicting the SDF value of a point under a given point cloud; represents the convolution module in the generator; y represents the sampling of the input shape; represents the convolutional module that maps the input shape y to the generator; SDF gt Represents the true SDF value, used to compare with the output of the generator.

[0079] The loss function for further optimizing the generative adversarial network framework is expressed as follows:

[0080]

[0081] in, represents the loss function for further optimizing the generative adversarial network framework, the goal is to make the generated SDF value as close to zero as possible to improve the surface approximation quality; Q represents a set of sampling points near the surface of the input shape; q represents a sampling point on the surface; G λ Represents the constructed generator, predicting the SDF value of a point under a given point cloud; The optimized parameters of the convolution module in the corresponding generator.

[0082] Example 2

[0083] The .obj file was tested using coverage (COV) as the test metric; 800 reconstructed instances were randomly selected from each category and evaluated using coverage (COV) as the test metric. The coverage reflects the degree of coverage of the shapes in the reference dataset by the generated dataset. The test results are shown in Table 1. In the comparison of the three categories (chair, table, and airplane), the method "Ours" of the present invention achieved the best results. Specifically, for the chair category, "Ours" performed best with a coverage of 71.6%. In the table category, it also stood out with the highest coverage of 80.9%. For the airplane category, "Ours" had the highest coverage of 73.8%. The performance comparison of the three methods (IMGAN, ShapeGAN, and Ours) highlights the competitiveness and advantages of the method "Ours" of the present invention in each category.

[0084] Table 1 COV (%)

[0085]

[0086] The contents not described in detail in the specification of the present invention belong to the prior art known to the professional and technical personnel in the field. Although the illustrative specific embodiments of the present invention are described above to facilitate the understanding of the present invention by the technical personnel in the field, it should be clear that the present invention is not limited to the scope of the specific embodiments. For the ordinary technical personnel in the field, as long as various changes are within the spirit and scope of the present invention defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A point cloud reconstruction method based on generative adversarial network, characterized in that: The following steps are involved: Step 1, data processing of ShapeNet dataset; Step 1.1, import the os module, torch module, and numpy module for data processing operations; Step 1.2, initialize the point cloud scaling factor and set the total number of sampling points; Step 1.3, set the number of random sampling points, set the number of signed distance sampling points, extract the position and normal vector of the random sampling points, and calculate the signed distance value, gradient and normal vector of the signed distance sampling points; Step 1.4, create a new point set from the data of the randomly sampled points, and convert the new point set into an octree structure; Step 1.5, saving the original point cloud data, the octree point cloud data, and the signed distance sampling point data as a data set; Step 2: construct a generative adversarial network framework, adjust the hyperparameters and weights of the neural network model, and obtain a trained neural network model; Step 2.1, set the voxel size of the generator and discriminator, as well as the latent vector dimension; Step 2.2, construct the generator, design the loss function of the generator, build the vector module of the generator, the convolutional network module of the generator and the generation module of the generator, wherein the vector module of the generator is composed of a linear layer, the convolutional network module of the generator is composed of four generation groups and an initial convolutional layer, and the generation module of the generator is composed of a nine-layer fully connected network with a periodic activation function; The vector module of the generator outputs a style vector, the convolutional network module of the generator outputs voxel features, and the generation module of the generator outputs a signed distance field; Step 2.3, build the discriminator and design the loss function of the discriminator. The discriminator consists of five discriminant modules and a four-layer fully connected network. First, build the discriminant module. Each discriminant module includes a residual convolution layer and two convolution layers with a convolution kernel size of 3. The residual convolution layer is used as a shortcut connection, and the convolution kernel size is 1. Then build a fully connected network consisting of four fully connected layers. The discriminator outputs a scalar value to judge whether the voxel points generated by the generator are correct or wrong; Step 2.4, input the training set obtained in step 1.5 into the generative adversarial network framework constructed in steps 2.2 and 2.3, the input latent vector is obtained by sampling from the standard Gaussian distribution, the hyperparameters and weights of the neural network model are adjusted, and the trained neural network model is obtained; Step 3, input the test set obtained in step 1.5 into the neural network model trained in step 2.4, and use the marching cube algorithm to reconstruct the shape of the point cloud data; Step 3.1, set the grid file save path and import the trained neural network model; Step 3.2, set 150,000 random latent vectors that follow a standard Gaussian distribution and calculate the mean of the latent space; Step 3.3, set the truncation parameter to the mean of the latent space, so that the mean of the latent space where the style vector generated by the generator is located gradually approaches the truncation parameter; Step 3.4, using the marching cubes algorithm to extract the signed distance values ​​output by the trained neural network model as isosurfaces, and reconstruct the three-dimensional shape; In step 3.5, the reconstructed three-dimensional shape is subjected to a spherical normalization operation and saved in the form of an .obj file in the mesh file saving path set in step 3.

1.

2. The point cloud reconstruction method based on generative adversarial network according to claim 1, characterized in that: The sampling range of the signed distance sampling points in step 1.3 includes the surface of the original point cloud and the interior of the original point cloud. The surface of the original point cloud adopts a random sampling method, and the interior of the original point cloud adopts a uniform sampling method. Finally, points are randomly selected from the signed distance sampling points until the number of points reaches 5000.

3. The point cloud reconstruction method based on generative adversarial network according to claim 1, characterized in that: In step 1.2, the total number of sampling points is 9000; in step 1.3, the number of random sampling points is 3500, and the number of signed distance sampling points is 5500.

4. The point cloud reconstruction method based on generative adversarial network according to claim 1, characterized in that: In step 1.5, the data set includes a training set and a test set, and the ratio of the training set to the test set is set to 7:

3.

5. The point cloud reconstruction method based on generative adversarial network according to claim 1, characterized in that: The dimension of the potential vector in step 2.1 is 512 dimensions.

6. The point cloud reconstruction method based on generative adversarial network according to claim 1, characterized in that: In step 2.2, the loss function of the generator is designed specifically that the training of the generator depends on the feedback of the discriminator, and the loss function of the generator is expressed as follows: L Z =γB gh -E a-A [log(X F (Q F (M(a))))]; Among them, L Z represents the loss function of the generator; γ represents the weight of the control regularization term; B gh represents the regularization term of the generator; E represents the constructed discriminator; a~A represents the sampling of latent vectors from the latent space A; X F represents the fully connected network of the discriminator, M represents the generated SDF; M(a) represents the SDF generated by the latent vector a; Q F Represents the discriminant module of the discriminator, which acts on the generated SDF, i.e., M(a).

7. The point cloud reconstruction method based on generative adversarial network according to claim 1, characterized in that: The vector module of the generator in step 2.2 is composed of linear layers, specifically: first build nine linear layers, transform the input feature vector through linear layers with equal learning rates, and generate a style vector for modulating the generator.

8. The point cloud reconstruction method based on generative adversarial network according to claim 1, characterized in that: The convolutional network module of the generator in the step 2.2 consists of four generation groups and an initial convolutional layer, specifically: construct four generation groups and an initial convolutional layer with a channel number of 512 and a convolution kernel size of 3, each generation group is composed of three convolutional layers, the input and output channel numbers of the first two convolutional layers are both 512, the convolution kernel size is 3, and the input channel number of the last convolutional layer is 512, the output channel number is 64, and the convolution kernel size is 1.

9. The point cloud reconstruction method based on generative adversarial network according to claim 1, characterized in that: The generation module of the generator in step 2.2 is composed of a nine-layer fully connected network with a periodic activation function, specifically: it is composed of a nine-layer fully connected network, the activation function of each hidden layer is a periodic function, the dimensions of the first four hidden layers are 512 dimensions, the dimensions of the fifth and sixth hidden layers are 256 dimensions, and the dimensions of the last three hidden layers are 512 dimensions.

10. The point cloud reconstruction method based on generative adversarial network according to claim 1, characterized in that: The loss function of the discriminator designed in step 2.3 is expressed as follows: in, represents the loss of the discriminator; E represents the constructed discriminator; a~A represents the sampling of latent vectors from the latent space A; X F represents the fully connected network of the discriminator; M represents the generated SDF; M(a) represents the SDF generated by the latent vector a; Q F represents the discriminant module of the discriminator, acting on the generated SDF, i.e., M(a); M~preal represents the real SDF sampled from the real data distribution; η F represents the weight of the regularization term controlling the discriminator; W(M; φ F ) represents the regularization term of the discriminator, and the parameter φ F Represents the SDF generated by the constraints.

Images