Identity recognition method and system based on depth spectral clustering and computer storage medium

Abstract

The invention discloses an identity recognition method and system based on depth spectral clustering and a computer storage medium, and relates to the technical field of identity recognition, and the method comprises the steps: obtaining the identity data of a user; extracting features in the identity data by using a deep network model to obtain identity features; the identity features are clustered through a spectral clustering model, the category to which the identity features belong is obtained, the spectral clustering model is established based on a sparse KAN network, norm regularization and entropy regularization constraints are introduced into the sparse KAN network, and meanwhile an orthogonal constraint mechanism is introduced into the last layer of the sparse KAN network. According to the method, norm regularization and entropy regularization constraint, orthogonalization constraint and manifold learning technologies are introduced, so that the identity recognition efficiency, stability and accuracy are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of identity recognition, and particularly relates to an identity recognition method and system based on deep spectral clustering and a computer storage medium. BACKGROUND

[0002] Identity recognition is to recognize the category of a user's identity, i.e., a legal user or an illegal user, through various data provided by the user, such as a face image, a fingerprint, voiceprint data, etc. At present, common identity recognition methods are mostly based on spectral clustering technology. Spectral clustering has been widely applied in various unsupervised learning tasks, including identity recognition, due to its powerful global feature capturing capability and flexible affinity measurement method. In recent years, the rise of deep learning has brought new research paradigms to unsupervised clustering. According to the implementation process of a deep learning algorithm, it can be divided into data embedding based on a neural network and loss function calculation based on spectral clustering.

[0003] However, the existing identity recognition method based on deep spectral clustering has the following defects: (1) The existing deep spectral clustering method adopts a fully connected layer to construct a network, which leads to an exponential increase in the number of parameters with the increase in network depth and input dimension, causing problems such as huge consumption of computing resources, risk of overfitting, etc., which directly leads to low processing efficiency of identity recognition.

[0004] (2) In the calculation of spectral clustering, the feature representation of data needs to be independent of each other in order to accurately reflect the similarity and difference between data points and thus achieve effective clustering. However, in the calculation of the loss function using the spectral clustering structure, deep learning cannot guarantee the independence between the obtained network outputs, leading to poor stability of the identity recognition result.

[0005] (3) The traditional spectral clustering method relies on a static Gaussian kernel to construct an affinity matrix, which cannot be dynamically adjusted according to specific task requirements and is easily disturbed by noise in the original data, leading to poor accuracy of the identity recognition result. SUMMARY

[0006] The embodiments of the present application provide an identity recognition method, system and computer storage medium based on deep spectral clustering, to solve the problems of low efficiency, poor stability and poor accuracy in identity recognition using the traditional spectral clustering technology in the prior art.

[0007] In one aspect, the embodiments of the present application provide an identity recognition method based on deep spectral clustering, comprising: obtaining identity data of a user; extracting features in the identity data using a deep network model to obtain identity features; The identity features are clustered by using a spectral clustering model to obtain a category to which the identity features belong, wherein the spectral clustering model is established based on a sparse KAN (Kolmogorov-Arnold network) network, norm regularization and entropy regularization constraints are introduced into the sparse KAN network, and an orthogonal constraint mechanism is also introduced into the last layer of the sparse KAN network; a loss function used by the sparse KAN network during training includes a regularization loss function and a spectral clustering loss function, the regularization loss function is established based on an activation function after the norm regularization and the entropy regularization constraints are introduced into the sparse KAN network, and the spectral clustering loss function is established based on an optimized affinity matrix; the optimization process of the affinity matrix includes: initializing the affinity matrix by using training samples, establishing an optimization objective function based on the initial affinity matrix, introducing a manifold learning technique into the optimization objective function, projecting the training samples into a low-dimensional space, and solving the optimization objective function to complete the optimization of the affinity matrix.

[0008] In another aspect, the embodiments of the present application also provide an identity recognition system based on deep spectral clustering, comprising: a data acquisition module configured to acquire identity data of a user; a deep network model configured to extract features in the identity data to obtain identity features; a spectral clustering model configured to cluster the identity features to obtain a category to which the identity features belong; The spectral clustering model is established based on a sparse KAN network, norm regularization and entropy regularization constraints are introduced into the sparse KAN network, and an orthogonal constraint mechanism is also introduced into the last layer of the sparse KAN network; a loss function used by the sparse KAN network during training includes a regularization loss function and a spectral clustering loss function, the regularization loss function is established based on an activation function after the norm regularization and the entropy regularization constraints are introduced into the sparse KAN network, and the spectral clustering loss function is established based on an optimized affinity matrix; the optimization process of the affinity matrix includes: initializing the affinity matrix by using training samples, establishing an optimization objective function based on the initial affinity matrix, introducing a manifold learning technique into the optimization objective function, projecting the training samples into a low-dimensional space, and solving the optimization objective function to complete the optimization of the affinity matrix.

[0009] In another aspect, the embodiments of the present application also provide a computer storage medium, which stores a plurality of computer instructions for causing a computer to execute the above method.

[0010] The identity recognition method, system and computer storage medium based on deep spectral clustering in the present application have the following advantages: (1) By introducing norm regularization and entropy regularization constraints, the activation function matrix of the KAN network is sparsified, realizing the sparsity of the parameters within the layer and avoiding the excessive weight of the single-layer matrix. At the same time, the diversity of the activation function distribution is ensured through entropy regularization, and the regularization weight is dynamically adjusted during the training process. In the early stage of training, the constraint force is reduced to retain the features, and in the later stage, the constraint is enhanced to improve the sparsity, further reducing the computational complexity of the network, and thus improving the efficiency of identity recognition.

[0011] (2) An orthogonalization constraint layer is added to the last layer of the network, and Cholesky decomposition is used to impose an orthogonal constraint on the embedding matrix to ensure the independence and decoupling of the feature representation, avoid feature redundancy in the feature learning process, and solve the problems of unclear optimization path, slow convergence speed and gradient explosion caused by feature redundancy in traditional methods. The convergence speed of the model during the training process is improved, the discriminative ability of the features is enhanced, and thus the stability of the identity recognition result is improved.

[0012] (3) The manifold learning technique is used to generate weighted coefficient adjustment samples to balance the global and local features, and these weighted features are projected into a low-dimensional space to extract more distinctive state feature representations, avoid feature redundancy, promote more effective processing and analysis, solve the strong dependence of traditional spectral clustering on static Gaussian kernel, avoid the gradient instability and insufficient global structure capture caused by the kernel width hyperparameter, and improve the accuracy of the identity recognition result. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings.

[0014] Figure 1 The flowchart of the identity recognition method based on deep spectral clustering provided by the embodiments of the present application. DETAILED DESCRIPTION

[0015] The technical solutions in the embodiments of the present application will be described clearly and completely in the embodiments of the present application combined with the drawings. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0016] Figure 1A flowchart illustrating the identity recognition method based on deep spectral clustering provided in this application embodiment. This application embodiment provides an identity recognition method based on deep spectral clustering, including the following steps: S100, obtain the user's identity data.

[0017] For example, identity data can be facial images, fingerprint images, or voiceprint signals. After obtaining this identity data, it needs to be cleaned and preprocessed. Data cleaning includes removing low-quality samples, duplicate samples, or mislabeled data, while preprocessing includes standardization and denoising and enhancement. Standardization involves uniformizing the data size, normalizing gray levels, or standardizing the mean and variance. Denoising and enhancement involve expanding the sample size through filtering, data augmentation, and other methods.

[0018] S110 uses a deep network model to extract features from identity data to obtain identity features.

[0019] For example, for image-based identity data, this application uses CNN (Convolutional Neural Network) to extract identity features, while for audio-based identity data, this application uses Transformer to extract identity features.

[0020] S120 utilizes a spectral clustering model to cluster identity features, obtaining the categories to which each feature belongs. The spectral clustering model is based on a sparse KAN network, which incorporates norm regularization and entropy regularization constraints. The final layer of the sparse KAN network also incorporates an orthogonal constraint mechanism. The loss functions used during training of the sparse KAN network include a regularization loss function and a spectral clustering loss function. The regularization loss function, after incorporating norm and entropy regularization constraints in the sparse KAN network, is based on an activation function. The spectral clustering loss function is based on an optimized affinity matrix. The optimization process for the affinity matrix includes: initializing the affinity matrix using training samples; establishing an optimization objective function based on the initial affinity matrix; incorporating manifold learning techniques into the optimization objective function; projecting the training samples into a low-dimensional space; and solving the optimization objective function to optimize the affinity matrix.

[0021] For example, the activation function matrix of the KAN network constructed in this application Represented as:

[0022] in, KAN ( x ) indicates that the input training sample, i.e., the identity feature, is x At that time, the results processed using the KAN network, L It is the number of layers of activation functions. l The layer number is the activation function. is the number of layers of the KAN network, The KAN network can have an arbitrary number of layers, and each layer can contain an arbitrary number of nodes. The activation function can be represented as:

[0023]

[0024] where, is the activation function in the th layer, l is the layer number index of the activation function, is the l th activation function in the th layer, and are the weights of the two parts, respectively, indicates the linear unit function, while is composed of multiple spline basis functions. Typically, the number of node vectors in the spline function is 8. Further, define the set of activation function matrices of each layer as Apply the norm constraint in the following way:

[0025] Intra-layer sparsity constraint: apply the 1 norm constraint to the activation function matrix of each layer, which sparsifies the parameters within a specific layer and reduces invalid connections. l Global balance constraint: apply the 2 norm constraint to the overall activation function matrix, which avoids optimization imbalance caused by excessively large single-layer matrix weights.

[0026] l The expression of the norm regularization is:

[0027]

[0028] where, denotes the norm regularization, L denotes the number of layers of the activation function, l is the layer number index of the activation function, denotes the activation function matrix composed of the activation functions in the l th layer, denotes taking the l 1 norm. This mechanism combines the advantages of the two norms, while achieving matrix sparsity and global structural balance.

[0029] ​​​To avoid the decline of expression ability caused by sparsification, entropy regularization is introduced, which is expressed as:

[0030] wherein, denotes the entropy regularization, L denotes the number of layers of the activation function, l is the layer number of the activation function, denotes the normalized value of the l activation function in the i , j layer, and is calculated as:

[0031] wherein, denotes the l 1 norm, l is the layer number of the activation function, k denotes the total number of nodes in the layer, m is the total number of nodes in the layer.

[0032] By maximizing the entropy value of the activation function distribution, the matrix is prevented from being singular, and the feature expression ability of the model for high-dimensional data is ensured. The regularization loss function is expressed as:

[0033] wherein, denotes the regularization loss function, t denotes the current training number, denotes the total training number, Lambda denotes the regularization weight parameter, denotes the norm regularization, denotes the entropy regularization. This strategy reduces the regularization strength at the beginning of training to avoid premature feature pruning, and gradually strengthens the constraint in the later period to improve the matrix sparsity.

[0034] Further, the present application introduces an orthogonal constraint mechanism in the sparse KAN network, optimizes the embedding space by forcing the independence of the feature vector, and improves the convergence stability and feature discrimination ability of the clustering task. The mapping relationship of the sparse KAN network is defined as wherein is the training sample, y is the output corresponding to the input x . To ensure the strong independence of the feature vector in the loss function, the orthogonal constraint is expressed as:

[0035] wherein, y denotes when the input training sample isx the output of the sparse KAN network, the upper index T denotes the transpose, denotes the expectation, denotes the identity matrix, c denotes the number of clusters.

[0036] The iteration training is performed by randomly selecting m training samples to form a small batch, and the orthogonal constraint is converted into a matrix form:

[0037] wherein, m is the number of randomly selected training samples, is an embedding matrix, is obtained by stacking by rows, denotes the identity matrix, c denotes the number of clusters. The orthogonalization calculation of the embedding matrix is realized by Cholesky decomposition, Cholesky decomposition is performed on the target matrix to obtain a lower triangular matrix , which satisfies , the orthogonal matrix is reconstructed through the decomposition result, and it is ensured that strictly satisfies the identity matrix constraint, and the optimization deviation caused by feature redundancy is avoided.

[0038] In the embodiments of the present application, in order to ensure the real-time optimization of the loss function, randomly select m samples as input each time the training iteration is performed, dynamically update the small batch data, enhance the adaptability of the model to the data distribution, and accelerate the convergence process of the orthogonal constraint.

[0039] Further, after the input of the training samples, the similarity measure inside the training samples is dynamically learned, and it is expected to learn the elements of the affinity matrix , that is, the affinity value , so that the smaller distance between the th training sample and the th training sample corresponds to a larger affinity value . For the global optimal similarity expression of the sample, that is, the initial optimization objective function can be expressed as:

[0040] wherein, denotes vector normalization, denotes the affinity vector of the th training sample to other training samples, m is the number of randomly selected training samples,x i and x j denote the i-th training sample, i and j denote the i-th training sample, denote the distance between the i-th training sample l 2-norm, is the affinity value in the affinity matrix is the regularization parameter, in order to guarantee the sparsity of the obtained affinity matrix, the regularization constraint is added in the traditional spectral clustering calculation, to ensure that the optimal solution of has exactly non-zero values.

[0041] The distance between the training sample and is defined as , which can be obtained as follows:

[0042] wherein, d i denote the Euclidean distance between the training sample and . If d i is reordered from small to large, the global optimal solution of the correlation matrix in the graph structure model, i.e. the initial value of the affinity matrix , can be finally obtained as follows:

[0043] wherein, is the affinity value in the affinity matrix , denotes the distance between the training sample and , and denotes the number of clusters. c

[0044] Considering the interference of noise in the training sample on the affinity matrix, the manifold learning technique is introduced to weight the training sample, so that the initial optimization objective function can be updated as follows:

[0045]

[0046] wherein, m is the number of randomly selected training samples, x i and x j denote the i-th training sample, i and j ​​training samples, Indicates taking l 2-norm, It is a diagonal matrix, and the diagonal elements of this diagonal matrix are... Used for feature weighting, balancing global and local features. Affinity matrix The affinity value in It is a regularization parameter. Represents vector standardization, Indicates the first The affinity vectors of each training sample to other training samples. By using manifold learning weighting, the influence of noise in the training samples is effectively reduced, and the ability of the affinity matrix to capture the local structure of the data is improved.

[0047] By introducing manifold learning techniques and projecting the training samples into a low-dimensional space, the aforementioned optimization objective function is updated again as follows:

[0048]

[0049] in, m This represents the number of training samples randomly selected during the training process. for W transpose, Denotes the projection subspace matrix, satisfying the orthogonality constraint. Represents a diagonal matrix. x i and x j They represent the first i and j training samples, Indicates taking l 2-norm, s ij Representing the affinity matrix S ij The affinity value in and Represents the regularization parameter. Represents the trace of a matrix. Represents vector standardization, Indicates the first The affinity vector of each training sample to other training samples. diagonal matrix diagonal elements in I Represents the identity matrix. The purpose of this method is to prevent the projection matrix from degenerating into a zero matrix, ensuring the effectiveness of the feature mapping. Through this projection mechanism, more discriminative low-dimensional features are extracted, feature redundancy is avoided, and the accuracy of subsequent clustering tasks is improved.

[0050] To obtain the global optimum of the objective function, gradient descent is used to solve it. During the solution process, the objective function is first fixed... and s ij Solving for the optimized solution yields the following results. W Then fix W and s ij Solving for the optimized solution yields the following results. Finally based on and W Solving for the optimal s ij .

[0051] Specifically, first fix and s ij ,right The optimization objective function is:

[0052] in, for W transpose, Denotes the projection subspace matrix. Represents the trace of a matrix. Represents a diagonal matrix. X Indicates input data, It is a lower triangular matrix. This represents the regularization parameter, calculated by... The former The eigenvectors corresponding to the smallest eigenvalues ​​are obtained, resulting in the optimized eigenvectors. .

[0053] Then fix and s ij ,right The objective function is:

[0054]

[0055] in, diagonal matrix The first in i diagonal elements, Indicates the first i indivual , For the intermediate matrix, Represents the regularization parameter. for W transpose, denotes a projection subspace matrix, denotes the i-th i , d denotes the data dimension of the input data, X denotes the input data, is a lower triangular matrix, denotes the dot product.

[0056] The optimized affinity matrix .

[0057] Further, the spectral clustering loss function is denoted as:

[0058] wherein, denotes the spectral clustering loss function, y i and y j denote the output of the sparse KAN network when the training sample is x i and x j , respectively, denotes the 2-norm, l s ij is the affinity value in the optimized affinity matrix S ij .

[0059] Based on the regularization loss function and the spectral clustering loss function, the loss function used in the training of the sparse KAN network in the present application can be obtained as follows: Loss

[0060] wherein, denotes the regularization loss function, denotes the spectral clustering loss function.

[0061] The embodiments of the present application further provide an identity recognition system based on deep spectral clustering, which comprises: a data acquisition module, configured to acquire identity data of a user; a deep network model, configured to extract features in the identity data to obtain identity features; a spectral clustering model, configured to cluster the identity features to obtain a category to which the identity features belong; ​​​The spectral clustering model is established based on a sparse KAN network, norm regularization and entropy regularization constraints are introduced into the sparse KAN network, and an orthogonal constraint mechanism is also introduced into the last layer of the sparse KAN network; the loss function used by the sparse KAN network in training includes a regularization loss function and a spectral clustering loss function; the regularization loss function is established based on an activation function after the norm regularization and entropy regularization constraints are introduced into the sparse KAN network; and the spectral clustering loss function is established based on an optimized affinity matrix; the optimization process of the affinity matrix includes: initializing the affinity matrix by using training samples, establishing an optimization objective function based on the initial affinity matrix, introducing a manifold learning technology into the optimization objective function, projecting the training samples into a low-dimensional space, and solving the optimization objective function to complete the optimization of the affinity matrix.

[0062] The embodiment of the present application further provides a computer storage medium, which stores a plurality of computer instructions for enabling a computer to execute the method described above.

[0063] Although the preferred embodiments of the present application have been described, those skilled in the art who are informed of the basic inventive concept can make additional changes and modifications to the embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0064] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and equivalent technologies thereof, the present application also intends to include these modifications and variations.

Claims

1. A method of identity recognition based on deep spectral clustering, characterized in that, The method comprises: obtaining identity data of a user; extracting features in the identity data by using a deep network model to obtain identity features; performing clustering on the identity features by using a spectral clustering model to obtain a category to which the identity features belong, wherein the spectral clustering model is established based on a sparse KAN network, norm regularization and entropy regularization constraints are introduced into the sparse KAN network, and an orthogonal constraint mechanism is also introduced into a last layer of the sparse KAN network, a loss function used by the sparse KAN network during training comprises a regularization loss function and a spectral clustering loss function, the regularization loss function is established based on an activation function after the norm regularization and the entropy regularization constraints are introduced into the sparse KAN network, the spectral clustering loss function is established based on an optimized affinity matrix, and an optimization process of the affinity matrix comprises: initializing the affinity matrix by using training samples, establishing an optimization objective function based on an initial affinity matrix, introducing a manifold learning technique into the optimization objective function, projecting the training samples into a low-dimensional space, and solving the optimization objective function to complete optimization of the affinity matrix.

2. The deep spectral clustering based identity recognition method of claim 1, wherein, The norm regularization constraint comprises: Intra-layer sparsity constraint: imposing l 1-norm constraint; global balance constraint: imposing l 2-norm constraint; An expression of the norm regularization is: wherein, denotes the L1 norm regularization, L denotes the number of layers of activation functions, l is the layer index of the activation function, denotes the i-th l activation function matrix composed of the layers of activation functions, denotes the i-th l 1 norm. 3.The deep spectral clustering based identity recognition method of claim 2, wherein, The entropy regularization is expressed as: wherein, denotes the entropy regularization, L denotes the number of layers of activation functions, l is the layer index of the activation function, denotes the normalized value of the l layer the i , j activation function.

4. The deep spectral clustering based identity recognition method of claim 3, wherein, The regularization loss function is expressed as: wherein, denotes the regularization loss function, t denotes the current training number, denotes the total training number, The orthogonal constraint is expressed as: denotes the regularization weight parameter, denotes the norm regularization, denotes the entropy regularization.

5. The deep spectral clustering based identity recognition method of claim 1, wherein, The spectral clustering loss function is expressed as: wherein, y denotes the output of the sparse KAN network when the input training sample is x , the upper index T denotes the transpose, denotes the expectation of , the upper index denotes the identity matrix, c denotes the number of clusters; By randomly selecting m a small batch of training samples to perform iterative training, the orthogonal constraint is converted into a matrix form: wherein, m is the number of randomly selected training samples, is the embedding matrix, is the number of clusters, is stacked by row, is the identity matrix, c is the number of clusters.

6. The deep spectral clustering based identity recognition method of claim 1, wherein, After the manifold learning technique is introduced and the training samples are projected into the low-dimensional space, the optimization objective function is expressed as: wherein, represents the spectral clustering loss function, m represents the number of training samples randomly selected during the training process, y i and y j respectively represent the output of the sparse KAN network when the training sample is x i and x j respectively represent the output of the sparse KAN network when the training sample is represents taking l 2-norm, s ij is the affinity value in the optimized affinity matrix S ij .

7. The deep spectral clustering based identity recognition method of claim 1, wherein, The method comprises: wherein, m denotes the number of randomly selected training samples in the training process, is the transpose of W denotes the projection subspace matrix, denotes the diagonal matrix, x i and x j denote the i-th and j-th training samples, i and j denote the i-th and j-th training samples, denotes the 2-norm, l s ij denotes the affinity value in the optimized affinity matrix S ij and denote the regularization parameters, denotes the trace of a matrix.​​​ 8. The deep spectral clustering based identity recognition method of claim 7, wherein, Solving the optimization objective function using gradient descent method, in the solving process, first fix and s ij , the solving obtains the optimized W , then fix W and s ij , the solving obtains the optimized , finally, based on and W solving the optimal s ij .

9. System for applying the method for identity recognition based on deep spectral clustering according to any of claims 1 to 8, characterized in that, a data acquisition module configured to obtain identity data of a user; a deep network model configured to extract features in the identity data to obtain identity features; a spectral clustering model configured to perform clustering on the identity features to obtain a category to which the identity features belong; the spectral clustering model is established based on a sparse KAN network, norm regularization and entropy regularization constraints are introduced into the sparse KAN network, and an orthogonal constraint mechanism is also introduced into a last layer of the sparse KAN network, a loss function used by the sparse KAN network during training comprises a regularization loss function and a spectral clustering loss function, the regularization loss function is established based on an activation function after the norm regularization and the entropy regularization constraints are introduced into the sparse KAN network, the spectral clustering loss function is established based on an optimized affinity matrix, and an optimization process of the affinity matrix comprises: initializing the affinity matrix by using training samples, establishing an optimization objective function based on an initial affinity matrix, introducing a manifold learning technique into the optimization objective function, projecting the training samples into a low-dimensional space, and solving the optimization objective function to complete optimization of the affinity matrix. The computer storage medium stores a plurality of computer instructions, and the plurality of computer instructions are used to make a computer execute the method in any one of claims 1-8.

10. A computer storage medium, characterized in that, ​

Images