Encoder self-decoding feature reconstruction point cloud method and electronic equipment

Abstract

The invention relates to a feature reconstruction point cloud method for encoder self-decoding. The method comprises the following steps: S1, processing a three-dimensional point cloud data set into high-dimensional data features by using a special method comprising a pooling step; s2, the high-dimensional data features are split, explicit features and implicit features are obtained, the explicit features comprise length, width, height, size and other information, and the implicit features comprise high-dimensional pooling features; s3, randomly generating points within the range of the explicit feature constraint, wherein the generation form of the points has randomness; and S4, generating a new implicit feature of the random point by using the same method when parameterized feature description is carried out, and forming a set by points conforming to the original implicit feature so as to obtain a new point cloud conforming to the parameterized feature, namely, completing the reconstruction process from the feature to the point cloud. The method has the advantages that the point cloud can be simply and efficiently generated and reconstructed through parameterized features, and meanwhile, a decoding method not containing new information is provided, so that encoding and decoding work for the point cloud has relative consistency.

Description

Technical Field

[0001] The present invention relates to a feature reconstruction point cloud method for encoder self-decoding and electronic equipment, belonging to the field of aircraft design and computer-aided design (CAD). Background Art

[0002] With the rapid development of artificial intelligence and machine learning, the methods and approaches for predicting aircraft aerodynamic performance have also developed simultaneously. Predicting the aerodynamic performance of different aircraft shapes requires a parametric description of the shape, and also requires a method to reconstruct the parametric description back into a 3D point cloud and 3D shape.

[0003] Currently, mainstream methods include implicit function-based, deep learning-based, and voxel-based methods. Implicit function-based methods treat the point cloud surface as the zero-isosurface of an implicit function and reconstruct the point cloud by fitting the implicit function. Common algorithms include Poisson surface reconstruction and the MLS implicit function method. The disadvantages of the Poisson surface reconstruction method are its high computational complexity and long time consumption. The disadvantages of the MLS implicit function method are its high computational complexity, low efficiency, and sensitive parameter settings. Deep learning-based methods use deep neural networks to learn the mapping relationship between latent space features and point clouds to achieve reconstruction. Common methods include PointNet and PointNet++. The disadvantages of the PointNet method are its poor ability to capture local information and its difficulty in processing large-scale point cloud data. The disadvantages of the PointNet++ method are its relatively complex network structure, long training time, and high hardware resource requirements. Voxel-based methods convert point clouds into voxel representations and perform feature learning and reconstruction in voxel space. Common algorithms include the 3DCNN method. Its disadvantages are that the voxelization process loses some detailed information, and it consumes a lot of computation and memory.

[0004] In order to make the process of reconstructing point clouds more convenient and universal, it is necessary to develop a more targeted and adaptive point cloud reconstruction method based on the parametric description method. Summary of the Invention

[0005] The technical problem solved by the present invention is: to overcome the shortcomings of the existing technology and provide a feature reconstruction point cloud method and electronic equipment for encoder self-decoding, which is used to solve the problem that the aircraft shape is difficult to reconstruct into point cloud and three-dimensional digital model after parameterized description.

[0006] The technical solution of the present invention is:

[0007] The present invention discloses a feature reconstruction point cloud method for encoder self-decoding, comprising:

[0008] According to the multi-dimensional point cloud of points taken on the aircraft surface, a high-dimensional data feature matrix of the aircraft shape is obtained;

[0009] Split the high-dimensional data feature matrix into explicit feature matrix and implicit feature matrix;

[0010] According to the explicit feature matrix, a random point set matrix is ​​constructed;

[0011] According to the implicit feature matrix, the random point set matrix is ​​dimensionally upgraded and pooled to obtain the pooled matrix;

[0012] Merge several pooled matrices to obtain the feature reconstructed point cloud matrix.

[0013] Furthermore, in the above method, the high-dimensional data feature matrix of the aircraft shape is obtained based on the multi-dimensional point cloud of points taken on the aircraft surface, specifically:

[0014] Use the function F[C,D] to increase the dimension of the point cloud [N,C] to obtain the matrix [N,D];

[0015] Merge the matrix [N, D] with the point cloud [N, C] to get the matrix [N, C + D];

[0016] Using the pooling method, the matrix [N, C+D] is pooled to obtain the high-dimensional data feature matrix [1, C+D] of the aircraft shape;

[0017] Among them, F[C,D] is the function that upgrades the matrix from C dimension to D dimension, C is the dimension of the point cloud, D is the dimension after dimensionality upgrade, and N is the number of point clouds.

[0018] Furthermore, in the above method, the high-dimensional data feature matrix is ​​split into an explicit feature matrix and an implicit feature matrix, specifically:

[0019] The high-dimensional data feature matrix [1, C+D] is split into an explicit feature matrix [1, C] and an implicit feature matrix [1, D]; the explicit feature [1, C] contains the maximum and minimum values ​​of the aircraft surface point set; the implicit feature [1, D] contains the high-dimensional pooling features.

[0020] Furthermore, in the above method, a random point set matrix is ​​constructed according to the explicit feature matrix, specifically:

[0021] Randomly generate M according to the maximum and minimum values ​​in the explicit feature matrix i points, forming a random point set matrix R i =[M i ,C]; i = 1…n; n is the number of random point set matrices;

[0022] Generate M i The data distribution of points in C dimensions is guaranteed to satisfy:

[0023] Among them, k represents the dimension, ε represents any small amount, and q m is the matrix [M i ,C], point p n is a point in the point cloud [N,C], M i , N are matrices [M i ,C] and the number of points in the point cloud [N,C].

[0024] Furthermore, in the above method, the random point set matrix is ​​subjected to dimension increase and pooling screening to obtain a pooled matrix, specifically:

[0025] The random point set matrix [M i ,C] is dimensionalized by function F[C,D] to obtain the matrix [M i ,D];

[0026] Using the pooling method, the matrix [M i ,D] and implicit features [1,D] are pooled and filtered to obtain the pooled matrix [M i ',D];

[0027] Wherein, i=1…n; n is the number of random point set matrices.

[0028] Furthermore, in the above method, the feature reconstructed point cloud matrix is ​​specifically:

[0029]

[0030] Wherein, i=1…n; n is the number of random point set matrices.

[0031] The present invention discloses an electronic device comprising a processor, wherein the processor is used to execute a feature reconstruction point cloud method of encoder self-decoding.

[0032] The beneficial effects of the present invention and the prior art are:

[0033] (1) The present invention adopts a method of constrained screening of randomly generated points to achieve simple and efficient point cloud generation and reconstruction for high-dimensional latent space features, solving the problems of complex networks, long training time, and high resource consumption of common reconstruction methods based on deep learning;

[0034] (2) The present invention adopts a method of constructing a decoder based on the reverse engineering of the encoder, realizing a decoding method that does not introduce additional shape information, thereby solving the problem that the implicit function and deep learning network contain a large amount of shape information, which weakens the shape generalization ability;

[0035] (3) The present invention uses the encoder self-decoding method to make the encoding and decoding tools for point clouds relatively consistent, solving the problem of the decoder's additional storage occupying a large amount of memory and storage;

[0036] (4) The existing point cloud reconstruction method uses two independent models to implement the two processes of parametric description and reconstruction, so that part of the shape information is stored in the decoder. The present invention uses the high-dimensional features of the point cloud extracted by the point cloud parametric description method including the pooling step, establishes a randomly generated pooling screening method based on the mathematical logic and data characteristics of the pooling step, and constructs a reconstructed point cloud decoder that does not contain additional information, thereby realizing a fast and convenient feature-based point cloud reconstruction process. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a schematic diagram of the point cloud of modelnet40_airplane_01 of the present invention;

[0038] Figure 2 This is a schematic diagram of the feature reconstruction point cloud screening set C1 corresponding to the first point Q1 of the present invention;

[0039] Figure 3 This is a schematic diagram of the feature reconstruction point cloud screening set C2 corresponding to the second point Q2 of the present invention;

[0040] Figure 4 This is a schematic diagram of the feature reconstruction point cloud screening set C3 corresponding to the third point Q3 of the present invention;

[0041] Figure 5 Schematic diagram of the screening set and initial point cloud corresponding to all points in the sampling point sample of the present invention;

[0042] Figure 6 This is a schematic diagram of the point cloud obtained by feature reconstruction of the present invention;

[0043] Figure 7 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0044] The present invention discloses a feature reconstruction point cloud method for encoder self-decoding, comprising:

[0045] According to the multi-dimensional point cloud of points taken on the aircraft surface, a high-dimensional data feature matrix of the aircraft shape is obtained;

[0046] Split the high-dimensional data feature matrix into explicit feature matrix and implicit feature matrix;

[0047] According to the explicit feature matrix, a random point set matrix is ​​constructed;

[0048] According to the implicit feature matrix, the random point set matrix is ​​dimensionally upgraded and pooled to obtain the pooled matrix;

[0049] Merge several pooled matrices to obtain the feature reconstructed point cloud matrix.

[0050] Preferably, a high-dimensional data feature matrix of the aircraft shape is obtained based on a multi-dimensional point cloud of points taken on the aircraft surface, specifically:

[0051] Use the function F[C,D] to increase the dimension of the point cloud [N,C] to obtain the matrix [N,D];

[0052] Merge the matrix [N, D] with the point cloud [N, C] to get the matrix [N, C + D];

[0053] Using the pooling method, the matrix [N, C+D] is pooled to obtain the high-dimensional data feature matrix [1, C+D] of the aircraft shape;

[0054] Among them, F[C,D] is the function that upgrades the matrix from C dimension to D dimension, C is the dimension of the point cloud, D is the dimension after dimensionality upgrade, and N is the number of point clouds.

[0055] Preferably, the high-dimensional data feature matrix is ​​split into an explicit feature matrix and an implicit feature matrix, specifically:

[0056] The high-dimensional data feature matrix [1, C+D] is split into an explicit feature matrix [1, C] and an implicit feature matrix [1, D]; the explicit feature [1, C] contains the maximum and minimum values ​​of the aircraft surface point set; the implicit feature [1, D] contains the high-dimensional pooling features.

[0057] Preferably, a random point set matrix is ​​constructed based on the explicit feature matrix, specifically:

[0058] Randomly generate M according to the maximum and minimum values ​​in the explicit feature matrix i points, forming a random point set matrix R i =[M i ,C]; i = 1…n; n is the number of random point set matrices;

[0059] Generate M i The data distribution of points in C dimensions is guaranteed to satisfy:

[0060] Among them, k represents the dimension, ε represents any small amount, and q m is the matrix [M i ,C], point p n is a point in the point cloud [N,C], M i , N are matrices [M i ,C] and the number of points in the point cloud [N,C].

[0061] Preferably, the random point set matrix is ​​subjected to dimension increase and pooling screening to obtain a pooled matrix, specifically:

[0062] The random point set matrix [Mi ,C] is dimensionalized by function F[C,D] to obtain the matrix [M i ,D];

[0063] Using the pooling method, the matrix [M i ,D] and implicit features [1,D] are pooled and filtered to obtain the pooled matrix [M i ',D];

[0064] Wherein, i=1…n; n is the number of random point set matrices.

[0065] Preferably, the feature reconstructed point cloud matrix is ​​specifically:

[0066]

[0067] Wherein, i=1…n; n is the number of random point set matrices.

[0068] The invention discloses an electronic device comprising a processor, wherein the processor is used for executing a feature reconstruction point cloud method of encoder self-decoding.

[0069] Example

[0070] like Figure 7 As shown, this embodiment provides a method for reconstructing a point cloud by encoder self-decoding features, including the following steps:

[0071] 1. A C-dimensional point cloud containing N points on the surface of an aircraft is represented as an [N,C] matrix. A special method including a pooling step is used. This type of method usually consists of two steps. The first step is to upgrade the [N,C] matrix to an [N,D] matrix through a model F[C,D] and merge it with the initial [N,C] matrix to obtain an [N,C+D] matrix. The second step is to convert the [N,C+D] matrix into a [1,C+D] matrix through a pooling step such as maximum or minimum pooling, thereby obtaining the high-dimensional data features of the aircraft shape corresponding to the [N,C] matrix.

[0072] 2. Split the high-dimensional data feature matrix [1, C + D], where [1, C] is the explicit feature and [1, D] is the implicit feature. The explicit feature [1, C] contains information such as the maximum and minimum values ​​of the aircraft surface point set, while the implicit feature [1, D] is the high-dimensional pooling feature;

[0073] 3. Randomly generate M points in the C-dimensional space constrained by the maximum and minimum values ​​of the explicit features [1, C] to form a random point set matrix R1 = [M1, C]. The random generation process ensures that the data distribution in C dimensions satisfies

[0074] 4. The matrix [M1, C] is increased in dimension by F[C, D] to obtain the matrix [M1, D]. According to the same pooling method as that for obtaining the matrix [1, C + D], the matrix [M1, D] and the implicit feature [1, D] are pooled and screened to obtain the matrix [M1', D] that satisfies the pooling process. Taking the maximum pooling as an example, the calculation mask = index ([M1, D] < [1, D]) is used. The matrix [M1', D] that satisfies the pooling process is [[M1, D][mask]];

[0075] 5. Through multiple rounds of steps 3-4, a sufficient number of C-dimensional space points can be obtained, that is, through multiple matrices R2~R n =[M2,C]~[M n ,C], and obtain the corresponding [M2',D]~[M n ', D], merge to get the feature reconstructed point cloud matrix

[0076] The same method as that used in parametric feature description is used to generate implicit new features of random points, and the point set that meets the original implicit features is converted into a new point cloud that meets the parametric features, thus completing the reconstruction process from features to point clouds.

[0077] In this embodiment, take the airplane point cloud model "modelnet40_airplane_01" in modelnet40 as an example. Figure 1 As shown in the figure, we use FPS sampling with pointnet++-like methods to obtain its parameterized features.

[0078] 1. Perform FPS sampling on the sample point cloud modelnet40_airplane_01 to obtain the corresponding downsampled point samples, which contain n points;

[0079] 2. Implement this feature reconstruction point cloud method on the point cloud within the spherical neighborhood of the first point Q1 in the downsampled point sample within the range of d = 0.07. First, use the model constructed by the pointnet++ algorithm to obtain the parameterized feature A1 of the spherical neighborhood within the first point;

[0080] 3. Randomly generate a point P1 in a spherical neighborhood with a range of d = 0.07, and perform a dimensionality-increasing operation on it using the same PointNet++ model to obtain B1;

[0081] 4. Use A1 to perform pooling and screening on B1 to obtain the screening set C1 of points in P1 that meet the characteristics of A1.

[0082] like Figure 2 As shown;

[0083] 5. Continue to perform steps 2, 3, and 4 on the second point Q2 in the downsampled point sample to obtain the filter set C2 corresponding to Q2, as shown in Figure 3 As shown;

[0084] 6. Perform steps 2, 3, and 4 on all subsequent points Q3, Q4, ..., Qn in the downsampled point sample, thereby obtaining the filter set C3, C4, ..., Cn corresponding to each point in the downsampled point sample, as shown in Figure 4 、 Figure 5 As shown;

[0085] 7. Merge C1, C2, C3, C4, ..., Cn to obtain the feature reconstructed point cloud corresponding to modelnet40_airplane_01, such as Figure 6 shown.

[0086] This method utilizes parametric feature extraction and point cloud processing logic, leveraging the symmetry and reversibility of pooling operations to reconstruct parametric features into point clouds. This method significantly improves reconstruction efficiency, enabling feature-to-point cloud reconstruction without requiring decoder training. It is a powerful tool for efficiently combining CAD models, point clouds, and parametric features.

[0087] The present invention is not limited to the reconstruction process implemented on a single point cloud feature, but can be extended to the feature reconstruction operation of any number of point cloud sets.

[0088] The above describes the method of the present invention in detail. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The above specific examples are intended only to facilitate understanding of the method and its core concept. It should be noted that any simple modifications to the above examples that do not depart from the technical essence of the present invention fall within the technical scope of the present invention.

[0089] The contents not described in detail in the specification of the present invention belong to the common knowledge of those skilled in the art.

Claims

1. A feature reconstruction point cloud method for encoder self-decoding, characterized in that: include: According to the multi-dimensional point cloud of points taken on the aircraft surface, a high-dimensional data feature matrix of the aircraft shape is obtained; Split the high-dimensional data feature matrix into explicit feature matrix and implicit feature matrix; According to the explicit feature matrix, a random point set matrix is ​​constructed; According to the implicit feature matrix, the random point set matrix is ​​dimensionally upgraded and pooled to obtain the pooled matrix; Merge several pooled matrices to obtain the feature reconstructed point cloud matrix.

2. The feature reconstruction point cloud method of encoder self-decoding according to claim 1 is characterized in that: The multi-dimensional point cloud of points taken on the aircraft surface is used to obtain the high-dimensional data feature matrix of the aircraft shape, specifically: Use the function F[C,D] to increase the dimension of the point cloud [N,C] to obtain the matrix [N,D]; Merge the matrix [N, D] with the point cloud [N, C] to get the matrix [N, C + D]; Using the pooling method, the matrix [N, C+D] is pooled to obtain the high-dimensional data feature matrix [1, C+D] of the aircraft shape; Among them, F[C,D] is the function that upgrades the matrix from C dimension to D dimension, C is the dimension of the point cloud, D is the dimension after dimensionality upgrade, and N is the number of points in the point cloud.

3. The feature-reconstructed point cloud method of encoder self-decoding according to claim 2, characterized in that: The high-dimensional data feature matrix is ​​split into an explicit feature matrix and an implicit feature matrix, specifically: The high-dimensional data feature matrix [1, C+D] is split into an explicit feature matrix [1, C] and an implicit feature matrix [1, D]; the explicit feature [1, C] contains the maximum and minimum values ​​of the aircraft surface point set; the implicit feature [1, D] contains the high-dimensional pooling features.

4. The feature-reconstructed point cloud method of encoder self-decoding according to claim 3, characterized in that: According to the explicit feature matrix, the random point set matrix is ​​constructed, specifically: Randomly generate M according to the maximum and minimum values ​​in the explicit feature matrix i points, forming a random point set matrix R i =[M i ,C]; i = 1…n; n is the number of random point set matrices; Generate M i The data distribution of points in C dimensions is guaranteed to satisfy: Among them, k represents the dimension, ε represents any small amount, and q m is the matrix [M i ,C], point p n is a point in the point cloud [N,C], M i , N are matrices [M i ,C] and the number of points in the point cloud [N,C].

5. The feature-reconstructed point cloud method of encoder self-decoding according to claim 4, characterized in that: The random point set matrix is ​​subjected to dimension increase and pooling screening to obtain a pooled matrix, specifically: The random point set matrix [M i ,C] is dimensionalized by function F[C,D] to obtain the matrix [M i ,D]; Using the pooling method, the matrix [M i ,D] and implicit features [1,D] are pooled and filtered to obtain the pooled matrix [M i ',D]; Wherein, i=1…n; n is the number of random point set matrices.

6. The feature-reconstructed point cloud method of encoder self-decoding according to claim 5, characterized in that: The feature reconstructed point cloud matrix is ​​specifically: Wherein, i=1…n; n is the number of random point set matrices.

7. An electronic device, characterized in that: The method comprises a processor configured to execute the feature-reconstructed point cloud method of encoder self-decoding according to any one of claims 1 to 6.

Images