Multi-view clustering method based on adaptive weighted tensor

Abstract

The invention discloses a multi-view clustering method based on an adaptive weighted tensor, and belongs to the technical field of image clustering analysis. The method comprises the following steps: acquiring a multi-view data set, and preprocessing images in the multi-view data set; extracting features of different visual angles from each initial image; performing standardization processing on the initial view angle features; constructing a target function by taking the data reconstruction error and the minimum low-rank constraint as targets, converting the target function into an augmented Lagrange multiplier method form, and solving a Lagrange function through a multiplication and division method in an alternating direction; and performing feature transformation on the to-be-processed image data through the optimal projection matrix, and executing clustering operation in the fusion feature space to obtain a clustered image. The method does not need to depend on label information, can achieve a robust clustering effect under the conditions of noise interference and complex sample distribution through multi-view feature fusion and low-dimensional projection modeling, and is suitable for automatic classification and analysis requirements of actual multi-source heterogeneous image data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image clustering analysis, and particularly relates to a multi-view image clustering method based on adaptive weighted tensor. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, the demand for image data processing is rapidly growing in various fields. The high complexity, structural diversity of image data and the interference in the process of acquisition, storage and dissemination bring challenges to computer vision tasks. Especially multi-view image data contains information from different perspectives or modalities, with different features and structures. How to effectively fuse data from different perspectives has become a technical problem.

[0003] To improve the accuracy and efficiency of image analysis, image segmentation technology is widely used in the preprocessing stage to effectively separate the target region from the background. However, traditional image segmentation methods often face problems such as blurred boundaries, insufficient segmentation accuracy, and sensitivity to noise when dealing with multi-modal or multi-view images, making it difficult to provide stable and reliable feature basis for subsequent clustering and recognition.

[0004] After image segmentation, how to effectively fuse image features from different perspectives or modalities and further achieve accurate clustering is a key technical difficulty in current image understanding. Existing clustering methods such as KMeans and GMM usually assume that data comes from a single perspective and use a fixed clustering strategy, which cannot fully utilize the feature information of multi-view data. These methods have poor robustness to noise and outliers, resulting in insufficient clustering accuracy and stability.

[0005] Weighted clustering methods attempt to assign weights to different perspectives, but ignore the differences between perspectives. In multi-view data clustering, how to adaptively adjust the weight according to the data quality of each perspective is still a challenge. Graph-based clustering methods, such as spectral clustering, can handle the relationship between perspectives, but have defects in noise processing and high computational complexity, making it difficult to deal with large-scale data.

[0006] Low-rank and low-dimensional feature extraction methods have been proposed, which can remove noise, reduce dimensionality and learn projection matrices to improve feature extraction and clustering accuracy. However, current low-rank methods cannot handle the fusion of multi-view image data, and are inefficient in high-dimensional data. SUMMARY

[0007] The purpose of the present application is to provide a multi-view clustering method based on adaptive weighted tensor to achieve accurate preprocessing and robust clustering analysis of complex image scenes.

[0008] Technical solution: The multi-view clustering method based on adaptive weighted tensor of the present application comprises the following steps:

[0009] obtaining a multi-view data set including images of different views and different categories of scenes;

[0010] preprocessing the images in the multi-view data set to obtain initial images;

[0011] extracting features of different views from each initial image to obtain initial view features of each image;

[0012] standardizing the initial view features to obtain data matrices, and merging all the data matrices to obtain an augmented matrix;

[0013] constructing a target function with a data reconstruction error and a low-rank constraint minimization as a target, converting the target function into an augmented Lagrange multiplier form, solving the Lagrange function by an alternating direction multiplication method to obtain an optimal projection matrix;

[0014] performing feature transformation on the image data to be processed by the optimal projection matrix, and performing clustering operation in a fusion feature space to obtain clustered images.

[0015] Further, the step of extracting features of different views from each initial image to obtain initial view features of each image includes:

[0016] performing dimension reduction processing on the initial image by principal component analysis, and retaining the first n0 principal components, wherein n0 is a natural number, and the obtained image features are denoted as PCA features;

[0017] performing gray processing on the initial image to obtain a gray image, extracting local binary pattern features from the gray image, and denoting the local binary pattern features as LBP features;

[0018] extracting features of the gray image by using a histogram of oriented gradients to obtain histogram of oriented gradients features, and denoting the histogram of oriented gradients features as HOG features;

[0019] The PCA features, the LBP features and the HOG features are used to form the initial view features.

[0020] Further, the data matrices obtained after standardizing the initial view features are respectively denoted as X rgb ,X lbp ,X hog The augmented matrix obtained after merging the data matrices is denoted as X = [X rgb ,X lbp ,X hog ], wherein X rgb represents the PCA features after standardization, X lbp represents the LBP features after standardization, and X hog represents the HOG features after standardization.

[0021] Further, the expression of the objective function is:

[0022]

[0023] In the formula, the first term after the equal sign represents the data reconstruction error term, the second term represents the low-rank constraint term, and the third term represents the L2 norm constraint term; X v represents the normalized feature matrix of the vth view, W represents the projection matrix, and W v represents the projection matrix of the vth view, V represents the number of views, and λ and β represent the regularization parameters. represents the square of the Frobenius norm, which is used to represent the data reconstruction error; || || * represents the nuclear norm, i.e., the sum of singular values, which is used to realize the low-rank constraint. represents the square of the L2 norm, which is used to make a parameter regularization penalty.

[0024] Further, the step of converting the objective function into the form of the augmented Lagrange multiplier method includes:

[0025] For the low-rank constraint term and the L2 norm constraint term of each view, the Lagrange multipliers Y v and Z corresponding to the low-rank constraint term and the L2 norm constraint term are introduced, and a Lagrange function is constructed, and the formula is:

[0026]

[0027] In the formula, represents the initial value of the projection matrix W v of the vth view, is the initial value of the projection matrix W, and <,> represents the inner product between matrices.

[0028] Further, the step of solving the Lagrange function by the alternating direction multiplication-division method to obtain the optimal projection matrix includes:

[0029] Step 51, set the maximum number of iterations;

[0030] Step 52, fix the variable W, and iteratively update the projection matrix W v of the vth view, at this time, the Lagrange function is represented as:

[0031]

[0032] The first term of the Lagrange function, the reconstruction error term, is expanded, and the Lagrange function is represented as:

[0033]

[0034] where T denotes transpose, trace denotes the trace of a matrix, and diag denotes the sum of diagonal elements of a matrix;

[0035] The matrix W is decomposed into: v v = U∑V T , where U denotes a left singular vector matrix, ∑ denotes a diagonal matrix composed of singular values, and V denotes a right singular vector matrix; T

[0036] The diagonal matrix ∑ composed of singular values is thresholded, and the formula is ∑' = max(∑ - λ, 0);

[0037] The reconstructed variable is:

[0038] Step 53, fix the variable W v , and update the projection matrix W iteratively, and the Lagrange function is represented as:

[0039]

[0040] After expanding the Lagrange function, we get:

[0041]

[0042] where the first term after the equal sign is the reconstruction error term, which represents the reconstruction error when the augmented matrix X is mapped to the low-dimensional space through the projection matrix W; the second term is the L2 regularization term, which controls the size of the projection matrix W;

[0043] Solve the linear equation system by gradient, and we get:

[0044]

[0045] where I denotes the identity matrix;

[0046] Step 54, fix the variables W v and W, and update the Lagrange multiplier, and the formula is:

[0047]

[0048] Z * = min(<Z, W - W * >),

[0049] Step 55, set the convergence condition as:||f(W (t+1) ) - f(W (t) ) || F < ε, ​

[0050] In the formula, f(·) is a target function used to describe the optimization process of the model, t represents the number of iteration steps, Wt represents the projection matrix output in the tth iteration step, Wt is the projection matrix output in the tth iteration step, and W is a set threshold value. (t) In the formula, f(·) is a target function used to describe the optimization process of the model, t represents the number of iteration steps, Wt represents the projection matrix output in the tth iteration step, Wt is the projection matrix output in the tth iteration step, and W is a set threshold value.

[0051] Step 56, according to the set projection matrix, the iteration of steps 52 to 54 is carried out until the projection matrix obtained satisfies the convergence condition, that is, the optimal projection matrix.

[0052] Further, the step of pre-processing the images in the multi-view data set comprises:

[0053] All images are processed into a standard size, and then normalized.

[0054] Beneficial effects: compared with the prior art, the present application has the following advantages:

[0055] 1. Introducing image segmentation as a preprocessing step effectively improves the feature extraction quality: in the present application, by introducing the image segmentation operation in the image processing flow, the target area in the image can be accurately extracted, the background redundancy and interference information are reduced, and a clearer and more stable input basis is provided for subsequent multi-view feature extraction and clustering, thereby improving the overall analysis accuracy;

[0056] 2. The adaptive adjustment of the view difference can be realized based on the weighted tensor model: the present application designs an independent mapping matrix for each view, and dynamically adjusts the contribution weight of each view in feature fusion through the joint optimization of low-rank constraint and reconstruction error, thereby improving the accuracy and effectiveness of multi-view data fusion;

[0057] 3. It has the ability of sparse structure guidance, and enhances the discriminative expression of multi-view features: the present application introduces a sparse modeling mechanism in the optimization process to compress and select the feature mapping results, effectively filters redundant information, retains high correlation features, and improves the stability and generalization ability of the clustering model;

[0058] 4. Unsupervised learning mechanism is adopted to realize high-quality clustering, which is suitable for complex image scenes: the present application does not need to rely on label information, and through multi-view feature fusion and low-dimensional projection modeling, it can realize stable clustering effect under the condition of noise interference and complex sample distribution, and is suitable for automatic classification and analysis of actual multi-source heterogeneous image data. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 A flowchart of an image clustering method provided by the embodiment of the present application is shown in the figure.

[0060] Figure 2The figure for comparing clustering effects of different methods under Scene-15 dataset. DETAILED DESCRIPTION

[0061] The embodiments of the present application will be further described below in conjunction with the drawings and examples. It can be understood that the specific embodiments described herein are only used to explain the embodiments of the present application, but not limit the embodiments of the present application. In addition, it should be noted that, for the convenience of description, only the parts related to the embodiments of the present application are shown in the drawings, but not all the structures.

[0062] In the following description, specific details are set forth, such as target system structures, techniques, etc., in order to provide a thorough understanding of the embodiments of the present application. However, it should be apparent to those skilled in the art that the embodiments of the present application can be practiced without these specific details. In other instances, well-known structures, devices, circuits, and methods have been described in detail in order to avoid obscuring the description of the embodiments of the present application.

[0063] It should be understood that the term "comprising" as used in the specification and the appended claims indicates the presence of the recited features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0064] It should also be understood that the term "and / or" as used in the specification and the appended claims indicates any combination of one or more of the associated listed items and all possible combinations of the items.

[0065] In addition, in the description of the specification and the appended claims, the terms "first", "second", and the like are used only to distinguish descriptions, and cannot be understood as indicating or implying relative importance.

[0066] In the specification of the present application, the reference "one embodiment" or "some embodiments" and the like means that the target features, structures or characteristics described in conjunction with the embodiment are included in one or more embodiments of the present application. Therefore, the statements "in one embodiment", "in some embodiments", "in other some embodiments", "in further some embodiments" and the like appearing in different places in the specification are not necessarily all referring to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized.

[0067] The multi-view clustering method based on adaptive weighted tensor described in the embodiments of the present application combines the advantages of the above two methods, and has the following advantages: Figure 1 As shown in the figure, the method comprises the following steps:

[0068] Step 1, obtaining a multi-view dataset, wherein the multi-view dataset comprises images of different views and different categories of scenes.

[0069] Obtaining images under different views, such as images containing front, side or back views, and obtaining images under different categories of scenes, to form a multi-view dataset.

[0070] Step 2, pre-processing the images in the multi-view dataset to obtain initial images.

[0071] Further, the step of pre-processing the images in the multi-view dataset comprises:

[0072] All images are processed to a standard size and then normalized.

[0073] In an example, taking the Scene-15 dataset as an example, the dataset contains 15 different categories of scene images, each category contains multiple samples, all images are processed to a standard size such as 224x224 to ensure data consistency, and then normalized.

[0074] Step 3, extracting features of different views from each initial image to obtain initial view features of each image.

[0075] Further, the step of extracting features of different views from each initial image to obtain initial view features of each image comprises:

[0076] Performing dimensionality reduction on the initial image by principal component analysis, retaining the first n0 principal components, and recording the obtained image features as PCA features; wherein n0 is a natural number;

[0077] Performing grayscale processing on the initial image to obtain a grayscale image, extracting local binary pattern features from the grayscale image, and recording the local binary pattern features as LBP features;

[0078] Performing feature extraction on the grayscale image using a histogram of oriented gradients to obtain histogram of oriented gradients features, and recording the histogram of oriented gradients features as HOG features;

[0079] Using the PCA features, the LBP features and the HOG features to form the initial view features.

[0080] In one example, the RGB image is processed by principal component analysis (PCA) for dimension reduction, and the first 50 principal components are retained to obtain the PCA feature of the RGB image. A local binary pattern (LBP) feature is extracted from the grayscale image, and the LBP feature is obtained by using an 8-neighborhood and a radius of 1 setting. A histogram of oriented gradients (HOG) is used for feature extraction of the grayscale image, and the HOG feature is obtained by setting a cell to 8x8 pixels and a block size to 1x1. All images obtain initial view features after the above feature extraction, and the initial view features include features of three different views.

[0081] Step 4, the initial view features are standardized to obtain a data matrix, and all data matrices are merged to obtain an augmented matrix.

[0082] Further, the data matrix obtained after the initial view features are standardized is denoted as X rgb ,X lbp ,X hog The augmented matrix obtained after the data matrix is merged is denoted as X = [X rgb ,X lbp ,X hog ], wherein X rgb represents the PCA feature after standardization, X lbp represents the LBP feature after standardization, and X hog represents the HOG feature after standardization.

[0083] Step 5, a target function is constructed by taking the data reconstruction error and the low-rank constraint minimum as the target, the target function is converted into an augmented Lagrange multiplier method form, the Lagrange function is solved by an alternating direction multiplication-division method to obtain an optimal projection matrix.

[0084] Further, the expression of the target function is:

[0085]

[0086] In the formula, the first term after the equal sign represents a data reconstruction error term, the second term represents a low-rank constraint term, and the third term represents an L2 norm constraint term; X v represents a standardized feature matrix of the vth view, and is used to represent the augmented matrix X = [X rgb ,X lbp ,X hog ], W represents a projection matrix, W v represents a projection matrix of the vth view, V represents the number of views, and λ and β represent regularization parameters; represents the square of the Frobenius norm, and is used to represent the data reconstruction error;* represents the nuclear norm, i.e. the sum of singular values, for realizing the low-rank constraint; represents the square of L2 norm, used for parameter regular penalty.

[0087] Further, the step of converting the objective function into the form of augmented Lagrange multiplier method comprises:

[0088] For the low-rank constraint term and the L2 norm constraint term of each view, introduce Lagrange multipliers Y v and Z corresponding to the low-rank constraint term and the L2 norm constraint term, and construct a Lagrange function, the formula is:

[0089]

[0090] In the formula, represents the initial value of the projection matrix W v of the vth view, is the initial value of the projection matrix W, <,> represents the inner product between matrices.

[0091] Further, the step of solving the Lagrange function by alternating direction multiplication and division to obtain the optimal projection matrix comprises:

[0092] Step 51, set the maximum number of iterations;

[0093] Step 52, fix the variable W, and iteratively update the projection matrix W v of the vth view,

[0094]

[0095] The first term in the Lagrange function is the reconstruction error term, which is expanded, and the Lagrange function is represented as:

[0096]

[0097] In the formula, T represents transposition, trace represents the trace of a matrix, and represents the sum of diagonal elements of a matrix;

[0098] The matrix W v is singular value decomposed into: W v = U∑V T , in which U represents a left singular vector matrix, ∑ represents a diagonal matrix composed of singular values, and V T represents a right singular vector matrix;

[0099] The diagonal matrix ∑ composed of singular values is further thresholded, the formula is: ∑' = max(∑-λ, 0), after the processing, the part of the singular values less than λ is set to zero, thereby realizing the sparsification and low-rankization;

[0100] The reconstructed variables are:

[0101] Step 53, fix variable W v If the projection matrix W is updated iteratively, then the Lagrange function can be expressed as:

[0102]

[0103] Expanding the Lagrange function yields:

[0104]

[0105] In the formula, the first term after the equal sign is the reconstruction error term, which represents the reconstruction error when the augmented matrix X is mapped to the low-dimensional space through the projection matrix W; the second term is the L2 regularization term, which controls the size of the projection matrix W to avoid overfitting.

[0106] By solving the linear equations using gradient descent, we can obtain:

[0107]

[0108] In the formula, I represents the identity matrix;

[0109] Step 54, fix variable W v And W, update the Lagrange multipliers, the formula is:

[0110]

[0111] Z * =min( <Z,W-W * >),

[0112] Step 55, set the convergence condition as: ||f(W) (t+1) )-f(W (t) )|| F <∈,

[0113] In the formula, f(·) is the objective function, which describes the optimization process of the model; t represents the iteration step; represents the projection matrix output in the t-th iteration; and W represents the projection matrix output in the t-th iteration. (t) It is the projection matrix output at step t, where a threshold is set.

[0114] Step 56: Iterate through steps 52 to 54 according to the set projection matrix until the calculated projection matrix satisfies the convergence condition, which is the optimal projection matrix.

[0115] Step 6: Perform feature transformation on the image data to be processed using the optimal projection matrix, and perform clustering operation in the fused feature space to obtain the clustered image.

[0116] The original image data is mapped to a low-dimensional feature space by feature transformation of the image data through an optimal projection matrix. Then the features from different perspectives are fused by a weighted tensor to form a unified feature representation. A clustering algorithm (such as KMeans) is applied in the fused feature space to cluster the images, thereby obtaining the final clustering result.

[0117] According to the set projection matrix, iteration is performed until the optimal projection matrix is obtained. When the obtained projection matrix meets the convergence condition, it is the optimal projection matrix. After the extracted features (PCA, LBP, HOG) of each perspective are standardized, they are spliced by column to generate an augmented matrix. The augmented matrix is multiplied by the optimal projection matrix to obtain a feature matrix Y, which contains the low-dimensional features of the image. The feature matrix Y is subjected to a clustering operation, and after the operation is completed, each image is assigned a class label, thereby realizing multi-perspective clustering of the images.

[0118] In this embodiment, the Scene-15 dataset is used to test the multi-perspective clustering algorithm of the present application, and the algorithm is compared with other classical clustering methods. The Scene-15 dataset contains 15 different categories of scene images, each category contains multiple samples, and all images are adjusted to a standard size and normalized in the preprocessing stage. Then, noise is added to the training data to simulate the interference in the real environment, and the robustness of the algorithm under noise interference is tested.

[0119] During the addition of noise, for each image in the dataset, salt and pepper noise is randomly selected to be added, the noise ratio ranges from 2% to 20% of the image pixels, and the noise is randomly distributed. 50% of the images are used as training data, and the remaining 50% of the images are used as test data to evaluate the performance of the algorithm.

[0120] Table 1 Clustering effect under Scene-15 dataset

[0121] Method ACC NMI Purity KMeans 0.2080 0.1582 0.7900 GMM 0.2080 0.1582 0.7900 Hierarchical 0.2380 0.2021 0.8040 Spectral 0.2660 0.1712 0.7900 DBSCAN 0.5140 0.1100 0.5680 Mean Shift 0.5180 0.0000 0.5180 BIRCH 0.2500 0.1597 0.7920 Co-Reg 0.8520 0.4012 0.8520 The invention 0.8880 0.5208 0.8880

[0122] As shown in Table 1, in this experiment, under the Scene-15 dataset, the traditional single-view clustering methods (KMeans, GMM, hierarchical, spectral, BIRCH) can only use single features, resulting in low accuracy ACC and information consistency NMI (ACC ≈ 0.20-0.27, NMI ≈ 0.15-0.20), while the density clustering (DBSCAN, Mean Shift) has improved the accuracy (ACC ≈ 0.51), but at the expense of purity (Purity) and NMI; in contrast, the multi-view collaborative Co-Reg method greatly improves the clustering effect by mutual regularization of different view information (ACC ≈ 0.85, NMI ≈ 0.40, Purity ≈ 0.85), and the present application most fully integrates the three-view features under adaptive weighting and low-rank sparse constraint, achieving the highest ACC (0.8880), NMI (0.5208) and Purity (0.8880), verifying its significant advantages in multi-source feature fusion and robust clustering.

[0123] Figure 2 The clustering results of eight clustering algorithms in the same PCA dimension reduction space are shown, and the present application (AWTLRC) can clearly separate the two types of samples closely and cleanly, with almost no misclassification; KMeans, GMM and hierarchical clustering have more boundary overlaps; Spectral can partially identify non-spherical structures but still have confusion; DBSCAN and Mean Shift either identify sparse points as noise or over-subdivide clusters; BIRCH performs well in the center area but has obvious misclassification at the edge, and overall only the method described in the present application is closest to the real class distribution after clustering. Figure 2 ​

Claims

1. A multi-view clustering method based on adaptive weighted tensors, characterized in that, The method comprises the following steps: Obtaining a multi-view data set comprising images of different views and different categories of scenes; Preprocessing the images in the multi-view data set to obtain initial images; Extracting features of different views from each initial image to obtain initial view features of each image; Standardizing the initial view features to obtain a data matrix, merging all the data matrices to obtain an augmented matrix; Constructing a target function with a data reconstruction error and a low-rank constraint minimization as a target, converting the target function into an augmented Lagrange multiplier method form, solving the Lagrange function by an alternating direction multiplication-division method to obtain an optimal projection matrix; Performing feature transformation on the image data to be processed by the optimal projection matrix and performing clustering operation in a fusion feature space to obtain clustered images.

2. The adaptive weighted tensor based multi-view clustering method of claim 1, wherein, The step of extracting features of different views from each initial image to obtain initial view features of each image comprises: Performing dimension reduction processing on the initial image by principal component analysis to retain the first n0 principal components, and the obtained image features are denoted as PCA features; wherein n0 is a natural number; Performing grayscale processing on the initial image to obtain a grayscale image, extracting a local binary pattern feature from the grayscale image, and denoted as an LBP feature; Performing feature extraction on the grayscale image by using a histogram of oriented gradients to obtain a histogram of oriented gradients feature, denoted as an HOG feature; The initial view features are composed of the PCA features, the LBP features and the HOG features.

3. The adaptive weighted tensor based multi-view clustering method of claim 2, wherein, The data matrix obtained after normalizing the initial view angle features is denoted as X rgb , lbp , hog The augmented matrix obtained after merging the data matrices is denoted as X = [X rgb , lbp , hog ], wherein X rgb represents the PCA features after normalization, X lbp represents the LBP features after normalization, and X hog represents the HOG features after normalization.

4. The adaptive weighted tensor based multi-view clustering method of claim 3, wherein, The expression of the target function is: In the formula, the first term represents the data reconstruction error term, the second term represents the low rank constraint term, and the third term represents the L2 norm constraint term; X v represents the normalized feature matrix of the vth view, W represents the projection matrix, and W v represents the projection matrix of the vth view, V represents the number of views, and λ and β represent the regularization parameters; represents the square of the Frobenius norm, which is used to represent the data reconstruction error; || || * represents the nuclear norm, i.e., the sum of singular values, which is used to realize the low rank constraint; represents the square of the L2 norm, which is used to make parameter regularization punishment.

5. The adaptive weighted tensor based multi-view clustering method of claim 4, wherein, The step of converting the target function into an augmented Lagrange multiplier method form comprises: The low-rank constraint term and the L2 norm constraint term for each view angle introduce Lagrange multipliers Y v and Z correspond to the low-rank constraint term and the L2 norm constraint term, and a Lagrange function is constructed, and the formula is: wherein denotes the v-th view projection matrix W v an initial value of the projection matrix W, is an initial value of the projection matrix W, <,> denotes the inner product between matrices.

6. The adaptive weighted tensor based multi-view clustering method of claim 5, wherein, The step of solving the Lagrange function by an alternating direction multiplication-division method to obtain an optimal projection matrix comprises: Step 51, setting a maximum number of iterations; Step 52, fix variable W, iteratively update the vth view projection matrix W v The Lagrange function at this time is represented as: Expanding the first reconstruction error term in the Lagrange function, the Lagrange function is expressed as: In the formula, T represents transposition, trace represents the trace of a matrix, and represents the sum of diagonal elements of a matrix; The matrix W v is singular value decomposed into W v = U∑V T , where U is a left singular vector matrix, ∑ is a diagonal matrix of singular values, and V T is a right singular vector matrix. Threshold processing is further performed on the diagonal matrix Σ composed of singular values, and the formula is Σ'=max(Σ-λ,0); The reconstructed variable is then: Step 53, fix variable W v , update projection matrix W iteratively, then the Lagrange function is represented as: After expanding the Lagrange function, the following is obtained: In the formula, the first term after the equal sign is the reconstruction error term, which represents the reconstruction error when the augmented matrix X is mapped to a low-dimensional space by the projection matrix W; the second term is an L2 regularization term, which controls the size of the projection matrix W; Gradient solving of a linear equation set can be performed to obtain: In the formula, I represents an identity matrix; Step 54, fix variable W v and W, update Lagrange multiplier, formula is: Z * = min(<Z, W - W * >), Step 55, set the convergence condition as: ||f(W (t+1) )-f(W (t) )|| <∈ F <∈, In the formula, f(·) is a target function used to describe the optimization process of the model, t represents the number of iteration steps, Wt represents the projection matrix output in the tth iteration step, W (t) is the projection matrix output in the tth iteration step, and is a set threshold value. Step 56, according to the set projection matrix, performing iteration by steps 52 to 54 until the obtained projection matrix meets the convergence condition, and the optimal projection matrix is obtained.

7. The adaptive weighted tensor based multi-view clustering method according to any one of claims 1 to 6, characterized in that, The step of preprocessing the images in the multi-view data set comprises: All images are processed to a standard size and then normalized.

Images