Multi-view consistent image clustering method and system based on embedded feature extraction

Abstract

The present invention proposes a multi-view consistent image clustering method and system based on embedded feature extraction, which relates to the technical field of multi-view image data clustering processing. By processing the data matrix of the original multi-view data, the sample embedding and feature embedding of each view are obtained, and the feature embedding of each view is subjected to group sparse constraints to achieve denoising of the feature embedding of the original multi-view data. The feature embedding and sample embedding of the original multi-view data are then connected through a bipartite graph to perform feature extraction. The embedded knowledge is transmitted between the feature embedding and the sample embedding through the bipartite graph, and the information in the view is fully learned. The redundant information and noise contained in the data features are effectively eliminated to maximize the consistency learning of the model for the multi-view data, so that the final learned consistency graph can be more accurate. A rank constraint is added to the Laplacian matrix corresponding to the final consistency graph to directly obtain the clustering result.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-view image data clustering processing, and more specifically, to a multi-view image data clustering method based on embedded feature extraction. Figure 1 Consistent image clustering method and system. Background Art

[0002] Data clustering refers to the process of dividing a data set without classification labels into several clusters. It is an unsupervised classification method. Its division is based on the principle of "birds of a feather flock together". The samples in the same cluster are as similar as possible, and the similarity between samples in different clusters is as low as possible.

[0003] With the development of the Internet, data has exploded, and there is a lot of important information hidden behind these surging data. Therefore, people hope to be able to analyze them at a higher level in order to make better use of these data. Clustering these unlabeled data for subsequent processing can save a lot of resources. Now, data can be described from multiple sources or several different feature sets. For example, an image can be described by different features, such as GIST, LBP, SIFT, HOG features, etc.; the same book can be translated into different languages; the change process of sensor signals can be described in time domain and frequency domain. These data are called multi-view data. Multi-view data contains more complementary information from different views, which can further improve the clustering performance of data clustering. Therefore, in recent years, multi-view data has been widely used in the research field of data clustering, such as using multi-view text data for article classification, news classification and spam detection.

[0004] Clustering of multi-view image data is more widely used, and multi-view image data is used for face recognition, biological species classification, retrograde detection, image retrieval, and automatic album classification, etc. However, for multi-view image data, since the acquisition of images is affected by factors such as light, target distance, and resolution, its features often contain redundant information and noise. Image quality defects caused by problems such as weather and equipment are difficult to improve during the image acquisition process. These features containing redundant information and noise will affect the data clustering effect and lead to incorrect image clustering results. Therefore, in the process of multi-view image clustering, how to effectively remove the redundant information and noise of multi-view image data features is very important to improve the accuracy of clustering.

[0005] Although the existing multi-view clustering methods have improved the effect of multi-view image clustering from various angles, such as a multi-view K multi-mean image clustering method based on adaptive weights disclosed in the prior art, which uses multiple sub-cluster centers to capture the data distribution of each cluster in each view, and can adaptively assign weights to each view, so as to reasonably fuse the complementary information and compatible information of different views to construct a shared bipartite graph, and finally impose Laplace rank constraints on the shared bipartite graph to divide it into C clusters to achieve clustering of image samples, and then use it to solve the problem of multi-view image recognition and detection, but they usually do not consider processing the original image multi-view data features, but directly learn similarity graphs from the original multi-view data features containing a large amount of redundant information and noise. The similarity graphs learned in this way become inaccurate because the data features contain too much redundant information and noise, especially for multi-view data such as images containing high-dimensional features. It is very unfavorable, because the higher the dimensionality of the data, the more redundant information and noise there are, resulting in the final clustering performance being severely limited and the image cannot be correctly clustered. Summary of the invention

[0006] In order to solve the problem of how to effectively remove the redundant information and noise of multi-view image data features in the multi-view image clustering process, the present invention provides a multi-view image clustering method based on embedded feature extraction. Figure 1 Consistent image clustering method and system.

[0007] In order to achieve the above technical effects, the technical solution of the present invention is as follows:

[0008] A multi-view feature extraction based on embedded Figure 1 A consistent image clustering method, the method comprising the following steps:

[0009] Determine an original multi-view data matrix, and construct a similarity matrix for each view based on sample elements in the multi-view data matrix;

[0010] Calculate the Laplacian matrix of each view according to the similarity matrix of each view, and obtain the sample embedding of each view based on the Laplacian matrix;

[0011] Perform singular value decomposition on the original multi-view data matrix to obtain the left singular matrix of each view, and construct the feature embedding of each view based on the left singular matrix of each view;

[0012] The feature embedding of each view is constrained to be group sparse, and the sample embedding and feature embedding are connected together through a bipartite graph for feature extraction;

[0013] Based on the sample embedding of each view, a block diagonal matrix is ​​formed, and the consistency graph of the block diagonal matrix is ​​learned through the cost function, and then the rank constraint is imposed on the consistency graph;

[0014] Integrate feature extraction and consistency graph learning into a unified learning framework, obtain the objective function and solve it to obtain the consistency graph;

[0015] Connected components are determined from the consistency graph to obtain image clustering results.

[0016] In this technical solution, as a whole, by processing the data matrix of the original multi-view data, the sample embedding and feature embedding of each view are obtained, and the feature embedding of each view is subjected to group sparse constraints to achieve denoising of the feature embedding of the original multi-view data. The feature embedding and sample embedding of the original multi-view data are then connected through a bipartite graph for feature extraction. The embedded knowledge is transferred between the feature embedding and the sample embedding through the bipartite graph, thereby promoting each other to fully learn the information within the view, and the influence of redundant information and noise contained in the data features is effectively eliminated. At the same time, due to the divergence cost, the consistency graph and the sample embeddings of different views are mutually regularized to maximize the model's consistency learning of multi-view data and learn more accurate multi-view features, so that the final learned consistency graph can be more accurate. A rank constraint is added to the Laplace matrix corresponding to the final consistency graph to directly obtain the clustering result without the need for subsequent processing, thereby reducing the uncertainty of the clustering result.

[0017] Preferably, let the original multi-view data matrix be represented as Among them, d v represents the feature dimension of the vth view, n represents the number of samples of the original data, V represents the number of views, X (v) represents the original multi-view data matrix;

[0018] set up represents the i-th sample element of the v-th view. The K-nearest neighbor algorithm is used to construct the similarity matrix of each view. The similarity matrix of each view is expressed as W v ,Depend on composition, represents the similarity between the i-th sample and the j-th sample in the v-th view, and the formula is:

[0019]

[0020] in, represents the set of k samples whose Euclidean distance is closest to the jth sample in the vth view.

[0021] Preferably, the process of calculating the Laplacian matrix of each view is:

[0022] S11. Obtain the degree matrix corresponding to the similarity matrix of the v-th view;

[0023] Let the degree matrix be represented by Dv , the expression is:

[0024]

[0025] S12. Based on the degree matrix and the similarity matrix of each view, the Laplacian matrix of each view is calculated, and the calculation expression is:

[0026] L v =D v -W v .

[0027] Preferably, when obtaining the sample embedding of each view based on the Laplacian matrix, a standard clustering algorithm is used on the Laplacian matrix, assuming H (v) Represents the sample embedding of the v-th view, and constructs the optimization model as:

[0028]

[0029]

[0030] Among them, tr() means (H (v) ) T L (v) H (v) The trace of the matrix, (H (v) ) T represents the transpose of the sample embedding of the vth view, n represents the number of samples, c represents the number of clusters, I represents the unit matrix, and then solves the sample embedding H (v) .

[0031] Preferably, the expression for performing singular value decomposition on the original multi-view data matrix is:

[0032]

[0033] Among them, U (v) represents the left singular matrix of the original data matrix for each view, d v represents the feature dimension of the vth view; Σ (v) is a diagonal matrix consisting of the singular values ​​of the original data matrix for each view, E (v) is the right singular value matrix of each original view data matrix,

[0034] When constructing the feature embedding of each view based on the left singular matrix of each view, take U (v) The first c feature vectors of construct the feature embedding A of each view (v) ,

[0035] Preferably, the 2,1 norm is used to perform group sparsity constraints on the feature embedding of each view, and the sample embedding and the feature embedding are connected together through a bipartite graph for feature extraction. The process expression satisfies:

[0036]

[0037] Among them, || || 2，1 represents the 2,1 norm; A (v) is the feature embedding of the v-th view, d v represents the feature dimension of the vth view; X (v) is the original data matrix of the vth view, n is the number of samples; H (v) is the sample embedding of the v-th view, c is the number of clusters.

[0038] Preferably, the sample is embedded in H (v) Transpose with sample embedding (H (v) ) T Multiply to get a block diagonal matrix The process of learning the consistency graph of the block diagonal matrix through the cost function and then rank-constraining the consistency graph satisfies:

[0039]

[0040] stS≥0,S1=1,rank(L S ) = nc

[0041] Among them, L S Represents the Laplace matrix corresponding to the consistency graph S, rank(L S ) = nc means that the rank constraint on the Laplacian matrix is ​​nc; represents the F-norm; β1 is the scale constant.

[0042] Preferably, the objective function is:

[0043]

[0044] The constraints are satisfied:

[0045]

[0046] Among them, β1 is the scale constant, β2 is the regularization parameter, and β3 is the trade-off factor;

[0047] When solving the objective function, the alternating direction multiplier method is used to iterate until the absolute value of the difference between the current objective function value and the previous objective function value is less than the set error threshold, the objective function converges, and the consistency diagram is obtained.

[0048] Preferably, the rank constraint on the Laplacian matrix corresponding to the consistency graph enables the learned consistency graph to achieve neighborhood allocation of multi-view data, the number of connected components of the consistency graph is equal to the number of classification clusters of the original data, each connected component corresponds to a classification cluster, and the clustering result is obtained from the consistency graph.

[0049] The present application also provides a multi-view based on embedded feature extraction Figure 1 A consistent image clustering system, the system comprising:

[0050] A similarity matrix construction module is used to determine the original multi-view data matrix and construct a similarity matrix for each view based on sample elements in the multi-view data matrix;

[0051] A sample embedding module is used to calculate the Laplacian matrix of each view according to the similarity matrix of each view, and obtain the sample embedding of each view based on the Laplacian matrix;

[0052] A feature embedding module is used to perform singular value decomposition on the original multi-view data matrix to obtain the left singular matrix of each view, and construct the feature embedding of each view based on the left singular matrix of each view;

[0053] The feature extraction module is used to perform group sparsity constraints on the feature embedding of each view and connect the sample embedding and feature embedding together through a bipartite graph for feature extraction;

[0054] The consistency graph learning module is based on the sample embedding of each view to form a block diagonal matrix, and the block diagonal matrix is ​​used for consistency graph learning through the cost function, and then the consistency graph is rank constrained;

[0055] The clustering module integrates feature extraction and consistency graph learning into a unified learning framework, obtains and solves the objective function, obtains the consistency graph, determines the connected components from the consistency graph, and obtains the image clustering results.

[0056] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0057] The present invention proposes a multi-view feature extraction method based on embedded feature extraction. Figure 1The consistent image clustering method and system processes the data matrix of the original multi-view data to obtain the sample embedding and feature embedding of each view, and performs group sparse constraints on the feature embedding of each view to achieve noise reduction of the feature embedding of the original multi-view data. The feature embedding and sample embedding of the original multi-view data are then connected through a bipartite graph for feature extraction. The embedded knowledge is transferred between the feature embedding and the sample embedding through the bipartite graph, which promotes each other to fully learn the information in the view, and the redundant information and noise contained in the data features are effectively eliminated. At the same time, due to the divergence cost, the consistency graph and the sample embeddings of different views are mutually regularized to maximize the consistency learning of the model for multi-view data, so that the final learned consistency graph can be more accurate. A rank constraint is added to the Laplace matrix corresponding to the final consistency graph to directly obtain the clustering result without the need for subsequent processing, thereby reducing the uncertainty of the clustering result. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 The multi-view based on embedded feature extraction proposed in Example 1 of the present invention Figure 1 Schematic diagram of the process of consistent image clustering method;

[0059] Figure 2 A flowchart showing the calculation of the Laplacian matrix of each view proposed in Embodiment 1 of the present invention;

[0060] Figure 3 The multi-view based embedding feature extraction proposed in Example 3 of the present invention Figure 1 Structure diagram of the consistent image clustering system. DETAILED DESCRIPTION

[0061] The drawings are for illustrative purposes only and should not be construed as limiting the present patent;

[0062] In order to better illustrate the present embodiment, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the actual size;

[0063] It is understandable to those skilled in the art that descriptions of certain well-known contents in the drawings may be omitted.

[0064] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0065] The positional relationships described in the drawings are only for illustrative purposes and should not be construed as limiting the present patent;

[0066] Example 1

[0067] like Figure 1 As shown, this embodiment proposes a multi-view method based on embedded feature extraction Figure 1A consistent image clustering method is provided, the method comprising the following steps:

[0068] S1. determining an original multi-view data matrix, and constructing a similarity matrix for each view based on sample elements in the multi-view data matrix;

[0069] In this embodiment, it can be understood that images captured at different camera angles are different views, and different views are multi-views. Assume that the original multi-view data matrix is ​​represented as Among them, d v represents the feature dimension of the vth view, n represents the number of samples of the original data, V represents the number of views, X (v) represents the original multi-view data matrix;

[0070] set up represents the i-th sample element of the v-th view. The K-nearest neighbor algorithm is used to construct the similarity matrix of each view. The similarity matrix of each view is expressed as W v ,Depend on composition, represents the similarity between the i-th sample and the j-th sample in the v-th view, and the formula is:

[0071]

[0072] in, represents the set of k samples whose Euclidean distance is closest to the jth sample in the vth view, where k is set to 10.

[0073] S2. Calculate the Laplacian matrix of each view according to the similarity matrix of each view, and obtain the sample embedding of each view based on the Laplacian matrix;

[0074] See also Figure 2 , the process of calculating the Laplacian matrix of each view is:

[0075] S11. Obtain the degree matrix corresponding to the similarity matrix of the v-th view;

[0076] Let the degree matrix be represented by D v , the expression is:

[0077]

[0078] S12. Based on the degree matrix and the similarity matrix of each view, the Laplacian matrix of each view is calculated, and the calculation expression is:

[0079] L v =D v -W v .

[0080] Based on the Laplacian matrix, when obtaining the sample embedding of each view, a standard clustering algorithm is used on the Laplacian matrix. Let H (v) Represents the sample embedding of the v-th view, and constructs the optimization model as:

[0081]

[0082]

[0083] Among them, tr() means (H (v) ) T L (v) H (v) The trace of the matrix, (H (v) ) T represents the transpose of the sample embedding of the vth view, n represents the number of samples, c represents the number of clusters, I represents the unit matrix, and then solves the sample embedding H (v) .

[0084] S3. Perform singular value decomposition on the original multi-view data matrix to obtain the left singular matrix of each view, and construct the feature embedding of each view based on the left singular matrix of each view;

[0085] The expression for singular value decomposition of the original multi-view data matrix is:

[0086]

[0087] Among them, U (v) represents the left singular matrix of the original data matrix for each view, d v represents the feature dimension of the vth view; Σ (v) is a diagonal matrix consisting of the singular values ​​of the original data matrix for each view, E (v) is the right singular value matrix of each original view data matrix,

[0088] When constructing the feature embedding of each view based on the left singular matrix of each view, take U (v) The first c feature vectors of construct the feature embedding A of each view (v) ,

[0089] S4. Perform group sparsity constraints on the feature embedding of each view, and connect the sample embedding and feature embedding together through a bipartite graph for feature extraction;

[0090] The 2,1 norm is used to constrain the feature embedding of each view to be group sparse, and the sample embedding and feature embedding are connected together through a bipartite graph for feature extraction. The process expression satisfies:

[0091]

[0092] Among them, || 2，1 represents the 2,1 norm; A (v) is the feature embedding of the v-th view, d v represents the feature dimension of the vth view; X (v) is the original data matrix of the vth view, n is the number of samples; H (v) is the sample embedding of the v-th view, c is the number of clusters.

[0093] S5. Based on the sample embedding of each view, a block diagonal matrix is ​​formed, and a consistency graph is learned for the block diagonal matrix through a cost function, and then a rank constraint is imposed on the consistency graph;

[0094] Embed the sample into H (v) Transpose with sample embedding (H (v) ) T Multiply to get a block diagonal matrix The process of learning the consistency graph of the block diagonal matrix through the cost function and then rank-constraining the consistency graph satisfies:

[0095]

[0096] stS≥0,S1=1,rank(L S ) = nc

[0097] Among them, L S Represents the Laplace matrix corresponding to the consistency graph S, rank(L S ) = nc means that the rank constraint on the Laplacian matrix is ​​nc; represents the F-norm; β1 is the scale constant.

[0098] S6. Integrate feature extraction and consistency graph learning into a unified learning framework, obtain the objective function and solve it to obtain the consistency graph;

[0099] The objective function is:

[0100]

[0101] The constraints are satisfied:

[0102]

[0103] Among them, β1 is the scale constant, β2 is the regularization parameter, and β3 is the trade-off factor. Here, β1, β2, and β3 are set to 30, 1, and 10 respectively.

[0104] When solving the objective function, the alternating direction multiplier method is used to iterate until the absolute value of the difference between the current objective function value and the previous objective function value is less than the set error threshold, the objective function converges, and the consistency graph is obtained. The specific steps are:

[0105] (1) Solve for H (v) , fixed A (v) and S, by discarding irrelevant terms, the objective function is expressed as:

[0106]

[0107]

[0108] Due to the orthogonality constraint, in fact H (v) lies on a Stiefel manifold. So H (v) The updates are as follows:

[0109]

[0110] The π operation represents the singular value decomposition (SVD) of the matrix. Indicates H (v) The matrix is ​​differentiated, and t1 is the step size.

[0111] In this embodiment, t1 is set to 0.1, and H (v) Iterate and update 10 times.

[0112] (2) Solve A (v) , fixed H (v) and S, by discarding irrelevant terms, the objective function can be expressed as:

[0113]

[0114] A variable G is introduced to separate the above problems. That is,

[0115]

[0116] For the above problem, the augmented Langrangian form is as follows:

[0117]

[0118] Where Y is the Lagrange multiplier. From the above formula, we can solve:

[0119] G=(β2D G +βI G ) -1 (βA (v) -Y)

[0120] Where DG is a diagonal matrix with each diagonal element being I G is the identity matrix of the same dimension as the G matrix.

[0121] We define the objective function with respect to A (v) The derivative of is set to 0, then:

[0122]

[0123] In addition, the update of Y is:

[0124] Y T =Y T-1 +β(GA (v) )

[0125] (3) Solve for S and fix H (v) and A (v) , by discarding irrelevant terms, the objective function can be expressed as:

[0126]

[0127] stS≥0,S1=1,rank(L S ) = nc

[0128] The rank constraint rank(L s ) = nc represents the Laplace matrix L corresponding to the consistency graph s The c smallest eigenvalues ​​of are all 0, which can be expressed as And it can be further expressed as And the constraints are I is the identity matrix. So the formula can be transformed into:

[0129]

[0130] stS≥0,S1=1

[0131] Where γ is a trade-off parameter used to implement rank(L s )=nc. In this embodiment, γ is initialized to 1. During the optimization iteration, if the connected component of S is greater than c, the value of γ is increased, otherwise it is decreased. Because each column in S is independent, the formula can be further converted into an unconstrained Lagrangian function:

[0132]

[0133] where η and ρ represent Lagrange multipliers, s j represents the j-th column vector of S, According to the KKT condition, we can solve s jfor:

[0134]

[0135] (4) Repeat (1) to (3) until the difference between the currently calculated objective function value and the last calculated objective function value is less than the set error threshold. At this time, the objective function is considered to have converged and step S7 is executed.

[0136] S7. Determine the connected components from the consistency graph and obtain the image clustering result.

[0137] Here, the rank constraint on the Laplacian matrix corresponding to the consistency graph enables the learned consistency graph to achieve neighborhood allocation of multi-view data. The number of connected components of the consistency graph is equal to the number of classification clusters of the original data. Each connected component corresponds to a classification cluster, and the clustering result is obtained from the consistency graph.

[0138] Example 2

[0139] This embodiment verifies the effectiveness of the method proposed in Example 1 by conducting simulation experiments on the method proposed in Example 1.

[0140] In this embodiment, the Yale dataset is used, which is a facial image dataset containing 165 photos of 15 people. Each person has 11 photos with different expressions and lighting conditions. Each picture is described by three features: 4096-dimensional intensity feature, 3304-dimensional LBP feature and 6750-dimensional Gabor feature.

[0141] In order to verify the effectiveness of the present invention, this embodiment tests the clustering performance of the proposed method on the Yale basic database, and compares it with five currently advanced multi-view clustering algorithms, namely: AASC (Affinity Aggregation for Spectral Clustering), MVGL (Graph Learning for Multiview Clustering), AWP (Multiview Clustering via Adaptively Weighted Procrustes), MCGC (Multiview Consensus Graph Clustering), GMC (Graph-Based Multi-View Clustering), and uses ACC (precision), NMI (normalized information), Recall (recall rate), F-score (F score), ARI (adjusted Rand coefficient) to evaluate the clustering performance. The comparison results are shown in Table 1.

[0142] Table 1

[0143]

[0144] It can be seen from Table 1 that the clustering performance of the present invention is significantly better than that of other comparative methods. The effectiveness of the present invention can be proved by the above simulation experiments.

[0145] Example 3

[0146] The present application also provides a multi-view based on embedded feature extraction Figure 1 Consistent image clustering system, see Figure 3 , the system comprising:

[0147] A similarity matrix construction module is used to determine the original multi-view data matrix and construct a similarity matrix for each view based on sample elements in the multi-view data matrix;

[0148] A sample embedding module is used to calculate the Laplacian matrix of each view according to the similarity matrix of each view, and obtain the sample embedding of each view based on the Laplacian matrix;

[0149] A feature embedding module is used to perform singular value decomposition on the original multi-view data matrix to obtain the left singular matrix of each view, and construct the feature embedding of each view based on the left singular matrix of each view;

[0150] The feature extraction module is used to perform group sparsity constraints on the feature embedding of each view and connect the sample embedding and feature embedding together through a bipartite graph for feature extraction;

[0151] The consistency graph learning module is based on the sample embedding of each view to form a block diagonal matrix, and the block diagonal matrix is ​​used for consistency graph learning through the cost function, and then the consistency graph is rank constrained;

[0152] The clustering module integrates feature extraction and consistency graph learning into a unified learning framework, obtains and solves the objective function, obtains the consistency graph, determines the connected components from the consistency graph, and obtains the image clustering results.

[0153] Obviously, the above embodiments of the present invention are only examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A multi-view consistent image clustering method based on embedded feature extraction, characterized in that: The method comprises the following steps: Determine an original multi-view data matrix, and construct a similarity matrix for each view based on sample elements in the multi-view data matrix; Calculate the Laplacian matrix of each view according to the similarity matrix of each view, and obtain the sample embedding of each view based on the Laplacian matrix; Perform singular value decomposition on the original multi-view data matrix to obtain the left singular matrix of each view, and construct the feature embedding of each view based on the left singular matrix of each view; The expression for singular value decomposition of the original multi-view data matrix is: in, represents the left singular matrix of the original data matrix for each view, , Indicates v The feature dimension of each view; is a diagonal matrix consisting of the singular values ​​of the original data matrix for each view, ; is the right singular value matrix of each original view data matrix, ; When constructing the feature embedding of each view based on the left singular matrix of each view, we take The first c feature vectors of construct the feature embedding of each view , ; The feature embedding of each view is constrained to be group sparse, and the sample embedding and feature embedding are connected together through a bipartite graph for feature extraction; The 2,1 norm is used to constrain the feature embedding of each view to be group sparse, and the sample embedding and feature embedding are connected together through a bipartite graph for feature extraction. The process expression satisfies: in, represents the 2,1 norm; For the v The feature embedding of each view, , Indicates The feature dimension of each view; For the The original data matrix of views, , n is the number of samples; For the Sample embeddings for views, , is the number of clusters; Based on the sample embedding of each view, a block diagonal matrix is ​​formed, and the consistency graph of the block diagonal matrix is ​​learned through the cost function, and then the rank constraint is imposed on the consistency graph; Integrate feature extraction and consistency graph learning into a unified learning framework, obtain the objective function and solve it to obtain the consistency graph; Connected components are determined from the consistency graph to obtain image clustering results.

2. The multi-view consistent image clustering method based on embedded feature extraction according to claim 1, characterized in that: Assume that the original multi-view data matrix is ​​represented as ,in, Indicates The feature dimensions of the view, n Represents the number of samples of the original data, V Indicates the number of views, X (v) represents the original multi-view data matrix; set up Indicates The i-th sample element of a view is constructed using the K nearest neighbor algorithm to construct the similarity matrix of each view. The similarity matrix of each view is expressed as ,Depend on composition, Indicates In the view i Samples and j The similarity between samples is expressed as: in, Indicates The set of k samples whose Euclidean distance to the j-th sample in the view is the closest.

3. The multi-view consistent image clustering method based on embedded feature extraction according to claim 2, characterized in that: The process of calculating the Laplacian matrix of each view is: S11. Find the The degree matrix corresponding to the similarity matrix of each view; Let the degree matrix be expressed as D v , the expression is: ； S12. Based on the degree matrix and the similarity matrix of each view, the Laplacian matrix of each view is calculated, and the calculation expression is: 。 4. The multi-view consistent image clustering method based on embedded feature extraction according to claim 3 is characterized in that: Based on the Laplacian matrix, when obtaining the sample embedding of each view, a standard clustering algorithm is used on the Laplacian matrix. Let Indicates The sample embedding of each view builds the optimization model as follows: in, express The trace of the matrix, Indicates The transpose of the sample embedding of the views, n represents the number of samples, represents the number of clusters, Represents the identity matrix, and then solves for the sample embedding .

5. The multi-view consistent image clustering method based on embedded feature extraction according to claim 1, characterized in that: Embed the sample Transpose with sample embedding Multiply to get a block diagonal matrix , the process of learning the consistency graph of the block diagonal matrix through the cost function and then rank constraining the consistency graph satisfies: in, represents the Laplace matrix corresponding to the consistency graph S, The rank constraint on the Laplacian matrix is ​​expressed as ; represents the F-norm; is a scale constant.

6. The multi-view consistent image clustering method based on embedded feature extraction according to claim 5, characterized in that: The objective function is: The constraints are satisfied: in, is the scale constant, Regularization parameter, is a weighing factor; When solving the objective function, the alternating direction multiplier method is used to iterate until the absolute value of the difference between the current objective function value and the previous objective function value is less than the set error threshold, the objective function converges, and the consistency diagram is obtained.

7. The multi-view consistent image clustering method based on embedded feature extraction according to claim 6, characterized in that: The rank constraint on the Laplacian matrix corresponding to the consistency graph enables the learned consistency graph to achieve neighborhood allocation of multi-view data. The number of connected components of the consistency graph is equal to the number of classification clusters of the original data. Each connected component corresponds to a classification cluster, and the clustering result is obtained from the consistency graph.

8. A multi-view consistent image clustering system based on embedded feature extraction, characterized in that: The system is used to implement the multi-view consistent image clustering method based on embedded feature extraction according to claim 1, comprising: A similarity matrix construction module is used to determine the original multi-view data matrix and construct a similarity matrix for each view based on sample elements in the multi-view data matrix; A sample embedding module is used to calculate the Laplacian matrix of each view according to the similarity matrix of each view, and obtain the sample embedding of each view based on the Laplacian matrix; A feature embedding module is used to perform singular value decomposition on the original multi-view data matrix to obtain the left singular matrix of each view, and construct the feature embedding of each view based on the left singular matrix of each view; The feature extraction module is used to perform group sparsity constraints on the feature embedding of each view and connect the sample embedding and feature embedding together through a bipartite graph for feature extraction; The consistency graph learning module is based on the sample embedding of each view to form a block diagonal matrix, and the block diagonal matrix is ​​used for consistency graph learning through the cost function, and then the consistency graph is rank constrained; The clustering module integrates feature extraction and consistency graph learning into a unified learning framework, obtains and solves the objective function, obtains the consistency graph, determines the connected components from the consistency graph, and obtains the image clustering results.

Images