End-to-end anchor point multi-view image clustering method under guidance of global principal component

Abstract

The invention discloses an end-to-end anchor point multi-view image clustering method under the guidance of global principal components, and particularly relates to the field of image processing. Comprising the following steps: acquiring a sub-view sample matrix of each view, and determining a feature mapping matrix; obtaining a global anchor point matrix and an anchor point diagram matrix; obtaining an anchor point fuzzy label matrix in the anchor point diagram matrix, and obtaining a sample fuzzy label matrix based on the anchor point fuzzy label matrix; establishing a target function based on the sub-view sample matrix, the feature mapping matrix, the global anchor point matrix, the anchor point diagram matrix, the anchor point fuzzy label matrix and the sample fuzzy label matrix; the feature mapping matrix is determined again, the sub-view sample matrixes of all the views are converted into the unified potential space again through the feature mapping matrix until the target function converges, and a final sample fuzzy label matrix is output. Unification of objective functions can be realized, so that model optimization is facilitated to obtain a globally optimal solution, and clustering precision is improved.

Description

Technical Field

[0001] The present application relates to a field, and in particular to an end-to-end anchor multi-view image clustering method guided by a global principal component. Background Art

[0002] The development of multimedia technology can effectively expand the source of sample features, thereby providing image analysis with more three-dimensional perspectives and richer details. As an important research topic in the field of multimedia data mining, multi-view clustering tasks aim to extract consistent features under each sub-view and complete data label learning based on the cluster distribution of potential public representations. Generally speaking, the classic multi-view clustering algorithm consists of three independent modules: sub-view feature extraction, public representation fusion, and discrete label learning. Specifically, it includes: strengthening the diversity of smooth multi-view subspace representations to standardize the common representation of views, and further parsing the low-rank correlation structure of consensus features to complete the clustering task; introducing anchor points to solve the problems of high spatial complexity and slow calculation speed of traditional multi-view clustering algorithms, and adaptively learning anchor weights to enhance the quality of anchor points; combining dictionary learning to incorporate intra-view self-representation updates and cross-view graph structure mining into a unified framework, and post-processing the unified graph structure to extract the final sample labels. Although these modular multi-view clustering algorithms can effectively extract global consensus features, the sub-view feature extraction, public representation fusion, and discrete label learning in the above methods are independent of each other, making it difficult to obtain the global optimal solution for objective function optimization, which leads to a decrease in clustering accuracy. Summary of the invention

[0003] The main purpose of this application is to provide an end-to-end anchor multi-view image clustering method guided by a global principal component, aiming to solve the problem of low clustering accuracy of existing clustering methods.

[0004] To achieve the above-mentioned purpose, the present application provides an end-to-end anchor multi-view image clustering method under the guidance of global principal components, comprising: step 1, obtaining a sub-view sample matrix of each view, and randomly generating an initial global anchor matrix and an initial anchor map matrix; step 2, determining a feature mapping matrix according to the sub-view sample matrix of each view, the initial global anchor matrix and the initial anchor map matrix; step 3, converting the sub-view sample matrices of all views to a unified latent space through the feature mapping matrix; step 4, extracting common anchor points from the sub-view sample matrices of all views in the unified latent space, and obtaining a global anchor matrix and an anchor map matrix; step 5, obtaining an anchor fuzzy label matrix in the anchor map matrix, And based on the anchor fuzzy label matrix, fuzzy label propagation is performed to obtain the sample fuzzy label matrix; step 6, based on the sub-view sample matrix, feature mapping matrix, global anchor matrix, anchor graph matrix, determine the structure loss and structure loss weight of each view; based on the anchor graph matrix, anchor fuzzy label matrix and sample fuzzy label matrix, determine the label transfer loss and transfer loss weight; establish the objective function according to the structure loss and structure loss weight of each view, label transfer loss and transfer loss weight; step 7, determine the feature mapping matrix according to the sub-view sample matrix, global anchor matrix and anchor graph matrix of each view, and return to step 3 until the objective function converges, and output the final sample fuzzy label matrix.

[0005] Optionally, the objective function is as follows:

[0006]

[0007]

[0008] In the formula, represents the subview sample matrix, ||·|| F represents the norm of matrix F, represents the feature mapping matrix under the vth subview, t represents the number of dimensions of the unified latent space, represents the global anchor matrix, m represents the number of global anchor points, represents the anchor graph matrix, and denotes the identity matrix, and Both represent column vectors whose elements are all 1, The column vectors of the feature map matrix satisfy the orthogonal constraint, Z T Z=I m Indicates that the column elements of the global anchor matrix satisfy the orthogonal constraint; represents the sample fuzzy label matrix, represents the anchor point fuzzy label matrix, F ≥ 0, F T 1 c =1 nIndicates that any element in the sample fuzzy label matrix satisfies the non-negative constraint and the sum of the column vector elements is 1, G ≥ 0, G T 1 c =1 m Indicates that any element in the anchor point fuzzy label matrix satisfies the non-negative constraint and the sum of the column vector elements is 1; α v represents the loss weight of the v-th view, and β represents the loss weight of label propagation.

[0009] Optionally, the feature map matrix in each view is updated as follows:

[0010]

[0011]

[0012] Where W v The optimal solution is expressed as W v =U v V v T , U v and V v Respectively represent X v B T Z T Perform singular value decomposition to obtain the left and right singular values.

[0013] Optionally, the global anchor matrix is ​​obtained by:

[0014]

[0015] Among them, the optimal solution of Z is expressed as Z = UV T , U and V represent Perform singular value decomposition to obtain the left and right singular values.

[0016] Optionally, the i-th column vector of the anchor map matrix is ​​obtained by the following formula:

[0017]

[0018] Among them, b i represents the i-th column vector in the anchor graph matrix, h i Representation Matrix The i-th column vector in .

[0019] Optionally, the anchor fuzzy label matrix is ​​obtained from the anchor map matrix according to the following formula:

[0020]

[0021] In the formula, Q represents the intermediate variable, g jrepresents the jth column vector in the anchor point fuzzy label matrix, η i Representation Matrix The j-th column vector in .

[0022] Optionally, the sample fuzzy label matrix is ​​obtained by the following formula:

[0023]

[0024] In the formula, f i represents the i-th column vector in the sample fuzzy label matrix, ξ i Representation Matrix The i-th column vector in , P is the intermediate variable, μ represents the regularization coefficient, Σ F represents the Lagrange multiplier matrix.

[0025] Optionally, the formula for the structural loss weight of each view is:

[0026]

[0027] The formula for transferring the loss weight is:

[0028]

[0029] In the formula, Represents the subview sample matrix.

[0030] Optionally, the basis for the objective function to converge is that the objective function reduction value is less than a convergence threshold.

[0031] Optionally, the reduced value of the objective function is determined according to the difference between the objective function value after the current round of iterative optimization and the objective function value after the previous round of iterative optimization.

[0032] Compared with the prior art, the beneficial effects of this application are as follows:

[0033] The end-to-end anchor multi-view image clustering method under the guidance of global principal components of the present invention converts samples under each view to a unified latent space through a feature mapping matrix, and reduces the information loss caused by feature mapping by minimizing the mapping loss function; incorporates sub-view feature extraction and public representation into a unified framework, alternately and collaboratively completes public anchor learning under the guidance of principal components and fuzzy label propagation based on topological information, realizes the unification of objective functions so as to facilitate model optimization to learn the global optimal solution and improve clustering accuracy, and completes end-to-end learning from image input to label output; performs fuzzy label propagation based on the topological information of the anchor graph matrix, and directly learns anchor label information and sample label information from the topological features of the anchor graph, without the need for post-processing to extract sample labels. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a flowchart of an end-to-end anchor multi-view image clustering method guided by global principal components in this application;

[0035] Figure 2 This is a flowchart of Embodiment 1 of the end-to-end anchor multi-view image clustering method guided by global principal components of the present application;

[0036] Figure 3 This is a result diagram of Example 1 of the end-to-end anchor point multi-view image clustering method guided by global principal components of the present application.

[0037] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solutions and advantages of this application clearer, the technical solutions in this application will be clearly and completely described below in conjunction with the drawings in this application. Obviously, the described embodiments are part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0039] The present invention provides an end-to-end anchor multi-view image clustering method guided by global principal components. Figure 1 As shown, the specific steps include:

[0040] Step S1, obtaining the sub-view sample matrix of each view, randomly generating an initial feature mapping matrix, an initial global anchor matrix, an initial anchor map matrix, an initial anchor fuzzy label matrix, and an initial sample fuzzy label matrix;

[0041] In this embodiment, the initial feature mapping matrix and the initial global anchor matrix under each subview can satisfy the orthogonal constraint by randomly generating column vectors, and the number of dimensions respectively satisfies and The matrix is ​​completed. The initial anchor map matrix, the initial sample fuzzy label matrix and the initial anchor fuzzy label matrix all satisfy the probability constraint. Their initialization can be achieved by randomly generating a matrix with column vectors that satisfy the non-negative constraint and the row sum is 1. The matrix dimensions satisfy and

[0042] Step S2, determining a feature mapping matrix according to the sub-view sample matrix of each view, the initial global anchor matrix and the initial anchor map matrix;

[0043] Specifically, the feature mapping matrix under each subview is determined as follows:

[0044]

[0045] in, represents the feature mapping matrix under the vth subview, t represents the number of dimensions of the unified latent space, and the constraint condition Indicates that the column vector of the feature mapping matrix satisfies the orthogonal constraint, that is, the unified potential space is an orthogonal space, and the feature mapping matrix W under different subviews v The updates are independent of each other, and the update function of the feature map matrix under the corresponding view can finally be written as follows:

[0046]

[0047] Among them, in the first iteration, W v is the initial feature mapping matrix, and W can be obtained by the above formula v The optimal solution can be expressed as W v =U v V v T AndU v and V v Respectively represent X v B T Z T The left and right singular values ​​are obtained by performing singular value decomposition. Traversing each view and calculating the optimal solution of the feature mapping matrix, the feature mapping matrix under each subview can be completed. Updates.

[0048] Step S3, converting the sub-view sample matrices of all views into a unified latent space through a feature mapping matrix;

[0049] Step S4, extracting common anchor points from the sub-view sample matrices of all views in the unified latent space to obtain a global anchor matrix and an anchor map matrix; the global anchor matrix is ​​obtained by the following formula:

[0050]

[0051] Among them, in the first iteration, Z is the initial global anchor matrix, B is the initial anchor map matrix, and Z T Z=I m Indicates that the column elements of the global anchor matrix satisfy the orthogonal constraint, that is, the anchor points are independent of each other, α v represents the loss weight of the v-th view, and the optimal solution of Z is expressed as Z = UV T , U and V represent Perform singular value decomposition to obtain the left and right singular values.

[0052] The anchor map matrix is ​​obtained by the following formula:

[0053]

[0054] Among them, in the first iteration, B is the initial anchor graph matrix, G and F are the initial anchor fuzzy label matrix and the initial sample fuzzy label matrix respectively; B ≥ 0 means that all elements in the anchor graph matrix satisfy the non-negative constraint. denotes the identity matrix, and Both represent column vectors whose elements are all 1, β represents the loss weight of label propagation, and b i represents the i-th column vector in the anchor graph matrix, h i Representation Matrix The i-th column vector in . Since the column vectors of the anchor graph matrix are independent of each other, the updates of the anchor features are independent of each other. Therefore, the update of the anchor graph can be completed by traversing and updating the column vectors. Taking the update of the i-th column vector as an example, its update function can be written as follows:

[0055]

[0056] This optimization problem can be solved by Lagrange multiplier method according to KKT constraints. Therefore, the update of B can be completed by traversing and updating n column vectors in the unified anchor graph matrix. The specific solution is as follows.

[0057] When the Lagrange multiplier method is used according to the KKT constraints, the above optimization problem can be converted into the following form:

[0058]

[0059] Among them, the vector γ ≥ 0 and the parameter λ are Lagrange multipliers, and there is a constraint b i ≥0 and γ T b i = 0. For vector b i Taking the derivative we get:

[0060]

[0061] b i =h i +λ1 m +γSubstitution We can get:

[0062]

[0063] Will We can get:

[0064]

[0065] Among them, the provisions as well as Can further obtain as well as Due to b i ,γ≥0 and γ T b i = 0, for any k (1≤k≤m), γ k b ik =0, that is, existence and and(·) + =max(x,0). At this time b ik The optimal solution depends on The value of , and The optimal value solution of can be equivalent to the solution of the zero point of the following function:

[0066]

[0067] in, And the function problem is a linear piecewise convex problem, and Newton iteration method can be used to obtain The optimal solution and iterative formula can be written as:

[0068]

[0069] Since then, The optimal solution of can be obtained through iterative calculation, and when The function value does not change more than 1×10 -2 When , the iterative calculation stops. After the optimal value is obtained, Finally, the optimized unified bipartite anchor graph matrix B is obtained.

[0070] Step S4, obtaining an anchor point fuzzy label matrix in the anchor point graph matrix, and performing fuzzy label propagation based on the anchor point fuzzy label matrix to obtain a sample fuzzy label matrix;

[0071] The anchor point fuzzy label matrix G is obtained as follows:

[0072]

[0073] In the formula, at the first iteration, G and F are the initial anchor point fuzzy label matrix and the initial sample fuzzy label matrix respectively; represents the anchor point fuzzy label matrix and the column vector represents the membership probability of the anchor point and each data cluster, G≥0,G T 1 c =1 mIndicates that any element in the anchor point fuzzy label matrix satisfies the non-negative constraint and the sum of the column vector elements is 1, that is, the probability that the anchor point corresponding to any column vector belongs to each data cluster is 1. To facilitate optimization, an intermediate variable Q is introduced and its matrix elements are made to approximate the anchor point fuzzy label matrix G. Combined with the Lagrange multiplier method to alternately optimize the above formula, the intermediate variable And the above optimization function can be further rewritten as:

[0074]

[0075] Among them, the variables to be optimized include the regularization coefficient υ, the Lagrange multiplier matrix Σ G , intermediate variable Q and anchor point fuzzy label matrix G. Specifically, in the present invention, the value of the regularization coefficient υ is uniformly initialized to 1×10 -2 And update according to υ=υ×1.2; Lagrange multiplier matrix Σ G Initialize the matrix uniformly to all zeros and follow Σ G =υ(GQ) to update; the update basis of the intermediate variables Complete. For the intermediate variable Q, the Lagrange multiplier matrix Σ G After updating the regularization coefficient υ, the optimization function of the anchor point fuzzy label matrix can be finally written as:

[0076]

[0077] Among them, g j represents the jth column vector in the anchor point fuzzy label matrix, η i Representation Matrix Since the updates of anchor labels are independent of each other, the optimization of the anchor fuzzy label matrix can be obtained by traversing m column vectors. Taking the update of the jth column vector as an example, its objective function can be written as follows:

[0078]

[0079] The above objective function can be solved by Lagrange multiplier method according to KKT constraint conditions. Therefore, the anchor point fuzzy label matrix can be obtained by traversing and updating the m column vectors in the anchor point fuzzy label matrix.

[0080] The sample fuzzy label matrix F is obtained by the following formula:

[0081]

[0082] In the formula, at the first iteration, G and F are the initial anchor point fuzzy label matrix and the initial sample fuzzy label matrix respectively; the condition F ≥ 0, F T 1 c =1 nIt means that any element in the sample fuzzy label matrix satisfies the non-negative constraint and the sum of the column vector elements is 1, that is, the sum of the probabilities that the samples corresponding to any column vector belong to each data cluster is 1. represents the sample fuzzy label matrix and the column vector represents the membership probability of the corresponding sample and each data cluster. To facilitate optimization, this embodiment introduces an intermediate variable P and makes its matrix elements approximate the sample fuzzy label matrix F. Combined with the Lagrange multiplier method to alternately optimize the above formula, the intermediate variable And the above optimization function can be further rewritten as:

[0083]

[0084] Among them, the variables to be optimized include the regularization coefficient μ, the Lagrange multiplier matrix Σ F , intermediate variable P and sample fuzzy label matrix F. Specifically, in the present invention, the regularization coefficient μ is uniformly initialized to 1×10 -2 And update according to μ = μ × 1.2; Lagrange multiplier matrix Σ F Initialize the matrix uniformly to all zeros and follow Σ F =μ(FP) to update; the update basis of the intermediate variables Completed. For the intermediate variable P, the Lagrange multiplier matrix Σ F After updating the regularization coefficient μ, the optimization function of the sample fuzzy label matrix can be finally written as:

[0085]

[0086] Among them, f i represents the i-th column vector in the sample fuzzy label matrix, ξ i Representation Matrix The i-th column vector in . Since the updates of sample labels are independent of each other, the optimization of the sample fuzzy label matrix can be obtained by traversing n column vectors. Taking the update of the i-th column vector as an example, its objective function can be written as follows:

[0087]

[0088] stf i ≥0,f i T 1 c =1

[0089] For the above objective function, a closed-form solution can be obtained by Lagrange multiplier method according to KKT constraint conditions. Therefore, the update of F can be completed by traversing and updating n column vectors in the sample fuzzy label matrix.

[0090] Step S5, based on the sub-view sample matrix, feature mapping matrix, global anchor matrix, and anchor graph matrix obtained above, the structural loss and structural loss weight of each view are determined; the formula of the structural loss weight of each view is:

[0091]

[0092] Step S6, determining the label transfer loss and transfer loss weight based on the anchor graph matrix, the anchor fuzzy label matrix and the sample fuzzy label matrix; the formula of the transfer loss weight is:

[0093]

[0094] Step S7, establishing an objective function according to the structural loss weight of each view, the structural loss of all views, the label transfer loss and the transfer loss weight;

[0095]

[0096]

[0097] In the formula, represents the subview sample matrix, ||·|| F represents the norm of matrix F, represents the feature mapping matrix under the vth subview, t represents the number of dimensions of the unified latent space, represents the global anchor matrix, m represents the number of global anchors, and the column vector of the global anchor matrix represents the corresponding anchor features. Represents the anchor graph matrix, which is used to describe the correlation between samples and anchor points in the unified latent space. denotes the identity matrix, and Both represent column vectors whose elements are all 1, The column vectors of the feature map matrix satisfy the orthogonal constraint, Z T Z=I m Indicates that the column elements of the global anchor matrix satisfy the orthogonal constraint; represents the sample fuzzy label matrix, represents the anchor point fuzzy label matrix, F ≥ 0, F T 1 c =1 n Indicates that any element in the sample fuzzy label matrix satisfies the non-negative constraint and the sum of the column vector elements is 1, that is, the sum of the probabilities that the samples corresponding to any column vector belong to each data cluster is 1; G≥0,G T 1 c =1 m Indicates that any element in the anchor point fuzzy label matrix satisfies the non-negative constraint and the sum of the column vector elements is 1; α vrepresents the loss weight of the v-th view, and β represents the loss weight of label propagation.

[0098] Step S8, using the sub-view sample matrix of each view, the global anchor matrix and the anchor map matrix obtained in step S4, the feature mapping matrix is ​​updated; and returning to step S3 until the objective function converges, outputting the sample fuzzy label matrix.

[0099] Specifically, the basis for the convergence of the objective function is that the reduction value of the objective function is less than the convergence threshold. The reduction value of the objective function is determined according to the difference between the objective function value after the current round of iterative optimization and the objective function value after the previous round of iterative optimization, which can be expressed as:

[0100] loss i =|obj i -obj i-1 |

[0101] Among them, loss i represents the reduction value of the objective function corresponding to the i-th round of iterative optimization, obj i represents the objective function value after the i-th round of iterative optimization, ||·|| F Represents absolute value operation.

[0102] Example 1

[0103] This embodiment uses the multi-view image dataset Mfeat as a clustering example. The number of samples is 2000, including 6 sub-view data, and the number of feature dimensions of each view is 216, 76, 64, 6, 240 and 47 respectively, and the number of image label types is 10. The convergence threshold is set to 1×10 -2 The specific clustering process is as follows: Figure 2 As shown in the figure, according to the final output sample fuzzy label matrix, the samples are assigned to the corresponding data clusters according to the maximum membership probability to obtain the predicted label Y pre By comparing the true label Y and the predicted label Y pre , the results are as follows Figure 3 As shown in the figure, the objective function decreases steadily with the increase of iteration number and converges at the 25th iteration. In the Mfeat dataset, the clustering accuracy (ACC) reaches 86.50%, the clustering normalized mutual information (NMI) reaches 87.78%, and the adjustable Rand coefficient (ARI) reaches 81.80%. It can be seen that the method of the present invention significantly improves the clustering accuracy of multi-view image data, can effectively extract consistent representation distribution in the latent space and accurately complete sample label learning.

[0104] The above are only preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A global principal component guided end-to-end anchor multi-view image clustering method, characterized in that: include: Step 1, obtain the sub-view sample matrix of each view, and randomly generate the initial global anchor matrix and the initial anchor map matrix; Step 2, determining a feature mapping matrix according to the sub-view sample matrix of each view, the initial global anchor matrix and the initial anchor map matrix; Step 3, convert the sub-view sample matrices of all views to a unified latent space through the feature mapping matrix; Step 4, extracting common anchor points from the sub-view sample matrices of all views in the unified latent space to obtain a global anchor matrix and an anchor map matrix; Step 5, obtaining an anchor point fuzzy label matrix in the anchor point graph matrix, and performing fuzzy label propagation based on the anchor point fuzzy label matrix to obtain a sample fuzzy label matrix; Step 6, determining the structural loss and structural loss weight of each view based on the sub-view sample matrix, the feature mapping matrix, the global anchor matrix, and the anchor graph matrix; Determine the label transfer loss and transfer loss weight based on the anchor graph matrix, the anchor fuzzy label matrix and the sample fuzzy label matrix; establish the objective function based on the structure loss and structure loss weight of each view, the label transfer loss and the transfer loss weight; Step 7, determine the feature mapping matrix according to the sub-view sample matrix of each view, the global anchor matrix, and the anchor map matrix, and return to step 3 until the objective function converges, and output the final sample fuzzy label matrix.

2. The end-to-end anchor multi-view image clustering method guided by global principal components according to claim 1, characterized in that: The objective function is as follows: In the formula, represents the subview sample matrix, ||·|| F represents the norm of matrix F, represents the feature mapping matrix under the vth subview, t represents the number of dimensions of the unified latent space, represents the global anchor matrix, m represents the number of global anchor points, represents the anchor graph matrix, and denotes the identity matrix, and Both represent column vectors whose elements are all 1, W v T W v =I t The column vectors of the feature map matrix satisfy the orthogonal constraint, Z T Z=I m Indicates that the column elements of the global anchor matrix satisfy the orthogonal constraint; represents the sample fuzzy label matrix, represents the anchor point fuzzy label matrix, F ≥ 0, F T 1 c =1 n Indicates that any element in the sample fuzzy label matrix satisfies the non-negative constraint and the sum of the column vector elements is 1, G ≥ 0, G T 1 c =1 m Indicates that any element in the anchor point fuzzy label matrix satisfies the non-negative constraint and the sum of the column vector elements is 1; α v represents the loss weight of the v-th view, and β represents the loss weight of label propagation.

3. The end-to-end anchor multi-view image clustering method guided by global principal components according to claim 2, characterized in that: The update method of the feature mapping matrix under each view is: Where W v The optimal solution is expressed as W v =U v V v T , U v and V v Respectively represent X v B T Z T Perform singular value decomposition to obtain the left and right singular values.

4. The end-to-end anchor multi-view image clustering method guided by global principal components according to claim 2, characterized in that: The global anchor matrix is ​​obtained by the following formula: Among them, the optimal solution of Z is expressed as Z = UV T , U and V represent Perform singular value decomposition to obtain the left and right singular values.

5. The end-to-end anchor multi-view image clustering method guided by global principal components according to claim 2, characterized in that: The i-th column vector of the anchor graph matrix is ​​obtained by the following formula: Among them, b i represents the i-th column vector in the anchor graph matrix, h i Representation Matrix The i-th column vector in .

6. The end-to-end anchor multi-view image clustering method guided by global principal components according to claim 2, characterized in that: The anchor fuzzy label matrix is ​​obtained in the anchor map matrix according to the following formula: In the formula, Q represents the intermediate variable, g j represents the jth column vector in the anchor point fuzzy label matrix, η i Representation Matrix The j-th column vector in .

7. The end-to-end anchor multi-view image clustering method guided by global principal components according to claim 2, characterized in that: The sample fuzzy label matrix is ​​obtained by the following formula: In the formula, f i represents the i-th column vector in the sample fuzzy label matrix, ξ i Representation Matrix The i-th column vector in , P is the intermediate variable, μ represents the regularization coefficient, Σ F represents the Lagrange multiplier matrix.

8. The end-to-end anchor multi-view image clustering method guided by global principal components according to claim 2, characterized in that: The formula for the structural loss weight of each view is: The formula for transferring the loss weight is: In the formula, Represents the subview sample matrix.

9. The end-to-end anchor multi-view image clustering method guided by global principal components according to claim 1, characterized in that: The basis for the objective function to converge is that the reduction value of the objective function is less than the convergence threshold.

10. The end-to-end anchor multi-view image clustering method guided by global principal components according to claim 9, characterized in that: The reduced value of the objective function is determined according to the difference between the objective function value after the current round of iterative optimization and the objective function value after the previous round of iterative optimization.

Images