WO2021114262A1 - Facial image clustering method and apparatus, and computer-readable storage medium - Google Patents

Abstract

Provided are a facial image clustering method and apparatus, and a computer-readable storage medium. The method comprises the following steps: acquiring training sample data of a facial image, the largest characteristic amount corresponding to the training sample data, and a data set to be clustered of a facial image to be clustered; inputting the training sample data and the largest characteristic amount into a BP-Deep NMF model, training the BP-Deep NMF model, and determining a base image amount of the facial image to be clustered; and on the basis of the base image amount and the data set to be clustered, classifying, by means of a preset clustering rule, the facial image to be clustered, and determining a clustering result of the data set to be clustered, so as to obtain a classification result of the facial image to be clustered.

Description

Face image clustering method, device and computer readable storage medium Technical field

This application relates to the field of face recognition technology, and in particular to a method, device and computer-readable storage medium for clustering face images.

Background technique

With the promotion of the application of face recognition and retrieval systems, the face image data in the system has increased dramatically, and face clustering technology has become an important foundation for improving the retrieval efficiency of the system. Face clustering is usually to cluster the face image information in the database into some different sub-categories, so that the similarity between the sub-categories is as small as possible, and the similarity within the sub-categories is as large as possible. The subcategories with high similarity to the searched target are identified one by one, and several records with the greatest similarity are retrieved.

The feature extraction step occupies an important position in face recognition and face clustering technology. Principal component analysis and singular value decomposition are relatively classic feature extraction methods, but the feature vectors proposed by these two methods usually contain negative elements, so when the original sample is non-negative data, these methods are not reasonable and interpretable . NMF (Non-negative Matrix Factorization) is a feature extraction method for processing non-negative data. It has a wide range of applications, such as hyperspectral data processing and face image recognition. In the process of matrix decomposition of non-negative data of the original sample, the NMF algorithm has non-negativity restrictions on the extracted features, that is, all components after decomposition are non-negative, so non-negative sparse features can be extracted. The essence of the NMF algorithm is to approximately decompose the non-negative matrix X into the product of the base image matrix W and the coefficient matrix H, that is, X≈WH, and both W and H are non-negative matrices. In this way, each column of matrix X can be expressed as a non-negative linear combination of matrix W column vectors, which is also in line with the construction basis of the NMF algorithm-the perception of the whole is composed of the perception of the parts that make up the whole (pure additive) . In recent years, scholars have proposed many algorithms for deforming NMF, such as local NMF algorithm that strengthens local restrictions, discriminative NMF algorithm that integrates discriminant information, and symmetric NMF algorithm for symmetric matrices. Although the NMF algorithm and its variants have achieved certain results, the method only considers the shallow information of the data. For data with rich features, the single-layer structure decomposed at one time cannot learn the representation of features from multiple angles.

At present, DL (Deep Learning, deep learning) has become the current research boom. Deep learning has established a deep neural network with a hierarchical structure, which brings hope to solving optimization problems related to deep structure. Inspired by the success of deep learning technology, some researchers have proposed the Deep Non-negative Matrix Factorization (Deep Non-negative Matrix Factorization) model based on the single-layer NMF algorithm. Deep NMF can be seen as decomposing a complex task into several simple tasks, and then processing them one after another in a multi-layer structure. At the same time, this deep decomposition method can explore the underlying feature representation in complex data, so as to extract features that are more complete and more discriminative than single-layer learning.

Although the existing Deep NMF model has a deep layered structure, this structure is generally constructed by simply reusing the single-layer NMF algorithm, and its performance does not meet the ideal requirements, and the calculation efficiency of the Deep NMF calculation method is not High, poor performance in face data clustering tasks. In particular, none of these methods are produced by using deep neural networks, so they cannot use the powerful feature expression and clustering capabilities of deep neural networks.

The above content is only used to assist the understanding of the technical solution of the application, and does not mean that the above content is recognized as prior art.

Summary of the invention

The main purpose of this application is to provide a face image clustering method, device, and computer-readable storage medium, aiming to solve the technical problem of poor feature expression ability and clustering ability of Deep NMF.

To achieve the above objective, the present application provides a face image clustering method. The face image clustering method includes the following steps:

Acquiring training sample data of the face image, the highest feature amount corresponding to the training sample data, and the to-be-clustered data set of the face image to be clustered;

Input the training sample data and the highest feature amount into a BP-Deep NMF model, train the BP-Deep NMF model, and determine the base image amount of the face image to be clustered;

Based on the amount of base images and the data set to be clustered, the face images to be clustered are classified by preset clustering rules, the clustering result of the data set to be clustered is determined, and the data set to be clustered is obtained The classification result of the face image.

In addition, in order to achieve the above object, the present application also provides a face image clustering device, the face image clustering device includes: a memory, a processor, and stored in the memory and running on the processor When the computer-readable instructions are executed by the processor, the following steps are implemented:

Acquiring training sample data of the face image, the highest feature amount corresponding to the training sample data, and the to-be-clustered data set of the face image to be clustered;

Input the training sample data and the highest feature amount into a BP-Deep NMF model, train the BP-Deep NMF model, and determine the base image amount of the face image to be clustered; and,

Based on the amount of base images and the data set to be clustered, the face images to be clustered are classified by preset clustering rules, the clustering result of the data set to be clustered is determined, and the data set to be clustered is obtained The classification result of the face image.

This application acquires the training sample data of the face image, the highest feature amount corresponding to the training sample data, and the to-be-clustered data set of the face image to be clustered; inputting the training sample data and the highest feature amount to In the BP-Deep NMF model, the BP-Deep NMF model is trained, and the base image amount of the face image to be clustered is determined; based on the base image amount and the data set to be clustered, the aggregation is performed by preset The class rule classifies the face images to be clustered, determines the clustering results of the data sets to be clustered, and obtains the classification results of the face images to be clustered. A high-performance depth based on BP neural network is proposed. Non-negative matrix factorization (BP-Deep NMF) model. The BP-Deep NMF model uses radial basis functions (RBF) to construct the input signal of the neural network. This input is equivalent to the highest-level feature in the deep non-negative matrix factorization, and uses the original The training data is used as the expected output of the network, and the optimization of the model adopts the BP neural network algorithm to update the network weight matrix. The finally trained BP-Deep NMF model can directly obtain the deep non-negative matrix decomposition of the data without fine-tuning the decomposition, and shows superior performance in the task of face data clustering.

Description of the drawings

FIG. 1 is a schematic structural diagram of a face image clustering device in a hardware operating environment involved in a solution of an embodiment of the present application;

FIG. 2 is a schematic flowchart of the first embodiment of the applicant's face image clustering method.

The realization, functional characteristics, and advantages of the purpose of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings.

Detailed ways

It should be understood that the specific embodiments described here are only used to explain the present application, and are not used to limit the present application.

As shown in FIG. 1, FIG. 1 is a schematic structural diagram of a face image clustering apparatus in a hardware operating environment involved in a solution of an embodiment of the present application.

As shown in Fig. 1, the face image clustering apparatus may include: a processor 1001, such as a CPU, a network interface 1004, a user interface 1003, a memory 1005, and a communication bus 1002. Among them, the communication bus 1002 is used to implement connection and communication between these components. The user interface 1003 may include a display screen (Display) and an input unit such as a keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory, or a non-volatile memory (non-volatile memory), such as a magnetic disk memory. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

Those skilled in the art can understand that the structure of the face image clustering device shown in FIG. 1 does not constitute a limitation on the face image clustering device, and may include more or less components than shown in the figure, or a combination of certain components. Components, or different component arrangements.

As shown in FIG. 1, the memory 1005 as a computer storage medium may include an operating system, a network communication module, a user interface module, and a face image clustering program.

In the face image clustering device shown in FIG. 1, the network interface 1004 is mainly used to connect to the back-end server and communicate with the back-end server; the user interface 1003 is mainly used to connect to the client (user side) and conduct data with the client Communication; and the processor 1001 can be used to call the face image clustering program stored in the memory 1005.

In this embodiment, the face image clustering device includes: a memory 1005, a processor 1001, and a face image clustering program stored on the memory 1005 and running on the processor 1001, wherein the processor When 1001 calls the face image clustering program stored in the memory 1005, and performs the following operations:

Acquiring training sample data of the face image, the highest feature amount corresponding to the training sample data, and the to-be-clustered data set of the face image to be clustered;

Inputting the training sample data and the highest feature amount into a BP-Deep NMF model, and training the BP-Deep NMF model to obtain the base image amount of the face image to be clustered;

Based on the amount of base images and the data set to be clustered, the face images to be clustered are classified by preset clustering rules, and the clustering result of the data set to be clustered is determined to obtain the data set to be clustered. The classification result of the human-like face image.

Further, the processor 1001 may call a face image clustering program stored in the memory 1005, and also perform the following operations:

Acquiring the first linear parameter corresponding to the activation function in the BP-Deep NMF model and the first weight of each neuron in the BP-Deep NMF model;

Determine the first output error of the output layer neuron in the BP-Deep NMF model based on the first linear parameter, the training sample data, and the first weight;

Determine the second output error of the input layer neuron and the hidden layer neuron in the BP-Deep NMF model based on the first output error, the first linear parameter, and the first weight;

Based on the first output error and the second output error, determine the second weight of each neuron in the BP-Deep NMF model after the training of the BP-Deep NMF model is completed, and set the second weight As the base image amount of the face image to be clustered.

Further, the processor 1001 may call a face image clustering program stored in the memory 1005, and also perform the following operations:

Determine the weight offset of each neuron in the BP-Deep NMF model based on the first output error and the second output error;

Based on the weight bias, determine the second weight of each neuron in the BP-Deep NMF model after the training of the BP-Deep NMF model is completed, and use the second weight as the person to be clustered The amount of the base image of the face image.

Further, the processor 1001 may call a face image clustering program stored in the memory 1005, and also perform the following operations:

Acquiring the learning rate of the BP-Deep NMF model;

Based on the weight offset and the learning rate, the non-negative weight of each neuron in the BP-Deep NMF model is determined by the projection of the projection gradient method;

Based on the non-negative weight, determine the second weight of each layer of neurons in the BP-Deep NMF model after the BP-Deep NMF model training is completed, and use the second weight as the face image Base image volume.

Further, the processor 1001 may call a face image clustering program stored in the memory 1005, and also perform the following operations:

Based on the training sample data and the non-negative weights, determine the second linear parameter corresponding to the activation function in the BP-Deep NMF model, and use the second linear parameter as the first linear parameter to perform all Said determining the first output error of the output layer neuron in the BP-Deep NMF model based on the first linear parameter, the training sample data and the first weight; The first linear parameter and the first weight determine the step of determining the second output error of the input layer neuron and the hidden layer neuron in the BP-Deep NMF model.

Further, the processor 1001 may call a face image clustering program stored in the memory 1005, and also perform the following operations:

Acquiring the number of times of training of the BP-Deep NMF model and the loss function value of the BP-Deep NMF model;

If the number of training times reaches the maximum number of training times or the loss function value is less than or equal to the model error threshold, stop training the BP-Deep NMF model, and use the non-negative weight as the second weight to determine the The second weight of each layer of neurons in the BP-Deep NMF model after the training of the BP-Deep NMF model is completed, and the second weight is used as the base image amount of the face image.

Further, the processor 1001 may call a face image clustering program stored in the memory 1005, and also perform the following operations:

Acquiring the face image, and converting the face image into training sample data;

Acquiring the face image to be clustered, and converting the face image to be clustered into a data set to be clustered;

Based on the training sample data, the highest feature amount corresponding to the training sample data is determined through a preset calculation method.

Further, the processor 1001 may call a face image clustering program stored in the memory 1005, and also perform the following operations:

Extracting hidden feature quantities of features of each layer in the to-be-clustered data set based on the base image amount and the to-be-clustered data set;

Based on the hidden feature quantity, the face image to be clustered is classified by a preset clustering rule, and the clustering result of the data set to be clustered is determined to obtain the clustering result of the face image.

The present application also provides a face image clustering method. Referring to FIG. 2, FIG. 2 is a schematic flowchart of the first embodiment of the method in this application. The face image clustering method includes the following steps:

A system architecture to which the embodiment of the present application is applicable includes but is not limited to one server or multiple servers. The server may be a network device such as a computer. The server can be an independent device or a server cluster formed by multiple servers. Preferably, the server can use cloud computing technology for information processing.

Although the existing Deep NMF model has a deep hierarchical structure, this structure is generally constructed by simply reusing the single-layer NMF algorithm, and its performance does not meet the ideal requirements, and the computational efficiency of these methods is not high. In particular, none of these methods are produced by using deep neural networks, so they cannot use the powerful feature expression and clustering capabilities of deep neural networks. In view of this, this patent introduces the neural network algorithm into the Deep NMF field, and proposes a new high-performance deep non-negative matrix factorization (BP-Deep NMF) clustering model based on BP neural network. First, deep non-negative matrix factorization is constructed. (BP-Deep NMF) model, and then use the BP-Deep NMF model to cluster the data to be clustered, that is, cluster the face images to be clustered. The BP-Deep NMF method uses the radial basis function (RBF) to construct the input signal of the neural network. The input is equivalent to the highest-level feature in the deep non-negative matrix factorization, and the original training data is used as the expected output of the network. The optimization of the model uses the BP neural network algorithm to update the network weight matrix. Finally, the trained neural network model can directly obtain the deep non-negative matrix decomposition of the data without fine-tuning the decomposition. The BP-Deep NMF model proposed in this application shows superior performance in the task of clustering face data.

In order to facilitate understanding, the terms that may be involved in the embodiments of the present application are defined and explained below.

Deep non-negative matrix factorization (Deep NMF): NMF is only a single-layer decomposition learning process. It learns the base matrix W and the feature matrix H by decomposing the data matrix X. The deep non-negative matrix factorization (Deep NMF) can further capture the intuitive hierarchical feature information hidden in the data set. _{The basic idea is: the matrix H 1} obtained by the first shallow decomposition of the matrix X can be decomposed into the matrices W ₂ and H _{2 again} , thereby expanding the original single-layer structure into a two-layer structure. By analogy, the shallow structure model can eventually be expanded to a multi-layer (deep) structure model. Specifically, the Deep NMF model decomposes the non-negative data matrix X into L+1 non-negative factor matrices, namely:

X≈W ₁ W ₂ …W _L H _L ,

among them

From the above formula, the hidden attributes H ₁ ,..., H _L of the data learned by the Deep NMF model can be expressed as H ₁ ≈W ₂ …W _L H _L , H ₂ ≈W ₃ …W _L H _L ,. .., H _L-1 ≈W _L H _L , said H ₁ ,..., H _L is the characteristic matrix under non-negative constraints. In order to reduce the total reconstruction error of the entire model, the Deep NMF algorithm generally performs fine-tuning of the entire network model after the layer-by-layer decomposition is completed, that is, the following loss function based on the F-norm is minimized:

BP neural network (Back-Propagation Algorithm of Neural Network): The neural network continuously changes the weight of the network through the sample signal, adjusts the error value of the network, and finally the output error reaches the expected error range. Among them, the BP neural network is a multi-layer feedforward network based on error back propagation, while the standard BP neural network includes an input layer, a hidden layer, and an output layer. The layers are connected to each other through neurons. The neurons in the same layer of the network are not connected to each other.

Step S10: Obtain the training sample data of the face image, the highest feature amount corresponding to the training sample data, and the to-be-clustered data set of the face image to be clustered;

In this embodiment, before constructing the BP-Deep NMF model, the face image used to train the BP-Deep NMF model is first converted into training sample data, and the radial basis function (RBF) is used to convert the training sample data into The highest feature quantity, the input signal of the radial basis function (RBF) to construct the neural network is the highest feature quantity. This input is equivalent to the highest level feature in the deep non-negative matrix factorization. Among them, the face image is used for training BP-Deep NMF The facial image information of the model includes facial image information with different expressions, different identities, or different facial features. After the face image is converted into training sample data and the training sample data is converted into the highest feature quantity, the converted training sample data and the highest feature quantity are obtained to construct the BP-Deep NMF model.

(其中为空间x,y中两个数据点，t≥0)生成算法最高层特征矩阵即最高特征量H＝(H _ij) _n×n，其中，最高特征量

当x _i与x _j属于同一个人的不同脸部特征或者不同表情的人脸图像时，x _i与x _j属于同一类，则H _ij＝k(x _i,x _j)；当x _i与x _j属于不同人的脸部特征或者表情的人脸图像时，x _i与x _j属于不同类，则H _ij＝0。因此，显然，最高特征量H是对角分块矩阵，这种结构本身具有很好的聚类性质。 Specifically, it is assumed that the number of training samples is n and the training sample data matrix, that is, the training sample data is X=(x ₁ , x ₂ ,..., x _n ). Utilize Radial Basis Function (RBF)

(Among them are two data points in space x and y, t≥0) Generate the highest feature matrix of the algorithm, that is, the highest feature quantity H=(H _ij ) _n×n , where the highest feature quantity

When x _i and x _j belong to the same person’s face images with different facial features or different expressions, x _i and x _j belong to the same category, then H _ij =k(x _i , x _j ); when x _i and x _{When j} belongs to facial images of different people's facial features or expressions, and x _i and x _j belong to different classes, then H _ij =0. Therefore, it is obvious that the highest feature quantity H is a diagonal block matrix, and this structure itself has good clustering properties.

After the construction of the BP-Deep NMF model is completed, the to-be-clustered data set of the face images to be clustered is first obtained for clustering the face images to be clustered.

Step S20: Input the training sample data and the highest feature amount into a BP-Deep NMF model, train the BP-Deep NMF model, and determine the base image amount of the face image to be clustered;

In this embodiment, after the training sample data of the face image and the highest feature amount corresponding to the training sample data are obtained, the BP-Deep NMF model is built, that is, the BP-Deep NMF model is trained. First, the training sample data X and the highest feature The quantity H is input into the BP-Deep NMF model, specifically, the training sample data X is used as the expected output of the BP neural network, that is, the original training sample data X is input into the output layer of the BP-Deep NMF model, and the highest feature quantity H is used as the input signal of the BP-Deep NMF model, that is, the highest feature quantity H is input to the input layer of the BP-Deep NMF model. After that, based on the training sample data and the highest feature amount, by training the BP-Deep NMF model, determine the base image amount W of the face images to be clustered after the BP-Deep NMF model training is completed. Specifically, input the highest feature amount H to Calculate the forward output of the BP neural network. After obtaining the forward output of the BP neural network, input the highest feature quantity to the output layer of the BP neural network, and combine the forward output of the BP neural network to find the reverse output of the BP neural network. Update the model parameters of the BP-Deep NMF model, that is, the weight of each neuron in the BP-Deep NMF model.

During the construction of the BP-Deep NMF model, the BP-Deep NMF model is continuously iterated until the number of training times reaches the maximum number of training times or the loss function value is less than or equal to a threshold, otherwise the BP-Deep NMF model continues to be iterated. When the number of training times reaches the maximum number of training times or the loss function value is less than or equal to a threshold, it indicates that the construction of the BP-Deep NMF model is completed. After the construction of the BP-Deep NMF model is completed, the base image amount W of the face image to be clustered of the BP-Deep NMF model is output.

Step S30, based on the amount of base images and the data set to be clustered, classify the face images to be clustered by preset clustering rules, determine the clustering result of the data set to be clustered, and obtain The classification result of the face image to be clustered.

运用预设聚类规则即k均值聚类法分别对样本第i层特征向量集

进行聚类确定待聚类数据集的聚类结果，最终输出聚类结果得到待聚类人脸图像的分类结果，隐藏特征量可以体现在人脸图像信息的身份或者表情或者姿势特征等。例如，基于姿势特征的一级隐藏特征，可以将待聚类人脸图像分类成不同姿势特征的聚类结果，例如头部朝向角度分别为0度、30度或者-30度等等角度的聚类结果；基于表情特征的二级隐藏特征，可以将待聚类人脸图像分类成不同表情特征的聚类结果，例如表情开心、愤怒或者疑惑等不同表情特征的聚类结果；基于身份特征的三级隐藏特征，可以将待聚类人脸图像分类成不同人的聚类结果，也就是说以人为分类依据将待聚类人脸图像分类。 In this embodiment, after the construction of the BP-Deep NMF model is completed, the base image amount W of the face image to be clustered of the BP-Deep NMF model is output, based on the base image amount W and the data set to be clustered Y=(y ₁ ,y ₂ ,...,y _{m ), calculate the ith layer feature vector of each sample y k} (k=1,...,m) in the data set to be clustered, which is the hidden feature quantity

Use the preset clustering rules, namely the k-means clustering method, to separately analyze the feature vector set of the i-th layer of the sample

Perform clustering to determine the clustering result of the data set to be clustered, and finally output the clustering result to obtain the classification result of the face image to be clustered. The hidden feature amount can be reflected in the identity or expression or posture characteristics of the face image information. For example, based on the first-level hidden features of posture features, the face images to be clustered can be classified into clustering results of different posture features, for example, the head orientation angles are 0 degrees, 30 degrees, or -30 degrees. Class results; based on the secondary hidden features of expression features, the face images to be clustered can be classified into clustering results of different expression features, such as the clustering results of different expression features such as happy, angry, or confused; based on identity features The three-level hidden features can classify the face images to be clustered into clustering results of different people, that is to say, classify the face images to be clustered on the basis of human classification.

The face image clustering method proposed in this embodiment proposes a high-performance deep non-negative matrix factorization (BP-Deep NMF) clustering model based on BP neural network. The BP-Deep NMF model uses radial basis functions (RBF) The input signal of the neural network is constructed. The input is equivalent to the highest-level feature in the deep non-negative matrix factorization, and the original training data is used as the expected output of the network. The optimization of the model adopts the BP neural network algorithm to update the network weight matrix. The finally trained BP-Deep NMF model can directly obtain the deep non-negative matrix decomposition of the data without fine-tuning the decomposition, and shows superior performance in the task of face data clustering.

Based on the first embodiment, a second embodiment of the applicant’s face image clustering method is proposed. In this embodiment, step S20 includes:

Step a: Obtain the first linear parameter corresponding to the activation function in the BP-Deep NMF model and the first weight of each neuron in the BP-Deep NMF model;

In this embodiment, first set the number of layers L of the BP neural network, that is, the BP neural network has a total of L layers of neurons, and f(x)=p ^1/L ·x(p>0) is used to construct the BP neural network The first linear parameter is the linear parameter p in the activation function f(x)=p ^1/L ·x. The first linear parameter p acts on the input signal of each neuron in the BP-Deep NMF model Converted into an output signal, the sum of the product of the input of the neuron and the weight corresponding to the input, and the activation function f(x) is applied to it to obtain the output of the neuron of this layer and feed it as the input to the next neural network Floor.

Step b: Determine the first output error of the output layer neuron in the BP-Deep NMF model based on the first linear parameter, the training sample data, and the first weight;

构建BP-Deep NMF模型开始，将最高特征量H输入至BP-Deep NMF模型的输入层，则BP神经网络输入层的输入为a ₀＝H，BP神经网络的层数为L，BP神经网络第l层加权重后的值为Z _l＝W _la _l-1，BP神经网络第l层的输出为a _l＝p ^1/L·Z _l，其中，l＝1,...,L。基于训练样本数据X、第一线性参数p和BP神经网络的第一权重W，输出BP神经网络输出层神经元的第一输出误差δ _L＝p ^1/L·(X-p·W _L·W _L-1·...·W ₁·H)。 In this embodiment, f(x)=p ^1/L ·x (parameter p>0) is the activation function of the BP neural network, and the bias vector b=0 is set, and the loss function of the BP neural network can be defined as :

Begin building the BP-Deep NMF model, input the highest feature amount H into the input layer of the BP-Deep NMF model, then the input of the BP neural network input layer is a ₀ = H, the number of layers of the BP neural network is L, and the BP neural network The weighted value of the l layer is Z _l =W _l a _l-1 , the output of the l layer of the BP neural network is a _l =p ^1/L ·Z _l , where l=1,...,L . Based on the training sample data X, the first linear parameter p, and the first weight W of the BP neural network, output the first output error of the neurons in the output layer of the BP neural network δ _L = p ^1/L · (Xp · W _L · W _{L -1} ·...·W ₁ ·H).

Step c: Determine the second output error of the input layer neuron and the hidden layer neuron in the BP-Deep NMF model based on the first output error, the first linear parameter and the first weight;

In this embodiment, after outputting the first output error of the neurons in the output layer of the BP neural network, based on the obtained first output error, the first linear parameter p, and the first weight W, the first output error in the BP-Deep NMF model is output The second output error of the input layer neuron or hidden layer neuron of the l layer, the calculation formula of the _{second output error δ l is as follows:}

Among them, l=L-1,L-2,...,1, and W is the weight of the BP neural network.

Step d: Based on the first output error and the second output error, determine the second weight of each neuron in the BP-Deep NMF model after the training of the BP-Deep NMF model is completed, and combine the The second weight is used as the base image amount of the face image to be clustered.

In this embodiment, based on the first output error and the second output error, by iterating the BP-Deep NMF model, the second weight of each neuron after the BP-Deep NMF model training is completed is determined, and the second weight is output, and The second weight after the training of the BP-Deep NMF model is used as the base image amount of the face image to be clustered to determine the base image amount W of the face image to be clustered, specifically, the highest feature amount H is input to obtain BP The forward output of the neural network, after the forward output of the BP neural network is obtained, the first output error and the second output error are determined based on the forward output to obtain the reverse output of the BP neural network to update the BP-Deep NMF model The model parameter of the BP-Deep NMF model is the weight of each neuron in the BP-Deep NMF model, and the second weight after the completion of the BP-Deep NMF model training is used as the base image amount of the face image to be clustered, so as to be further based on this The amount of base image clusters the face images to be clustered.

Further, in an embodiment, the determining the first output error of each neuron in the BP-Deep NMF model after the training of the BP-Deep NMF model is completed based on the first output error and the second output error Two weights, and the step of using the second weight as the base image amount of the face image to be clustered includes:

Step e: Determine the weight offset of each neuron in the BP-Deep NMF model based on the first output error and the second output error;

的计算公式如下： In this embodiment, after determining the first output error δ _L and the second output error δ _l , each neuron of the BP neural network calculates _{according to the output first output error δ L} and the second output error δ _l The weight bias of each neuron in the BP-Deep NMF model, the weight bias

The calculation formula is as follows:

为第l-1层神经元的输出a _l-1的转置，δ _l为第二输出误差，以及参数l为l＝L,...,1。 among them,

Is the transpose _{of the output a l-1} of the neuron of the l-1th layer _{, δ l} is the second output error, and the parameter l is l=L,...,1.

Step f: Based on the weight bias, determine the second weight of each neuron in the BP-Deep NMF model after the training of the BP-Deep NMF model is completed, and use the second weight as the waiting The amount of base images for clustering face images.

In this embodiment, based on the weight offset, the BP-Deep NMF model is iterated to determine the second weight of each neuron after the BP-Deep NMF model training is completed, and the second weight is output, and the BP-Deep NMF After the model training is completed, the second weight is used as the base image amount of the face image to be clustered to determine the base image amount W of the face image to be clustered. Specifically, the highest feature amount H is input to obtain the forward direction of the BP neural network Output, after obtaining the forward output of the BP neural network, determine the first output error and the second output error based on the forward output to obtain the reverse output of the BP neural network to update the model parameter of the BP-Deep NMF model, namely BP -The weight of each neuron in the Deep NMF model for subsequent output. The second weight after the completion of the BP-Deep NMF model training is used as the base image amount of the face image to be clustered, so that the base image amount can be further calculated based on the base image amount. The face images to be clustered are clustered.

Further, in an embodiment, wherein, based on the weight bias, the second weight of each neuron in the BP-Deep NMF model is determined after the training of the BP-Deep NMF model is completed, and The step of using the second weight as the base image amount of the face images to be clustered includes:

Step g: Obtain the learning rate of the BP-Deep NMF model;

Step h, based on the weight offset and the learning rate, determine the non-negative weight of each neuron in the BP-Deep NMF model through the projection of the projection gradient method;

In this embodiment, after determining the weight offset of the BP-Deep NMF model, the learning rate r of the BP-Deep NMF model is obtained for updating the weight parameters of the BP neural network, and the BP is updated based on the weight offset and the learning rate The calculation formula of the weight parameter of the neural network is as follows:

Among them, l=L,...,1.

After updating the weight parameters of the BP neural network, the updated weight parameters of the BP neural network are projected by the projection gradient method to project into non-negative weight parameters. The projection weight parameters determine the non-negative parameters of each neuron in the BP-Deep NMF model. Negative weight.

Step i: Based on the non-negative weight, determine the second weight of each layer of neurons in the BP-Deep NMF model after the training of the BP-Deep NMF model is completed, and use the second weight as the person The amount of the base image of the face image.

In this embodiment, based on the projected non-negative weights, by iterating the BP-Deep NMF model, determine the second weight of each neuron after the BP-Deep NMF model training is completed, output the second weight, and add the BP-Deep NMF model. After Deep NMF model training is completed, the second weight is used as the base image amount of the face image to be clustered to determine the base image amount W of the face image to be clustered. Specifically, the highest feature amount H is input to obtain the BP neural network Forward output, after obtaining the forward output of the BP neural network, determine the first output error and the second output error based on the forward output to obtain the reverse output of the BP neural network to update the model parameters of the BP-Deep NMF model That is, the weight of each neuron in the BP-Deep NMF model is used to output the second weight after the completion of the BP-Deep NMF model training as the base image amount of the face image to be clustered, so as to be further based on the base image The face images to be clustered are clustered.

Further, in an embodiment, after the step of determining the non-negative weight of each neuron in the BP-Deep NMF model through the projection of the projection gradient method based on the weight offset and the learning rate, further include:

Step j: Determine a second linear parameter corresponding to the activation function in the BP-Deep NMF model based on the training sample data and the non-negative weight, and use the second linear parameter as the first linear parameter , Performing the determining the first output error of the output layer neuron in the BP-Deep NMF model based on the first linear parameter, the training sample data, and the first weight, and determining the first output error based on the first output The error, the first linear parameter, and the first weight determine the step of determining the second output error of the input layer neuron and the hidden layer neuron in the BP-Deep NMF model.

In this embodiment, after the step of determining the non-negative weight of each neuron in the BP-Deep NMF model through the projection of the projection gradient method based on the weight offset and the learning rate, that is, After obtaining the non-negative weight of each neuron in the BP-Deep NMF model, update the linear parameters of the activation function in the BP-Deep NMF model, that is, determine the second linearity corresponding to the activation function in the BP-Deep NMF model Parameters, taking the second linear parameter as the first linear parameter, the update formula of the linear parameter of the activation function is as follows:

It is understandable that in order to determine the optimal parameter p, sub-optimization of the following problem needs to be solved:

Using gradient descent method to solve the parameter p, there are:

是C _deep关于p ^(t)的导数，可以计算得： Where ρ(p ^(t) ) is the step vector of p,

Is the derivative of C _{deep with} respect to p ^(t) , which can be calculated as:

In order to ensure the non-negativity of ^{p (t+1), let:}

Combining the above formula, the update formula of the linear parameter of the activation function can be obtained.

Further, in an embodiment, the second weight of neurons in each layer of the BP-Deep NMF model after the completion of the training of the BP-Deep NMF model is determined based on the non-negative weight, and the The step of using the second weight as the base image amount of the face image includes:

Step k: Obtain the number of training times of the BP-Deep NMF model and the loss function value of the BP-Deep NMF model;

Step 1. If the number of training times reaches the maximum number of training times or the value of the loss function is less than or equal to the model error threshold, stop training the BP-Deep NMF model, use the non-negative weight as the second weight, and determine The second weight of each layer of neurons in the BP-Deep NMF model after the training of the BP-Deep NMF model is completed, and the second weight is used as the base image amount of the face image.

In this embodiment, the number of training times of the BP-Deep NMF model and the loss function value of the BP-Deep NMF model are obtained to detect whether the iteration of the BP-Deep NMF model is completed. It is detected that the number of training times reaches the maximum number of training times and the loss function value is less than or equal to the model error threshold. If the loss function value is less than or equal to the model error threshold value, ie C _deep ≤ε or the number of iterations, that is, the number of training times reaches the maximum number of training times I _max , indicating BP-Deep NMF model iteration is complete, the iteration is stopped BP-Deep NMF model output weight NMF model parameter BP-Deep weight matrix _{W i (i = 1, ...} , L).

Otherwise, if the number of training times has not reached the maximum number of training times, continue to obtain the training sample data of the face image and the highest feature amount corresponding to the training sample data, and input the training sample data and the highest feature amount into the BP-Deep NMF model, and based on the training For the sample data and the highest feature amount, the BP-Deep NMF model is trained to determine the base image amount of the face images to be clustered after the BP-Deep NMF model training is completed.

Or, the loss function value is greater than the model error threshold, continue to obtain the training sample data of the face image and the highest feature amount corresponding to the training sample data, and input the training sample data and the highest feature amount into the BP-Deep NMF model, and based on the training sample Data and the highest feature amount, through training the BP-Deep NMF model, determine the base image amount of the face image to be clustered after the BP-Deep NMF model training is completed.

It can be seen that once the BP neural network training is completed, the following deep non-negative matrix decomposition can be automatically obtained: X≈W _L W _L-1 …W ₁ H W _i ≥0, i=1, 2,...,LH≥0.

The face image clustering method proposed in this embodiment proposes a high-performance deep non-negative matrix factorization (BP-Deep NMF) clustering model based on BP neural network. The BP-Deep NMF method uses radial basis functions (RBF) The input signal of the neural network is constructed. The input is equivalent to the highest-level feature in the deep non-negative matrix factorization, and the original training data is used as the expected output of the network. The optimization of the model adopts the BP neural network algorithm to update the network weight matrix. The finally trained BP-Deep NMF model can directly obtain the deep non-negative matrix decomposition of the data without fine-tuning the decomposition, and shows superior performance in the task of face data clustering.

Based on the first embodiment, a third embodiment of the applicant’s face image clustering method is proposed. In this embodiment, step S10 includes:

Step m: Obtain the face image, and convert the face image into training sample data;

In this embodiment, the face image is the face image information used to train the BP-Deep NMF model, including face image information with different expressions, different identities, or different facial features. Preparations before constructing the BP-Deep NMF model Work, convert the face image used to train the BP-Deep NMF model into training sample data for subsequent construction of the BP-Deep NMF model. The training sample data has data labels, and the data labels are generally different people. That is, the face image includes the face image information of different people.

Step n: Obtain the face image to be clustered, and convert the face image to be clustered into a data set to be clustered;

In this embodiment, after the construction of the BP-Deep NMF model is completed, the face image to be clustered is obtained, and the face image to be clustered is converted into a data set to be clustered for subsequent processing of the face image to be clustered. Clustering.

Step o, based on the training sample data, determine the highest feature amount corresponding to the training sample data through a preset calculation method.

(其中为空间x,y中两个数据点，t≥0)生成算法最高层特征矩阵即最高特征量H＝(H _ij) _n×n，其中，最高特征量

当x _i与x _j属于同一个人的不同脸部特征或者不同表情的人脸图像时，x _i与x _j属于同一类，则H _ij＝k(x _i,x _j)；当x _i与x _j属于不同人的脸部特征或者表情的人脸图像时，x _i与x _j属于不同类，则H _ij＝0。因此，显然，最高特征量H是对角分块矩阵，这种结构本身具有很好的聚类性质。 In this embodiment, before constructing the BP-Deep NMF model, the radial basis function (RBF) is used to convert the training sample data into the highest feature quantity for subsequent construction of the BP-Deep NMF model. The radial basis function (RBF) The input signal of the neural network is the highest feature quantity, which is equivalent to the highest level feature in the deep non-negative matrix factorization. Assume that the number of training samples is n and the training sample data matrix, that is, the training sample data is X=(x ₁ , x ₂ ,..., x _n ). Utilize Radial Basis Function (RBF)

(Among them are two data points in space x and y, t≥0) Generate the highest feature matrix of the algorithm, that is, the highest feature quantity H=(H _ij ) _n×n , where the highest feature quantity

When x _i and x _j belong to the same person’s face images with different facial features or different expressions, x _i and x _j belong to the same category, then H _ij =k(x _i , x _j ); when x _i and x _{When j} belongs to facial images of different people's facial features or expressions, and x _i and x _j belong to different classes, then H _ij =0. Therefore, it is obvious that the highest feature quantity H is a diagonal block matrix, and this structure itself has good clustering properties.

Further, in an embodiment, the face image to be clustered is classified by a preset clustering rule based on the amount of base images and the data set to be clustered, and the data to be clustered is determined The steps of obtaining the classification result of the face image to be clustered include:

Step p, extracting the hidden feature amount of each layer feature in the to-be-clustered data set based on the base image amount and the to-be-clustered data set;

隐藏特征量

的计算公式如下： In this embodiment, after the construction of the BP-Deep NMF model is completed, the base image amount W of the face image to be clustered of the BP-Deep NMF model is output, based on the base image amount W and the data set to be clustered Y=(y ₁ ,y ₂ ,...,y _{m ), calculate the i-th layer feature vector of each sample y k} (k=1,...,m) in the data set to be clustered, that is, the hidden feature quantity

Hidden feature

The calculation formula is as follows:

Step q: Based on the hidden feature quantity, classify the face image to be clustered by preset clustering rules, determine the clustering result of the data set to be clustered, and obtain the clustering result of the face image.

进行聚类确定待聚类数据集的聚类结果，最终输出聚类结果得到待聚类人脸图像的分类结果，隐藏特征量可以体现在人脸图像信息的身份或者表情或者姿势特征等。例如，基于姿势特征的一级隐藏特征，可以将待聚类人脸图像分类成不同姿势特征的聚类结果，例如头部朝向角度分别为0度、30度或者-30度等等角度的聚类结果；基于表情特征的二级隐藏特征，可以将待聚类人脸图像分类成不同表情特征的聚类结果，例如表情开心、愤怒或者疑惑等不同表情特征的聚类结果；基于身份特征的三级隐藏特征，可以将待聚类人脸图像分类成不同人的聚类结果，也就是说以人为分类依据将待聚类人脸图像分类。 In this embodiment, based on the hidden feature amount, the preset clustering rule, namely the k-means clustering method, is used to separately analyze the feature vector set of the i-th layer of the sample.

Perform clustering to determine the clustering result of the data set to be clustered, and finally output the clustering result to obtain the classification result of the face image to be clustered. The hidden feature amount can be reflected in the identity or expression or posture characteristics of the face image information. For example, based on the first-level hidden features of posture features, the face images to be clustered can be classified into clustering results of different posture features, for example, the head orientation angles are 0 degrees, 30 degrees, or -30 degrees. Class results; based on the secondary hidden features of expression features, the face images to be clustered can be classified into clustering results of different expression features, such as the clustering results of different expression features such as happy, angry, or confused; based on identity features The three-level hidden features can classify the face images to be clustered into clustering results of different people, that is to say, classify the face images to be clustered on the basis of human classification.

The face image clustering method proposed in this embodiment proposes a high-performance deep non-negative matrix factorization (BP Deep NMF) clustering model based on BP neural network. The BP Deep NMF method uses radial basis functions (RBF) to construct neural networks. The input signal of the network is equivalent to the highest-level feature in the deep non-negative matrix factorization, and the original training data is used as the expected output of the network. The optimization of the model uses the BP neural network algorithm to update the network weight matrix. The finally trained BP-Deep NMF model can directly obtain the deep non-negative matrix decomposition of the data without fine-tuning the decomposition, and shows superior performance in the task of face data clustering.

In addition, an embodiment of the present application also proposes a computer-readable storage medium that stores a human face image clustering program, and the human face image clustering program may also be executed by a processor to implement the above-mentioned face image clustering program. The steps of each embodiment of the image clustering method.

The expanded content of the specific implementation of the computer storage medium of this application is basically the same as each embodiment of the above-mentioned face image clustering method, and will not be repeated here.

It should be noted that in this article, the terms "include", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements not only includes those elements, It also includes other elements that are not explicitly listed, or elements inherent to the process, method, article, or system. If there are no more restrictions, the element defined by the sentence "including a..." does not exclude the existence of other identical elements in the process, method, article, or system that includes the element.

The serial numbers of the foregoing embodiments of the present application are for description only, and do not represent the superiority or inferiority of the embodiments.

Through the description of the above implementation manners, those skilled in the art can clearly understand that the above-mentioned embodiment method can be implemented by means of software plus the necessary general hardware platform, of course, it can also be implemented by hardware, but in many cases the former is better.的实施方式。 Based on this understanding, the technical solution of this application essentially or the part that contributes to the existing technology can be embodied in the form of a software product, and the computer software product is stored in a storage medium (such as ROM/RAM) as described above. , Magnetic disks, optical disks), including several instructions to make a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) execute the method described in each embodiment of the present application. The above are only the preferred embodiments of the application, and do not limit the scope of the patent for this application. Any equivalent structure or equivalent process transformation made using the content of the description and drawings of the application, or directly or indirectly applied to other related technical fields , The same reason is included in the scope of patent protection of this application.

Claims (16)

A face image clustering method, wherein the face image clustering method includes the following steps: Acquiring training sample data of the face image, the highest feature amount corresponding to the training sample data, and the to-be-clustered data set of the face image to be clustered; Input the training sample data and the highest feature amount into a BP-Deep NMF model, train the BP-Deep NMF model, and determine the base image amount of the face image to be clustered; and, Based on the amount of base images and the data set to be clustered, the face images to be clustered are classified by preset clustering rules, the clustering result of the data set to be clustered is determined, and the data set to be clustered is obtained The classification result of the face image. The face image clustering method according to claim 1, wherein said inputting said training sample data and said highest feature amount into a BP-Deep NMF model, training said BP-Deep NMF model, and determining The step of the amount of base images of the face images to be clustered includes: Acquiring the first linear parameter corresponding to the activation function in the BP-Deep NMF model and the first weight of each neuron in the BP-Deep NMF model; Determine the first output error of the output layer neuron in the BP-Deep NMF model based on the first linear parameter, the training sample data, and the first weight; Determine the second output error of the input layer neuron and the hidden layer neuron in the BP-Deep NMF model based on the first output error, the first linear parameter, and the first weight; and, Based on the first output error and the second output error, determine the second weight of each neuron in the BP-Deep NMF model after the training of the BP-Deep NMF model is completed, and set the second weight As the base image amount of the face image to be clustered. 3. The face image clustering method according to claim 2, wherein, based on the first output error and the second output error, it is determined that the BP-Deep NMF model is trained after the completion of the BP-Deep NMF model training. The second weight of each neuron in the NMF model, and the step of using the second weight as the base image amount of the face image to be clustered includes: Determining the weight offset of each neuron in the BP-Deep NMF model based on the first output error and the second output error; and, Based on the weight bias, determine the second weight of each neuron in the BP-Deep NMF model after the training of the BP-Deep NMF model is completed, and use the second weight as the person to be clustered The amount of the base image of the face image. The face image clustering method of claim 3, wherein the weight bias is used to determine the value of each neuron in the BP-Deep NMF model after the training of the BP-Deep NMF model is completed. The second weight, and the step of using the second weight as the base image amount of the face image to be clustered includes: Acquiring the learning rate of the BP-Deep NMF model; Based on the weight offset and the learning rate, the non-negative weight of each neuron in the BP-Deep NMF model is determined by the projection of the projection gradient method; and, Based on the non-negative weight, determine the second weight of each layer of neurons in the BP-Deep NMF model after the BP-Deep NMF model training is completed, and use the second weight as the face image Base image volume. 5. The face image clustering method according to claim 4, wherein, based on the weight offset and the learning rate, the projection of the projection gradient method is used to determine the value of each neuron in the BP-Deep NMF model. After the step of non-negative weighting, it also includes: Based on the training sample data and the non-negative weights, determine the second linear parameter corresponding to the activation function in the BP-Deep NMF model, and use the second linear parameter as the first linear parameter to perform all Said determining the first output error of the output layer neuron in the BP-Deep NMF model based on the first linear parameter, the training sample data and the first weight; The first linear parameter and the first weight determine the step of determining the second output error of the input layer neuron and the hidden layer neuron in the BP-Deep NMF model. The face image clustering method according to claim 4, wherein the determination is based on the non-negative weight to determine the level of neurons in each layer of the BP-Deep NMF model after the training of the BP-Deep NMF model is completed. The second weight, and the step of using the second weight as the base image amount of the face image includes: Acquiring the number of training times of the BP-Deep NMF model and the loss function value of the BP-Deep NMF model; and, If the number of training times reaches the maximum number of training times or the value of the loss function is less than or equal to the model error threshold, stop training the BP-Deep NMF model, use the non-negative weight as the second weight, and determine the BP -The second weight of each layer of neurons in the BP-Deep NMF model after the Deep NMF model training is completed, and the second weight is used as the base image amount of the face image. The face image clustering method according to claim 1, wherein the acquiring training sample data of the face image, the highest feature amount corresponding to the training sample data, and the to-be-clustered data set of the face image to be clustered The steps include: Acquiring the face image, and converting the face image into training sample data; Acquiring the face image to be clustered, and converting the face image to be clustered into a data set to be clustered; and, Based on the training sample data, the highest feature amount corresponding to the training sample data is determined through a preset calculation method. The face image clustering method according to claim 1, wherein the face image to be clustered is classified by a preset clustering rule based on the amount of the base image and the data set to be clustered , The step of determining the clustering result of the data set to be clustered and obtaining the classification result of the face image to be clustered includes: Extracting the hidden feature quantities of the features of each layer in the data set to be clustered based on the base image amount and the data set to be clustered; and, Based on the hidden feature amount, the face image to be clustered is classified by a preset clustering rule, the clustering result of the data set to be clustered is determined, and the clustering result of the face image is obtained. The face image clustering method according to claim 2, wherein the face image to be clustered is classified by a preset clustering rule based on the amount of the base image and the data set to be clustered , The step of determining the clustering result of the data set to be clustered and obtaining the classification result of the face image to be clustered includes: Extracting the hidden feature quantities of the features of each layer in the data set to be clustered based on the base image amount and the data set to be clustered; and, Based on the hidden feature amount, the face image to be clustered is classified by a preset clustering rule, the clustering result of the data set to be clustered is determined, and the clustering result of the face image is obtained. The face image clustering method according to claim 3, wherein the face image to be clustered is classified by a preset clustering rule based on the amount of the base image and the data set to be clustered , The step of determining the clustering result of the data set to be clustered and obtaining the classification result of the face image to be clustered includes: Extracting the hidden feature quantities of the features of each layer in the data set to be clustered based on the base image amount and the data set to be clustered; and, Based on the hidden feature amount, the face image to be clustered is classified by a preset clustering rule, the clustering result of the data set to be clustered is determined, and the clustering result of the face image is obtained. The face image clustering method according to claim 4, wherein the face image to be clustered is classified by a preset clustering rule based on the amount of the base image and the data set to be clustered , The step of determining the clustering result of the data set to be clustered and obtaining the classification result of the face image to be clustered includes: Extracting the hidden feature amount of each layer feature in the to-be-clustered data set based on the base image amount and the to-be-clustered data set; and, Based on the hidden feature amount, the face image to be clustered is classified by a preset clustering rule, the clustering result of the data set to be clustered is determined, and the clustering result of the face image is obtained. The face image clustering method of claim 5, wherein the face image to be clustered is classified based on the amount of the base image and the data set to be clustered by using a preset clustering rule , The step of determining the clustering result of the data set to be clustered and obtaining the classification result of the face image to be clustered includes: Extracting the hidden feature quantities of the features of each layer in the data set to be clustered based on the base image amount and the data set to be clustered; and, Based on the hidden feature amount, the face image to be clustered is classified by a preset clustering rule, the clustering result of the data set to be clustered is determined, and the clustering result of the face image is obtained. The face image clustering method according to claim 6, wherein the face image to be clustered is classified based on the amount of the base image and the data set to be clustered by using a preset clustering rule , The step of determining the clustering result of the data set to be clustered and obtaining the classification result of the face image to be clustered includes: Extracting the hidden feature quantities of the features of each layer in the data set to be clustered based on the base image amount and the data set to be clustered; and, Based on the hidden feature amount, the face image to be clustered is classified by a preset clustering rule, the clustering result of the data set to be clustered is determined, and the clustering result of the face image is obtained. The face image clustering method according to claim 7, wherein the face image to be clustered is classified by a preset clustering rule based on the amount of the base image and the data set to be clustered , The step of determining the clustering result of the data set to be clustered and obtaining the classification result of the face image to be clustered includes: Extracting the hidden feature quantities of the features of each layer in the data set to be clustered based on the base image amount and the data set to be clustered; and, Based on the hidden feature amount, the face image to be clustered is classified by a preset clustering rule, the clustering result of the data set to be clustered is determined, and the clustering result of the face image is obtained. A face image clustering device, wherein the face image clustering device includes: a memory, a processor, and computer-readable instructions stored on the memory and running on the processor, the computer When the readable instruction is executed by the processor, the following steps are implemented: Acquiring training sample data of the face image, the highest feature amount corresponding to the training sample data, and the to-be-clustered data set of the face image to be clustered; Input the training sample data and the highest feature amount into a BP-Deep NMF model, train the BP-Deep NMF model, and determine the base image amount of the face image to be clustered; and, Based on the amount of base images and the data set to be clustered, the face images to be clustered are classified by preset clustering rules, the clustering result of the data set to be clustered is determined, and the data set to be clustered is obtained The classification result of the face image. A computer-readable storage medium, wherein computer-readable instructions are stored on the computer-readable storage medium, and when the computer-readable instructions are executed by a processor, the following steps are implemented: Acquiring training sample data of the face image, the highest feature amount corresponding to the training sample data, and the to-be-clustered data set of the face image to be clustered; Input the training sample data and the highest feature amount into a BP-Deep NMF model, train the BP-Deep NMF model, and determine the base image amount of the face image to be clustered; and, Based on the amount of base images and the data set to be clustered, the face images to be clustered are classified by preset clustering rules, the clustering result of the data set to be clustered is determined, and the data set to be clustered is obtained The classification result of the face image.

Images