Non-comprehensive multi-view face image clustering method based on mark point and tensor multidimensional graph learning

By adopting a method based on marker points and tensor multidimensional graph learning, the clustering problem of non-comprehensive multi-view face image data is solved, improving the clustering accuracy and stability, reducing computational complexity, and realizing more efficient multi-view face image data processing.

CN119048784APending Publication Date: 2024-11-29SHENZHEN WANZHIDA TECH CO LTD

Patent Information

Application Number
CN202411059424.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-03
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

Existing multi-view clustering methods can only handle clustering problems of comprehensive multi-view data, and cannot effectively solve clustering problems of non-comprehensive multi-view face image data.

Method used

We employ a method based on marker points and tensor multidimensional graph learning. By extracting existence samples from each viewpoint, we learn viewpoint-bound marker points, construct a viewpoint-bound marker point graph, perform orthogonal nonnegative matrix decomposition, stack soft indicator matrices and minimize the tensor Schattenp norm, combine a unified bipartite graph framework and tensor low-rank constraints, fuse information between viewpoints, and finally perform singular value decomposition and K-means clustering.

Benefits of technology

It improves the clustering accuracy, stability, and computational efficiency of non-comprehensive multi-view face image datasets, effectively captures high-order correlations and complementary information between viewpoints, and reduces computational complexity and memory consumption.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure FT_1
    Figure FT_1
  • Figure FT_2
    Figure FT_2
  • Figure BDA0004977396260000021
    Figure BDA0004977396260000021
Patent Text Reader

Abstract

A non-comprehensive multi-view face image clustering method based on mark point and tensor multidimensional graph learning belongs to the field of face image clustering processing in multimedia information processing, and comprises the following steps: firstly, extracting existing samples in each view, and learning view specific mark points from the existing samples; and representing the distribution of all samples including non-comprehensive samples. Selecting mark points by using a direct alternate sampling method, constructing a mark point diagram with a specific view angle, and performing orthogonal non-negative matrix factorization on the mark point diagram of each view angle to obtain a soft indication matrix with a specific view angle; and stacking the soft indication matrixes of all the visual angles into a tensor, and capturing high-order correlation and complementary information among the visual angles by minimizing a tensor Schattenp-norm. The incompleteness between the visual angles is effectively processed through a unified bipartite graph frame and tensor low-rank constraint, and then the soft indication matrix of each visual angle is fused to obtain a final consensus indication matrix. And carrying out clustering analysis by using singular value decomposition and K-means clustering to obtain a final clustering result. Therefore, the adaptability and robustness of the model are promoted. Compared with other methods, the method is higher in accuracy and more stable in performance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of face image data clustering processing in multimedia information processing, and specifically relates to a non-comprehensive multi-view face image clustering method based on landmark point and tensor multi-dimensional graph learning. Background Art

[0002] In the field of face recognition and multimedia information processing, face image data is a widely available data with important application value. Multi-view clustering analysis of face image data can mine the characteristics of faces under different angles, expressions and lighting conditions, as well as potential category information such as similarities and differences between different faces. Among the many learning techniques for multi-view data, multi-view clustering is one of the most popular techniques. It can use information from multiple perspectives to perform unsupervised classification of data and can be applied to a variety of practical applications such as identity authentication, sentiment analysis and face recognition. However, in the process of collecting face image data, face information may be lost, resulting in inconsistent observability of each sample under different perspectives, thus causing the problem of non-comprehensive multi-view face image data. However, the existing multi-view clustering methods can only handle the clustering problem of comprehensive multi-view data. Therefore, an effective and novel technical method is needed to solve the clustering problem of non-comprehensive multi-view face image data. Summary of the invention

[0003] In order to solve the above problems, the present invention provides a non-comprehensive multi-view face image clustering method based on landmark point and tensor multi-dimensional graph learning, characterized in that the method comprises the steps of:

[0004] Extract the existence samples in each view for the non-comprehensive multi-view face image dataset;

[0005] Learning viewpoint-limited landmarks from existing samples to represent the distribution of all samples including non-comprehensive samples, using direct alternating sampling to select landmarks and constructing a viewpoint-limited landmark map;

[0006] Perform orthogonal non-negative matrix decomposition on the landmark point map of each view to obtain the view-limited soft indicator matrix;

[0007] The soft indicator matrices of each view are stacked into a tensor, and the high-order correlation and complementary information between views are captured by minimizing the tensor Schattenp-norm;

[0008] Through a unified bipartite graph framework and tensor low-rank constraints, the incompleteness between perspectives is effectively handled, and then the soft indicator matrix of each perspective is fused to obtain the final consensus indicator matrix;

[0009] The obtained consensus indicator matrix is ​​subjected to singular value decomposition, and the first c components of the left singular vector are selected for K-means clustering (c is the number of clusters) to obtain the final clustering result. Finally, the clustering accuracy of the final clustering result is calculated.

[0010] Furthermore, the existence samples of each perspective extracted from the non-comprehensive face image dataset are defined as:

[0011]

[0012] Among them, X represents the sample feature matrix of all viewing angles, X (v) is the existence feature matrix of view v, n v is the number of non-comprehensive samples of view v, d v is the characteristic dimension of view v, and m is the number of viewpoints.

[0013] Furthermore, the expression formula of the alternate sampling method for selecting landmark points is:

[0014]

[0015] in, represents the newly selected landmark point of the vth perspective, represents the i-th sample of the viewing angle v, represents the jth landmark point selected by the view v, and S represents the currently selected landmark point set. First, a sample is randomly selected as the landmark point and set to y1. Then, for each sample x i ,pass Calculate the minimum distance to all landmarks in the current landmark set S, and then select the sample with the largest distance to the current landmark set as the latest landmark of the current perspective.

[0016] Furthermore, the expression formula of the orthogonal non-negative matrix decomposition is:

[0017]

[0018] in, Represents the landmark map matrix of the v-th view, where n is the number of data points, m is the number of landmarks, and k is the number of clusters. Represents the soft indicator matrix of the v-th view. The landmark point representation matrix representing the v-th view. Denotes the identity matrix G (v) The column orthogonality constraint.

[0019] Furthermore, the expression formula for minimizing the tensor Schattenp-norm is:

[0020]

[0021] in, is a tensor The i-th front slice of is the Schatten p-norm of the matrix.

[0022] Furthermore, the unified bipartite graph framework combined with the tensor low-rank constraint is expressed as:

[0023]

[0024] Among them, B (v) =X (v) A (v) Indicates the construction of landmark points for incomplete data, F (v) ' is the soft indicator matrix of a specific perspective obtained through the unified bipartite graph framework and tensor low-rank constraint, A (v) is the landmark point selection matrix of the vth view, α v is the weight parameter of the vth perspective, β is the regularization parameter, and F is the final consensus indicator matrix.

[0025] Furthermore, according to the final clustering result, the clustering accuracy is calculated.

[0026] The present invention provides a non-comprehensive multi-view face image data clustering method based on the integration of local and global anchor graphs, which has the following advantages:

[0027] (1) The method adopts a multi-view clustering framework, making full use of the view consistency and difference of the non-comprehensive multi-view face image dataset to effectively mine the potential cluster structure. By combining the information of each view, it can more accurately capture the features of the face image at different viewpoints and improve the clustering accuracy.

[0028] (2) The method adopts the direct alternating sampling (DAS) technique to select landmarks and construct a view-specific landmark map. In this way, not only the computational complexity and memory consumption are reduced, but also the impact of non-comprehensive multi-view face image data is effectively mitigated. The DAS method can more stably select representative landmarks and avoid the instability caused by the traditional K-means method.

[0029] (3) The method adopts a tensor low-rank constraint to capture the high-order correlation and complementary information between views by minimizing the tensor Schattenp-norm. The tensor low-rank constraint can effectively handle the incompleteness between views and ensure the stability and accuracy of the clustering results. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0031] Figure 1 It is a flowchart of a non-comprehensive multi-view face image clustering method based on landmark point and tensor multi-dimensional graph learning provided by the present invention;

[0032] Figure 2 This is a schematic diagram of part of the ORL face image dataset; DETAILED DESCRIPTION

[0033] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings. It should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, the description of well-known structures and technologies is omitted to avoid unnecessary confusion of the concept of the present invention.

[0034] like Figure 1 The present invention provides a non-comprehensive multi-view face image clustering method based on landmark point and tensor multi-dimensional graph learning, and the method steps are as follows:

[0035] Step S110: First, extract the existing samples of each view from the non-comprehensive multi-view face image dataset. The existing samples are composed of a set of feature matrices and a mask matrix M∈{0,1} n×m Indicates that n v is the number of non-comprehensive samples of view v, d v is the feature dimension of the view v, m is the number of views, specifically, if there is a j-th view of the i-th sample, then M[i,j]=1.

[0036] Step S120: By learning the view-limited landmark points from the existing samples and selecting the landmark points by the direct alternating sampling method, a view-limited landmark point map is constructed to represent the distribution of all samples. The expression formula is as follows:

[0037]

[0038] Among them, x i represents the i-th sample, y j represents the jth selected landmark point, and S represents the currently selected landmark point set. First, a sample is randomly selected as the landmark point and set to y1. Then, for each sample x i, calculate the minimum distance to all landmark points in the current landmark point set S, and then select the sample with the largest distance to the current landmark point set as the new landmark point.

[0039] Step S130: by performing orthogonal non-negative matrix decomposition on the landmark point graph of each view, a soft indicator matrix limited by the view angle is obtained, and its expression formula is as follows:

[0040]

[0041] in, Represents the landmark map matrix of the v-th view, where n is the number of data points and m is the number of landmark points. Represents the soft indicator matrix of the v-th view. The landmark point representation matrix representing the v-th view. Denotes the identity matrix G (v) The column orthogonality constraint.

[0042] Step S140: stack the soft indicator matrices of each view into a tensor, and capture the high-order correlation and complementary information between views by minimizing the tensor Schattenp-norm. The expression formula is as follows:

[0043]

[0044] in, is a tensor The i-th front slice of is the Schatten p-norm of the matrix.

[0045] Step S150: Through a unified bipartite graph framework and tensor low-rank constraints, the incompleteness between perspectives is effectively handled to obtain the final consensus indicator matrix, which is expressed as follows:

[0046]

[0047] Among them, B (v) =X (v) A (v) Represents the construction of landmark points for incomplete data, F (v) ' is the view-specific soft indicator matrix obtained through the unified bipartite graph framework and tensor low-rank constraint, A (v) is the landmark point selection matrix of the vth view, α v is the weight parameter of the vth perspective, β is the regularization parameter, and F is the final consensus indicator matrix.

[0048] Step S160: Perform singular value decomposition on the obtained consensus indicator matrix, and select the first c components of the left singular vector for K-means clustering (c is the number of clusters) to obtain the final clustering result. Finally, calculate the clustering accuracy of the final clustering result.

[0049] Through this implementation case, the existing samples in each perspective are first extracted. The perspective-limited landmark points are learned from these existing samples to represent the distribution of all samples including non-comprehensive samples, and the landmark points are selected by the direct alternating sampling method to construct the perspective-limited landmark point map. Next, the landmark point map of each perspective is subjected to orthogonal non-negative matrix decomposition to obtain the perspective-limited soft indicator matrix. Then, the soft indicator matrices of each perspective are stacked into a tensor, and the high-order correlation and complementary information between perspectives are captured by minimizing the tensor Schattenp-norm. Through a unified bipartite graph framework and tensor low-rank constraints, the incompleteness between perspectives is effectively handled, and then the soft indicator matrices of each perspective are fused to obtain the final consensus indicator matrix. The obtained consensus indicator matrix is ​​subjected to singular value decomposition, and the first c components of the left singular vector are selected for K-means clustering (c is the number of clusters) to obtain the final clustering result. Finally, the clustering accuracy of the final clustering result is calculated.

[0050] To further illustrate, assuming that a non-comprehensive multi-view face image dataset is clustered according to this embodiment, a clustering result with a higher clustering accuracy than most methods will be obtained.

[0051] Detailed description of the invention Results

[0052] This implementation uses a publicly available face image dataset and simulates a non-comprehensive face image dataset according to different non-comprehensive sample ratios. The details of the dataset are described as follows:

[0053] ORL face image is a face dataset of 40 people. The dataset has the following three perspectives:

[0054] Perspective 1 is grayscale intensity. Grayscale intensity perspective is the most basic image representation. The value of each pixel represents the brightness of the pixel. In the ORL dataset, the size of each image is 92×112 pixels and the grayscale level is 256.

[0055] View 2 is the local binary pattern, which is a texture description operator that generates a binary pattern by comparing the grayscale values ​​of each pixel with its surrounding pixels.

[0056] View 3 is the Gabor filter, which is a linear filter that can capture edge, texture, and frequency information in an image.

[0057] In order to verify the superiority of this implementation (OURS), this implementation is compared with several existing methods for clustering animal image data, including AWMVC, BMVC, FMVCC, DCCAE and other methods. The clustering accuracy (ACC) of these methods for five non-comprehensive sample ratios on the above public datasets will be compared. The specific data comparison is shown in Table 1.

[0058] Table 1 Clustering accuracy of face dataset (%)

[0059]

[0060] By comparing the data in the above table, it can be clearly seen that OURS achieves the best performance and significantly improves the clustering accuracy of the non-comprehensive multi-view face image dataset. The quantitative results fully demonstrate the superiority of OURS, because OURS can better capture the perspective consistency information and perspective difference information in the non-comprehensive multi-view face image dataset. OURS effectively captures the high-order correlation and complementary information between perspectives by combining direct alternating sampling, positive and negative matrix decomposition, and tensor Schattenp-norm constraints, thereby improving the robustness to face image data. A large number of experiments have shown that this method is superior to existing methods. Regarding the setting of the parameters of this embodiment, in all experiments, the number of landmark points n a It is set to 5c, c is the number of clusters, the number of neighbors k is set to 0.05c, and the regularization parameter β is set to 0.1.

[0061] This embodiment proposes a non-comprehensive multi-view face image clustering method based on landmark points and tensor multi-dimensional graph learning, which is used to perform cluster analysis on face image data. By extracting existing samples in the non-comprehensive multi-view face image data set, the landmark points are selected by a direct alternating sampling method and a landmark point graph is constructed, and an orthogonal non-negative matrix decomposition is performed to obtain a soft indicator matrix limited by the perspective. The soft indicator matrices of each perspective are stacked into a tensor, and the high-order correlations are captured by minimizing the tensor Schattenp-norm. The consensus indicator matrix is ​​fused, and the final clustering result is obtained by singular value decomposition and K-means clustering, and the clustering accuracy is calculated. The experimental results on five non-comprehensive sample ratios of the public data set ORL show that this embodiment has higher clustering accuracy and better superiority than other methods.

Claims

1. A non-comprehensive multi-view face image clustering method based on landmark point and tensor multi-dimensional graph learning, characterized in that: The method comprises the steps of: Extract the existence samples in each view for the non-comprehensive multi-view face image dataset; Learning viewpoint-limited landmarks from existing samples to represent the distribution of all samples including non-comprehensive samples, using direct alternating sampling to select landmarks and constructing a viewpoint-limited landmark map; Perform orthogonal non-negative matrix decomposition on the landmark point map of each view to obtain the view-limited soft indicator matrix; The soft indicator matrices of each view are stacked into a tensor, and the high-order correlation and complementary information between views are captured by minimizing the tensor Schattenp-norm; Through a unified bipartite graph framework and tensor low-rank constraints, the non-comprehensiveness between perspectives is effectively handled, and then the soft indicator matrix of each perspective is fused to obtain the final consensus indicator matrix; The obtained consensus indicator matrix is ​​subjected to singular value decomposition, and the first c components of the left singular vector are selected for K-means clustering (c is the number of clusters) to obtain the final clustering result. Finally, the clustering accuracy of the final clustering result is calculated.

2. According to claim 1, a non-comprehensive multi-view face image clustering method based on landmark point and tensor multi-dimensional graph learning is characterized in that: Extract the presence samples of each view for a non-comprehensive multi-view face image dataset: Among them, X represents the sample feature matrix of all viewing angles, X (v) is the existence feature matrix of view v, n v is the number of non-comprehensive samples of view v, d v is the feature dimension of view v, and m is the number of viewpoints.

3. The non-comprehensive multi-view face image clustering method based on landmark point and tensor multi-dimensional graph learning according to claim 1, characterized in that: A direct alternating sampling method is used to select landmark points for each view to construct a landmark point map limited by the view angle. The alternating sampling method for selecting landmark points is defined as follows: in, represents the newly selected landmark point of the vth perspective, represents the i-th sample of the viewing angle v, represents the jth landmark point selected by the view v, S represents the currently selected landmark point set, first randomly select a sample as the landmark point and set it to y1, then for each sample x i ,pass Calculate the minimum distance to all landmarks in the current landmark set S, and then select the sample with the largest distance to the current landmark set as the latest landmark of the current perspective.

4. The non-comprehensive multi-view face image clustering method based on landmark point and tensor multi-dimensional graph learning according to claim 1, characterized in that: The landmark point map of each view is subjected to orthogonal non-negative matrix decomposition to obtain the soft indicator matrix limited by the view, where the expression formula of the orthogonal non-negative matrix decomposition is as follows: in, Represents the landmark map matrix of the vth view, where n is the number of data points, m is the number of landmark points, and k is the number of clusters. represents the soft indicator matrix of the v-th view, The landmark points representing the v-th viewing angle represent the matrix, Denotes the identity matrix G (v) The column orthogonality constraint.

5. The non-comprehensive multi-view face image clustering method based on landmark point and tensor multi-dimensional graph learning according to claim 1, characterized in that: The soft indicator matrices of each view are stacked into a tensor, and the high-order correlation and complementary information between views are captured by minimizing the tensor Schattenp-norm. The definition of minimizing the tensor Schattenp-norm is as follows: in, is a tensor The i-th front slice of is the Schatten p-norm of the matrix.

6. The non-comprehensive multi-view face image clustering method based on landmark point and tensor multi-dimensional graph learning according to claim 1, characterized in that: Through the unified bipartite graph framework and tensor low-rank constraints, the incompleteness between perspectives is effectively handled. Finally, the soft indicator matrix of each perspective is fused to obtain the final consensus indicator matrix. The unified bipartite graph framework combined with tensor low-rank constraints and consensus matrix is ​​defined as follows: Among them, B (v) =X (v) A (v) Indicates the construction of landmark points for incomplete data, F (v) ' is the soft indicator matrix of a specific perspective obtained through the unified bipartite graph framework and tensor low-rank constraint, A (v) is the landmark point selection matrix of the vth view, α v is the weight parameter of the vth perspective, β is the regularization parameter, and F is the final consensus indicator matrix.

7. The non-comprehensive multi-view face image clustering method based on landmark point and tensor multi-dimensional graph learning according to claim 1, characterized in that: According to the obtained clustering results, the clustering accuracy on the non-comprehensive multi-view face image dataset is calculated.

Citation Information

Patent Citations

  • Mark point positioning method and device, computer device and storage medium

    CN110176035A

  • Face image clustering method with incomplete feature description

    CN114973370A

  • Subspace clustering method for multi-view incomplete image

    CN116524222A

  • Deep online anchor point subspace learning-based incomplete multi-view large-scale animal image clustering method

    CN117994550A

  • High-order correlation preserved incomplete multi-view subspace clustering method and system

    WO2022228348A1

Cited By

  • Multi-view agricultural image clustering method based on enhanced multi-order similarity learning

    CN122023384A