Semi-supervised learning method and system

By optimizing the affinity matrix, label matrix, and extreme learning neural network parameters within a semi-supervised learning framework, the problem that existing methods cannot adapt to the limitations of high-dimensional manifold structures and graph structures is solved, enabling efficient processing and learning of image data.

CN113139505BActive Publication Date: 2024-11-12AEROSPACE NETWORK SECURITY TECH (SHENZHEN) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202110513201.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-05-11
Publication Date
2024-11-12
Estimated Expiration
2041-05-11

AI Technical Summary

Technical Problem

Existing graph-based semi-supervised learning methods cannot simultaneously adapt to high-dimensional manifold structures and restrict graph structures, leading to overfitting and failing to meet practical engineering needs.

Method used

A semi-supervised learning framework is adopted. By forming an image data matrix, normalization and Euclidean distance calculation are performed to generate a distance matrix, which is then input into a single-layer neural network. The affinity matrix and label matrix are optimized using preset activation and cost functions. Combined with graph convolutional layers and extreme learning neural networks, iterative optimization is performed to finally generate the label matrix output by the graph convolutional layer.

Benefits of technology

It enhances image data processing and learning capabilities, taking into account both the limitations of graph structures and the high-dimensional manifold structure with multiple features, thus avoiding algorithm overfitting.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113139505B_ABST
    Figure CN113139505B_ABST
Patent Text Reader

Abstract

The present disclosure provides a semi-supervised learning method and system; the method includes: forming an image data matrix from image data, normalizing it, obtaining the normalized Euclidean distance of the image data, forming a distance matrix with the Euclidean distance, and inputting it into a single-layer neural network; allowing the single-layer neural network to use the received distance matrix to generate an affinity matrix according to a preset activation function, and optimizing the affinity matrix by iteration of a preset cost function; allowing the image data matrix to be combined with the affinity matrix, and obtaining a label matrix output by a graph convolution layer according to a preset convolution function, and optimizing the label function by iteration of the preset cost function; sending the label matrix output by the convolution layer and the image data matrix to an extreme learning neural network, and optimizing the extreme learning neural network parameter matrix by iteration of the preset cost function.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] Embodiments of the present disclosure relate to the technical field of deep learning, and more particularly to a method and system for semi-supervised learning. Background Art

[0002] In graph-based semi-supervised learning methods, a semi-supervised learning framework based on graph convolutional neural networks with popular regularization is often used. However, its disadvantage is that it cannot meet the processing requirements of data with multiple features. Alternatively, a graph learning convolutional neural network semi-supervised learning framework is used in the prior art, but its disadvantage is that it does not restrict the graph structure, resulting in overfitting of the algorithm. It can be seen that the above two existing graph-based semi-supervised learning methods cannot meet actual engineering needs.

[0003] Based on this, a solution is needed that can simultaneously implement the above two semi-supervised learning methods. Summary of the invention

[0004] In view of this, the purpose of the present disclosure is to propose a method and system for semi-supervised learning.

[0005] Based on the above purpose, the present disclosure provides a method for semi-supervised learning, including: a training method for a semi-supervised learning framework, including:

[0006] The image data is formed into an image data matrix, and normalized, the Euclidean distance of the normalized image data is obtained, the Euclidean distance is formed into a distance matrix, and is input into a single-layer neural network;

[0007] The single-layer neural network generates an affinity matrix using the received distance matrix according to a preset activation function, and optimizes the affinity matrix using iteration of a preset cost function;

[0008] Combine the image data matrix with the affinity matrix, and obtain the label matrix output by the graph convolution layer according to the preset convolution function. Use the iteration of the preset cost function to optimize the label function.

[0009] The label matrix and image data matrix output by the convolutional layer are sent to the extreme learning neural network, and the extreme learning neural network parameter matrix is ​​optimized by iteration of the preset cost function for classification of image data.

[0010] Furthermore, the affinity matrix is ​​optimized by iterating a preset cost function, including:

[0011] The label function and the extreme learning neural network parameter matrix are fixed, and the gradient descent algorithm is used to iterate the first combination of the cost function to optimize the affinity matrix;

[0012] Utilize the iteration of the preset cost function to optimize the label function, including:

[0013] The affinity matrix and the extreme learning neural network parameter matrix are fixed, and the gradient descent algorithm is used to iterate the second combination of the cost function to optimize the label matrix;

[0014] Utilize the iteration of the preset cost function to optimize the extreme learning neural network parameter matrix, including:

[0015] Fix the affinity matrix and the label matrix, and use the third combination of cost functions to optimize the extreme learning neural network parameter matrix; and

[0016] In response to determining that the cost function does not appear to converge, continuing optimization of the extreme learning neural network parameter matrix, the affinity matrix, and the label matrix;

[0017] In response to determining that the cost function exhibits convergence, the optimization of the extreme learning neural network parameter matrix, the affinity matrix, and the label matrix is ​​stopped, and the optimization results are retained.

[0018] Furthermore, the cost function is expressed as:

[0019]

[0020] Wherein, a first combination consisting of a second function term, a third function term and a fourth function term is adopted in the cost function to optimize the affinity matrix;

[0021] A second combination consisting of the first function term, the second function term and the fifth function term is adopted in the cost function to optimize the label matrix.

[0022] The third combination of the fifth function term is adopted in the cost function to optimize the extreme learning neural network parameter matrix;

[0023] Furthermore, the fifth function term is used in the cost function to optimize the extreme learning neural network parameter matrix, including:

[0024] Formula

[0025]

[0026] Take the partial derivative of W in the equation and set the derivative result equal to zero to get the value of W.

[0027] Where W is the extreme learning neural network parameter matrix.

[0028] Furthermore, a gradient descent algorithm is used for the cost function to optimize the label matrix, including:

[0029] Using image data, a label matrix is ​​established as the initialization value for optimizing the label matrix;

[0030] Among them, the label data in the image data is encoded using unique hot encoding as the label data;

[0031] A vector that sets the unlabeled data in the image data to zero;

[0032] The same data label represents the same category of image data, and each category of image data has no less than one data sample.

[0033] Furthermore, an affinity matrix is ​​generated according to a preset activation function, including:

[0034] Using formula

[0035]

[0036] As the activation function of a single-layer neural network, the data matrix is ​​mapped to an affinity matrix.

[0037] Furthermore, the label matrix output by the graph convolution layer is obtained according to the preset convolution function, including:

[0038] In the graph convolutional layer, the formula is used

[0039] H (l+1) =f(SH (l) Z (l+1) )

[0040] And combined with the rules of affinity matrix, the image data matrix is ​​mapped to the output matrix of graph convolution layer;

[0041] Use the softmax classifier to map the output values ​​of the graph convolutional layer output matrix into probability values ​​and use them as the label matrix;

[0042] Wherein, at least one convolutional layer is set in the graph convolutional layer neural network.

[0043] Based on the same inventive concept, the present disclosure also provides a semi-supervised learning system, including: an image data preprocessing layer, a single-layer neural network, a graph convolutional layer neural network, and an extreme learning neural network;

[0044] Among them, the label function and the extreme learning neural network parameter matrix are fixed, and the gradient descent algorithm is used for the cost function to optimize the affinity matrix;

[0045] Utilize the iteration of the preset cost function to optimize the label function, including:

[0046] The affinity matrix and the extreme learning neural network parameter matrix are fixed, and the gradient descent algorithm is used for the cost function to optimize the label matrix;

[0047] Utilize the iteration of the preset cost function to optimize the extreme learning neural network parameter matrix, including:

[0048] Fix the affinity matrix and label matrix, and optimize the extreme learning neural network parameter matrix using the cost function; and

[0049] In response to determining that the cost function does not appear to converge, continuing optimization of the extreme learning neural network parameter matrix, the affinity matrix, and the label matrix;

[0050] In response to determining that the cost function exhibits convergence, the optimization of the extreme learning neural network parameter matrix, the affinity matrix, and the label matrix is ​​stopped, and the optimization results are retained.

[0051] Furthermore, the affinity matrix is ​​optimized by iterating a preset cost function, including:

[0052] The label function and the extreme learning neural network parameter matrix are fixed, and the gradient descent algorithm is used for the cost function to optimize the affinity matrix;

[0053] Utilize the iteration of the preset cost function to optimize the label function, including:

[0054] The affinity matrix and the extreme learning neural network parameter matrix are fixed, and the gradient descent algorithm is used for the cost function to optimize the label matrix;

[0055] Utilize the iteration of the preset cost function to optimize the extreme learning neural network parameter matrix, including:

[0056] Fix the affinity matrix and label matrix, and optimize the extreme learning neural network parameter matrix using the cost function; and

[0057] In response to determining that the cost function does not appear to converge, continuing optimization of the extreme learning neural network parameter matrix, the affinity matrix, and the label matrix;

[0058] In response to determining that the cost function exhibits convergence, the optimization of the extreme learning neural network parameter matrix, the affinity matrix, and the label matrix is ​​stopped, and the optimization results are retained.

[0059] Based on the same inventive concept, the present disclosure also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the semi-supervised learning method as described in any one of the above items is implemented.

[0060] From the above, it can be seen that the method and system for semi-supervised learning provided by the present invention, based on popular regularization and graph learning convolutional neural network, comprehensively consider the mechanism of extreme learning neural network to perform graph-based semi-supervised learning, taking into account the restriction of graph structure by graph convolutional neural network based on popular regularization and the advantage that graph learning convolutional neural network can adapt to the high-dimensional popular structure of graph data with multiple features, thereby realizing the improvement of image data processing and learning capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in the present disclosure or related technologies, the drawings required for use in the embodiments or related technical descriptions are briefly introduced below. Obviously, the drawings described below are only embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0062] Figure 1 A flowchart of a semi-supervised learning method according to an embodiment of the present disclosure;

[0063] Figure 2 A schematic diagram of a semi-supervised learning system according to an embodiment of the present disclosure;

[0064] Figure 3 It is a schematic diagram of the structure of an electronic device according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0065] In order to make the objectives, technical solutions and advantages of the present disclosure more clearly understood, the present disclosure is further described in detail below in combination with specific embodiments and with reference to the accompanying drawings.

[0066] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the embodiments of the present disclosure should be understood by people with ordinary skills in the field to which the present disclosure belongs. The "first", "second" and similar words used in the embodiments of the present disclosure do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprising" and similar words mean that the elements or objects appearing before the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects.

[0067] As described in the background technology section, existing semi-supervised learning methods and systems are still difficult to meet actual engineering needs.

[0068] In the process of implementing the present disclosure, the applicant discovered that the main problem with existing semi-supervised learning methods is that they cannot simultaneously take into account the characteristics of adapting to high-dimensional manifold structures and restricting graph structures to avoid algorithm overfitting.

[0069] Existing graph-based semi-supervised learning methods include: GCNMR (graph convolutional neural network based on manifold regularization) semi-supervised learning, and GLCN (graph learning convolutional neural network) semi-supervised learning. Among them, the GCNMR semi-supervised learning method for establishing graph structure adopts the K-nearest neighbor method based on Euclidean distance. Although this method has a simple structure, under actual engineering conditions, since the data usually has multiple features, in high-dimensional structures, the data structure is usually a manifold structure, it cannot meet actual engineering needs; the current GLCN semi-supervised learning does not restrict the graph structure, which can easily lead to overfitting of the algorithm and affect the performance of the GLCN network.

[0070] The technical method of the embodiment of the present disclosure is described in detail below through specific embodiments and in combination with a flowchart of the semi-supervised learning method.

[0071] refer to Figure 1 The semi-supervised learning method of an embodiment of the present disclosure optimizes the affinity matrix, label matrix and extreme learning neural network parameter matrix, and specifically includes the following steps:

[0072] Step S101, forming an image data matrix from image data, and normalizing it, obtaining the normalized Euclidean distance of the image data, forming the Euclidean distance into a distance matrix, and inputting it into a single-layer neural network.

[0073] In some embodiments, the selected data set includes 5 categories of pet images, namely Abyssinian cats, American bulldogs, pit bulls, basset hounds and beagles, with a total of 998 images, including labeled data and unlabeled data, and there are approximately 200 pet images of each category; each category of labeled data represents a category of pet images, and each category of labeled data contains no less than 1 data sample, and its goal is to divide pet images according to different breeds.

[0074] First, the image dataset in the database is preprocessed, including:

[0075] First, all images are modified into images with uniform pixels, and all image data are combined into an image data matrix X, which is used as the input of the graph convolutional layer neural network in the following section;

[0076] In this embodiment, all images may be modified into images of 128*128 pixels.

[0077] Further, all pet images are combined into an image data matrix X;

[0078] Further, the label matrix F is established using the image data, and the value Y of the label matrix established using the image data is used as the initialization value when the label matrix F is used for the first iteration operation in the following part, that is, Y = F(0);

[0079] Furthermore, according to the different categories to which the image data belongs, one-hot encoding is used for the label data in all image data as the label of this type of data in the label matrix Y. For unlabeled data, it is set to a vector of 0 in the label matrix Y.

[0080] Furthermore, each image data in the image data matrix X is pre-processed for normalization. In this embodiment, the specific normalization method used is not limited, and the Min-Max normalization method shown in the following formula can be selected:

[0081]

[0082] Among them, x * is the normalized value after normalization, max represents the maximum value of all data in a certain feature, min represents the minimum value of all data in a certain feature, and x represents the value represented by a certain feature in a certain data;

[0083] Alternatively, a normalization method such as Z-Score can be used for calculation.

[0084] Further, the Euclidean distance of the image data in the image data matrix X is calculated, the Euclidean distance is formed into a distance matrix, and the distance matrix is ​​sent to the single-layer neural network as the input of the single-layer neural network.

[0085] Step S102: Instruct the single-layer neural network to use the received distance matrix to generate an affinity matrix according to a preset activation function, and optimize the affinity matrix by iterating a preset cost function.

[0086] In some embodiments, the distance matrix is ​​received by a single-layer neural network, where the single-layer neural network is composed of a plurality of neurons.

[0087] Furthermore, the activation function shown in the following formula is preset in the single-layer neural network:

[0088]

[0089] Among them, x i and x j is the element in the image data matrix X, a is the weight vector in the single-layer neural network, Relu() represents the linear rectification function, and n is the number of neurons in the single-layer neural network.

[0090] Furthermore, according to the rules of the above activation function, the distance matrix in the single-layer neural network is mapped to an affinity matrix and represented by S.

[0091] Furthermore, in the semi-supervised learning framework of this embodiment, a cost function shown in the following formula is preset:

[0092]

[0093] Among them, L GLCNMR-ELM represents the cost function, F represents the label matrix, S represents the affinity matrix, M represents the distance matrix, O represents the output of the hidden layer neurons in the extreme learning neural network, X represents the image data matrix, L represents the Laplace matrix, λ and β are the preset hyperparameters, and y represents the label of the image data.

[0094] Furthermore, when optimizing different matrices, different items can be selected in the above optimization function to be combined for iterative calculation;

[0095] In some embodiments, for the optimization of the affinity matrix in a single-layer neural network, a combination of the second, third, and fourth function terms in the above cost function formula can be selected as a first combination of the cost function for optimizing the affinity matrix S. The specific formula is as follows:

[0096]

[0097] Specifically, the label matrix F is fixed, wherein when the affinity matrix S is optimized for the first time, the initialization value Y of the obtained label matrix F is used, that is, Y=F(0);

[0098] Furthermore, the extreme learning neural network parameter matrix W is fixed, wherein when the affinity matrix S is optimized for the first time, the number of hidden layer neurons can be initialized to K, and the parameter matrix W from the input layer to the hidden layer of the extreme learning neural network and the bias vector therein can be randomly generated.

[0099] Furthermore, the above formula is iterated using a gradient descent algorithm;

[0100] Furthermore, the optimized affinity matrix S is input into the graph convolutional layer neural network.

[0101] Step S103: combine the image data matrix with the affinity matrix, obtain the label matrix output by the graph convolution layer according to a preset convolution function, and optimize the label function by iterating the preset cost function.

[0102] In some embodiments, the image data matrix X established in the above section is input into a graph convolutional layer neural network;

[0103] Wherein, at least one graph convolution layer is set in the graph convolution layer neural network. In this embodiment, two graph convolution layers are set, namely a first graph convolution layer and a second graph convolution layer;

[0104] Furthermore, the first convolution function of the following formula is preset in the first graph convolution layer:

[0105] H (1) =f(SXZ (1) )

[0106] Among them, S represents the affinity matrix, X represents the image data matrix, Z is the graph convolution layer parameter matrix, and l represents the order of the convolution layer.

[0107] In the second graph convolution layer, the second convolution function with the following formula is preset:

[0108] H (l+1) =f(SH (l) Z (l+1) )

[0109] Among them, S represents the affinity matrix, H represents the output of the previous graph convolutional layer, Z is the graph convolutional layer parameter matrix, and l represents the order of the convolutional layer.

[0110] Furthermore, the graph convolutional neural network receives the affinity matrix S obtained in the above part, and combines it with the image data matrix X, and sequentially performs convolution operations of the above graph convolutional layers to obtain the output matrix of the last graph convolutional layer;

[0111] Furthermore, since the output values ​​of each element in the output matrix of the graph convolution layer are not in the form of probability values, it is necessary to send the output matrix of the graph convolution layer to the softmax classifier, and use the softmax classifier to map the output values ​​of each element in the output matrix of the graph convolution layer to probability values;

[0112] Furthermore, the matrix composed of the above probability values ​​is used as the label matrix F.

[0113] In some embodiments, the above cost function is further used to optimize the label matrix F. Specifically, a combination of the first, second, and fifth function terms in the above cost function formula can be selected as a second combination of the cost function for optimizing the label matrix F. The specific formula is as follows:

[0114]

[0115] Specifically, the affinity matrix S is fixed, and the extreme learning neural network parameter matrix W is fixed, and the above formula is iterated using the gradient descent algorithm.

[0116] Furthermore, the optimized label matrix F is output to the extreme learning neural network in the next part.

[0117] Step S104: sending the label matrix and the image data matrix output by the convolutional layer to the extreme learning neural network, and optimizing the extreme learning neural network parameters by iterating the preset cost function.

[0118] In some embodiments, the label matrix F output by the softmax classifier is received by the hidden layer neurons in the extreme learning neural network, and the extreme learning neural network parameter matrix W is optimized, wherein the hyperparameter matrix of the extreme learning neural network includes the number of hidden layer neurons.

[0119] Furthermore, when optimizing the parameter matrix W of the extreme learning neural network, the fifth function item can be selected in the cost function as the third combination of the cost function for optimization, and expressed as the following formula:

[0120]

[0121] Wherein, W is the extreme learning neural network parameter matrix, F represents the label matrix, and O represents the output of the hidden layer neurons in the extreme learning neural network;

[0122] Specifically, the affinity matrix S and the label matrix F are fixed, and the optimization objective is set to a strictly convex function. At this time, the minimum value of the extreme learning neural network parameter matrix W has a unique solution; and the partial derivative of W in the above formula is calculated, and the result after the partial derivative is further equal to 0, which is specifically expressed as the following formula:

[0123]

[0124] Furthermore, the value of W can be obtained.

[0125] Furthermore, it is judged whether the complete cost function converges. When the cost function converges, it indicates that the optimization training of the semi-supervised learning framework has been completed, and the obtained extreme learning neural network parameter matrix W can be used to generate an extreme learning neural network model.

[0126] And taking the label matrix F as the result of label diffusion, the extreme learning neural network model and the label matrix F are output.

[0127] Furthermore, if the cost function does not show convergence, it means that the optimization training of the semi-supervised learning framework is not completed, and the above optimization training process of the affinity matrix S, the label matrix F and the extreme learning neural network parameter matrix W is repeated until the complete cost function shows convergence.

[0128] Furthermore, the obtained parameter matrix W of the extreme learning neural network is deployed in the network, and the obtained label matrix is ​​used as the label of the data set containing the current image data, and the label is used to classify the image data to divide the pictures into different breeds of pets.

[0129] In another method embodiment of the present disclosure, a dataset of face images is used to illustrate the method of semi-supervised learning in the present disclosure:

[0130] Specifically, the YALE-B face dataset includes facial photos of 38 people, each of whom has 64 shooting effects under different lighting. Its goal is to use the Pokémon neural network to identify and classify the 38 people in the YALE-B face dataset.

[0131] First, the image dataset in the database is preprocessed, including:

[0132] First, all images are modified into images with uniform pixels, and all image data are combined into an image data matrix X, which is used as the input of the graph convolutional layer neural network in the following section;

[0133] In this embodiment, all images may be modified into images of 128*128 pixels.

[0134] Further, all pet images are combined into an image data matrix X;

[0135] Further, the label matrix F is established using the image data, and the value Y of the label matrix established using the image data is used as the initialization value when the label matrix F is used for the first iteration operation in the following part, that is, Y = F(0);

[0136] Furthermore, according to the different categories to which the image data belongs, one-hot encoding is used for the label data in all image data as the label of this type of data in the label matrix Y. For unlabeled data, it is set to a vector of 0 in the label matrix Y.

[0137] Furthermore, each image data in the image data matrix X is pre-processed for normalization. In this embodiment, the specific normalization method used is not limited, and the Min-Max normalization method shown in the following formula can be selected:

[0138]

[0139] Among them, x * is the normalized value after normalization, max represents the maximum value of all data in a certain feature, min represents the minimum value of all data in a certain feature, and x represents the value represented by a certain feature in a certain data;

[0140] Alternatively, a normalization method such as Z-Score can be used for calculation.

[0141] Further, the Euclidean distance of the image data in the image data matrix X is calculated, the Euclidean distance is formed into a distance matrix, and the distance matrix is ​​sent to the single-layer neural network as the input of the single-layer neural network.

[0142] Furthermore, the activation function shown in the following formula is preset in the single-layer neural network:

[0143]

[0144] Among them, x i and x j is the element in the image data matrix X, a is the weight vector in the single-layer neural network, Relu() represents the linear rectification function, and n is the number of neurons in the single-layer neural network.

[0145] Furthermore, according to the rules of the above activation function, the distance matrix in the single-layer neural network is mapped to an affinity matrix and represented by S.

[0146] Furthermore, in the semi-supervised learning framework of this embodiment, a cost function shown in the following formula is preset:

[0147]

[0148] Among them, L GLCNMR-ELM represents the cost function, F represents the label matrix, S represents the affinity matrix, M represents the distance matrix, O represents the output of the hidden layer neurons in the extreme learning neural network, X represents the image data matrix, L represents the Laplace matrix, λ and β are the preset hyperparameters, and y represents the label of the image data.

[0149] Furthermore, when optimizing different matrices, different items can be selected in the above optimization function to be combined for iterative calculation;

[0150] In some embodiments, for the optimization of the affinity matrix in a single-layer neural network, a combination of the second, third, and fourth function terms in the above cost function formula can be selected as a first combination of the cost function for optimizing the affinity matrix S. The specific formula is as follows:

[0151]

[0152] Specifically, the label matrix F is fixed, wherein when the affinity matrix S is optimized for the first time, the initialization value Y of the obtained label matrix F is used, that is, Y=F(0);

[0153] Furthermore, the extreme learning neural network parameter matrix W is fixed, wherein when the affinity matrix S is optimized for the first time, the number of hidden layer neurons can be initialized to K, and the parameter matrix W from the input layer to the hidden layer of the extreme learning neural network and the bias vector therein can be randomly generated.

[0154] Furthermore, the above formula is iterated using a gradient descent algorithm;

[0155] Furthermore, the optimized affinity matrix S is input into the graph convolutional layer neural network.

[0156] In some embodiments, the image data matrix X established in the above section is input into a graph convolutional layer neural network;

[0157] Wherein, at least one graph convolution layer is set in the graph convolution layer neural network. In this embodiment, two graph convolution layers are set, namely a first graph convolution layer and a second graph convolution layer;

[0158] Furthermore, the first convolution function of the following formula is preset in the first graph convolution layer:

[0159] H (1) =f(SXZ (1) )

[0160] Among them, S represents the affinity matrix, X represents the image data matrix, Z is the graph convolution layer parameter matrix, and l represents the order of the convolution layer.

[0161] In the second graph convolution layer, the second convolution function with the following formula is preset:

[0162] H (l+1) =f(SH (l) Z (l+1) )

[0163] Among them, S represents the affinity matrix, H represents the output of the previous graph convolutional layer, Z is the graph convolutional layer parameter matrix, and l represents the order of the convolutional layer.

[0164] Furthermore, the graph convolutional neural network receives the affinity matrix S obtained in the above part, and combines it with the image data matrix X, and sequentially performs convolution operations of the above graph convolutional layers to obtain the output matrix of the last graph convolutional layer;

[0165] Furthermore, since the output values ​​of each element in the output matrix of the graph convolution layer are not in the form of probability values, it is necessary to send the output matrix of the graph convolution layer to the softmax classifier, and use the softmax classifier to map the output values ​​of each element in the output matrix of the graph convolution layer to probability values;

[0166] Furthermore, the matrix composed of the above probability values ​​is used as the label matrix F.

[0167] In some embodiments, the above cost function is further used to optimize the label matrix F. Specifically, a combination of the first, second, and fifth function terms in the above cost function formula can be selected as a second combination of the cost function for optimizing the label matrix F. The specific formula is as follows:

[0168]

[0169] Specifically, the affinity matrix S is fixed, and the extreme learning neural network parameter matrix W is fixed, and the above formula is iterated using the gradient descent algorithm.

[0170] Furthermore, the optimized label matrix F is output to the extreme learning neural network in the next part.

[0171] In some embodiments, the label matrix F output by the softmax classifier is received by the hidden layer neurons in the extreme learning neural network, and the extreme learning neural network parameter matrix W is optimized, wherein the hyperparameter matrix of the extreme learning neural network includes the number of hidden layer neurons.

[0172] Furthermore, when optimizing the parameter matrix W of the extreme learning neural network, the fifth function item can be selected in the cost function as the third combination of the cost function for optimization, and expressed as the following formula:

[0173]

[0174] Wherein, W is the extreme learning neural network parameter matrix, F represents the label matrix, and O represents the output of the hidden layer neurons in the extreme learning neural network;

[0175] Specifically, the affinity matrix S and the label matrix F are fixed, and the optimization objective is set to a strictly convex function. At this time, the minimum value of the extreme learning neural network parameter matrix W has a unique solution; and the partial derivative of W in the above formula is calculated, and the result after the partial derivative is further equal to 0, which is specifically expressed as the following formula:

[0176]

[0177] Furthermore, the value of W can be obtained.

[0178] Furthermore, it is judged whether the complete cost function converges. When the cost function converges, it indicates that the optimization training of the semi-supervised learning framework has been completed, and the obtained extreme learning neural network parameter matrix W can be used to generate an extreme learning neural network model.

[0179] And taking the label matrix F as the result of label diffusion, the extreme learning neural network model and the label matrix F are output.

[0180] Furthermore, if the cost function does not show convergence, it means that the optimization training of the semi-supervised learning framework is not completed, and the above optimization training process of the affinity matrix S, the label matrix F and the extreme learning neural network parameter matrix W is repeated until the complete cost function shows convergence.

[0181] Furthermore, the obtained parameter matrix W of the extreme learning neural network is deployed in the network, and the obtained label matrix is ​​used as the label containing the current YALE-B face dataset, and the label is used to classify the face image data therein to divide the images according to different individuals.

[0182] It can be seen that the semi-supervised learning method of the embodiment of the present invention is based on popular regularization and graph learning convolutional neural network, and comprehensively considers the mechanism of extreme learning neural network to perform graph-based semi-supervised learning, taking into account the restriction of graph structure by graph convolutional neural network based on popular regularization and the advantage that graph learning convolutional neural network can adapt to the high-dimensional popular structure of graph data with multiple features, thereby realizing the improvement of image data processing and learning capabilities.

[0183] It should be noted that the method of the embodiment of the present disclosure can be performed by a single device, such as a computer or a server. The method of the present embodiment can also be applied in a distributed scenario and completed by multiple devices cooperating with each other. In the case of such a distributed scenario, one of the multiple devices can only perform one or more steps in the method of the embodiment of the present disclosure, and the multiple devices will interact with each other to complete the described method.

[0184] It should be noted that the above describes some embodiments of the present disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the above embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0185] Based on the same inventive concept, corresponding to any of the above-mentioned embodiments and methods, an embodiment of the present disclosure further provides a semi-supervised learning system.

[0186] refer to Figure 2 , the semi-supervised learning system includes: an image data preprocessing layer, a single-layer neural network, a graph convolutional layer neural network and an extreme learning neural network;

[0187] The image data preprocessing layer S201 is configured to: form an image data matrix from the image data, normalize it, obtain the Euclidean distance of the normalized image data, form a distance matrix with the Euclidean distance, and input it into a single-layer neural network;

[0188] The single-layer neural network S202 is configured to: generate an affinity matrix using the received distance matrix according to a preset activation function, and optimize the affinity matrix using iteration of a preset cost function;

[0189] The graph convolutional layer neural network S203 is configured to: combine the image data matrix with the affinity matrix, obtain a label matrix output by the graph convolutional layer according to a preset convolution function, and optimize the label function by iterating the preset cost function;

[0190] The extreme learning neural network S204 is configured to: send the label matrix and the image data matrix output by the convolution layer to the extreme learning neural network, and optimize the extreme learning neural network parameter matrix by iterating the preset cost function.

[0191] As an optional embodiment, the image data preprocessing layer S201 is specifically configured as follows:

[0192] In some embodiments, the selected data set includes 5 categories of pet images, namely Abyssinian cats, American bulldogs, pit bulls, basset hounds and beagles, with a total of 998 images, including labeled data and unlabeled data, and there are approximately 200 pet images of each category; each category of labeled data represents a category of pet images, and each category of labeled data contains no less than 1 data sample, and its goal is to divide pet images according to different breeds.

[0193] First, the image dataset in the database is preprocessed, including:

[0194] First, all images are modified into images with uniform pixels, and all image data are combined into an image data matrix X, which is used as the input of the graph convolutional layer neural network in the following section;

[0195] In this embodiment, all images may be modified into images of 128*128 pixels.

[0196] Further, all pet images are combined into an image data matrix X;

[0197] Further, the label matrix F is established using the image data, and the value Y of the label matrix established using the image data is used as the initialization value when the label matrix F is used for the first iteration operation in the following part, that is, Y = F(0);

[0198] Furthermore, according to the different categories to which the image data belongs, one-hot encoding is used for the label data in all image data as the label of this type of data in the label matrix Y. For unlabeled data, it is set to a vector of 0 in the label matrix Y.

[0199] Furthermore, each image data in the image data matrix X is pre-processed for normalization. In this embodiment, the specific normalization method used is not limited, and the Min-Max normalization method shown in the following formula can be selected:

[0200]

[0201] Among them, x * is the normalized value after normalization, max represents the maximum value of all data in a certain feature, min represents the minimum value of all data in a certain feature, and x represents the value represented by a certain feature in a certain data;

[0202] Alternatively, a normalization method such as Z-Score can be used for calculation.

[0203] Further, the Euclidean distance of the image data in the image data matrix X is calculated, the Euclidean distance is formed into a distance matrix, and the distance matrix is ​​sent to the single-layer neural network as the input of the single-layer neural network.

[0204] As an optional embodiment, the single-layer neural network S202 is specifically configured as follows:

[0205] In some embodiments, the distance matrix is ​​received by a single-layer neural network, where the single-layer neural network is composed of a plurality of neurons.

[0206] Furthermore, the activation function shown in the following formula is preset in the single-layer neural network:

[0207]

[0208] Among them, x i and x j is the element in the image data matrix X, a is the weight vector in the single-layer neural network, Relu() represents the linear rectification function, and n is the number of neurons in the single-layer neural network.

[0209] Furthermore, according to the rules of the above activation function, the distance matrix in the single-layer neural network is mapped to an affinity matrix and represented by S.

[0210] Furthermore, in the semi-supervised learning framework of this embodiment, a cost function shown in the following formula is preset:

[0211]

[0212] Among them, L GLCNMR-ELM represents the cost function, F represents the label matrix, S represents the affinity matrix, M represents the distance matrix, O represents the output of the hidden layer neurons in the extreme learning neural network, X represents the image data matrix, L represents the Laplace matrix, λ and β are the preset hyperparameters, and y represents the label of the image data.

[0213] Furthermore, when optimizing different matrices, different items can be selected in the above optimization function to be combined for iterative calculation;

[0214] In some embodiments, for the optimization of the affinity matrix in a single-layer neural network, a combination of the second, third, and fourth function terms in the above cost function formula can be selected as a first combination of the cost function for optimizing the affinity matrix S. The specific formula is as follows:

[0215]

[0216] Specifically, the label matrix F is fixed, wherein when the affinity matrix S is optimized for the first time, the initialization value Y of the obtained label matrix F is used, that is, Y=F(0);

[0217] Furthermore, the extreme learning neural network parameter matrix W is fixed, wherein when the affinity matrix S is optimized for the first time, the number of hidden layer neurons can be initialized to K, and the parameter matrix W from the input layer to the hidden layer of the extreme learning neural network and the bias vector therein can be randomly generated.

[0218] Furthermore, the above formula is iterated using a gradient descent algorithm;

[0219] Furthermore, the optimized affinity matrix S is input into the graph convolutional layer neural network.

[0220] As an optional embodiment, the graph convolutional layer neural network S203 is specifically configured as follows:

[0221] In some embodiments, the image data matrix X established in the above section is input into a graph convolutional layer neural network;

[0222] Wherein, at least one graph convolution layer is set in the graph convolution layer neural network. In this embodiment, two graph convolution layers are set, namely a first graph convolution layer and a second graph convolution layer;

[0223] Furthermore, the first convolution function of the following formula is preset in the first graph convolution layer:

[0224] H (1) =f(SXZ (1) )

[0225] Among them, S represents the affinity matrix, X represents the image data matrix, Z is the graph convolution layer parameter matrix, and l represents the order of the convolution layer.

[0226] In the second graph convolution layer, the second convolution function with the following formula is preset:

[0227] H (l+1) =f(SH (l) Z (l+1) )

[0228] Among them, S represents the affinity matrix, H represents the output of the previous graph convolutional layer, Z is the graph convolutional layer parameter matrix, and l represents the order of the convolutional layer.

[0229] Furthermore, the graph convolutional neural network receives the affinity matrix S obtained in the above part, and combines it with the image data matrix X, and sequentially performs convolution operations of the above graph convolutional layers to obtain the output matrix of the last graph convolutional layer;

[0230] Furthermore, since the output values ​​of each element in the output matrix of the graph convolution layer are not in the form of probability values, it is necessary to send the output matrix of the graph convolution layer to the softmax classifier, and use the softmax classifier to map the output values ​​of each element in the output matrix of the graph convolution layer to probability values;

[0231] Furthermore, the matrix composed of the above probability values ​​is used as the label matrix F.

[0232] In some embodiments, the above cost function is further used to optimize the label matrix F. Specifically, a combination of the first, second, and fifth function terms in the above cost function formula can be selected as a second combination of the cost function for optimizing the label matrix F. The specific formula is as follows:

[0233]

[0234] Specifically, the affinity matrix S is fixed, and the extreme learning neural network parameter matrix W is fixed, and the above formula is iterated using the gradient descent algorithm.

[0235] Furthermore, the optimized label matrix F is output to the extreme learning neural network in the next part.

[0236] As an optional embodiment, the extreme learning neural network S204 is specifically configured as follows:

[0237] In some embodiments, the label matrix F output by the softmax classifier is received by the hidden layer neurons in the extreme learning neural network, and the extreme learning neural network parameter matrix W is optimized, wherein the hyperparameter matrix of the extreme learning neural network includes the number of hidden layer neurons.

[0238] Furthermore, when optimizing the parameter matrix W of the extreme learning neural network, the fifth function item can be selected in the cost function as the third combination of the cost function for optimization, and expressed as the following formula:

[0239]

[0240] Wherein, W is the extreme learning neural network parameter matrix, F represents the label matrix, and O represents the output of the hidden layer neurons in the extreme learning neural network;

[0241] Specifically, the affinity matrix S and the label matrix F are fixed, and the optimization objective is a strictly convex function. At this time, the minimum value of the extreme learning neural network parameter matrix W has a unique solution; therefore, the partial derivative of W in the above formula is taken, and the result after the partial derivative is further equal to 0, then the optimal solution of W can be obtained, which is specifically expressed as the following formula:

[0242]

[0243] Furthermore, the value of W can be obtained.

[0244] Furthermore, it is judged whether the complete cost function converges. When the cost function converges, it indicates that the optimization training of the semi-supervised learning framework has been completed, and the obtained extreme learning neural network parameter matrix W can be used to generate an extreme learning neural network model.

[0245] And taking the label matrix F as the result of label diffusion, the extreme learning neural network model and the label matrix F are output.

[0246] Furthermore, if the cost function does not show convergence, it means that the optimization training of the semi-supervised learning framework is not completed, and the above optimization training process of the affinity matrix S, the label matrix F and the extreme learning neural network parameter matrix W is repeated until the complete cost function shows convergence.

[0247] Furthermore, the obtained parameter matrix W of the extreme learning neural network is deployed in the network, and the obtained label matrix is ​​used as the label of the data set containing the current image data, and the label is used to classify the image data to divide the pictures into different breeds of pets.

[0248] For the convenience of description, the above devices are described in terms of functions and are divided into various modules. Of course, when implementing the embodiments of the present disclosure, the functions of each module can be implemented in the same or multiple software and / or hardware.

[0249] The device of the above embodiment is used to implement the corresponding semi-supervised learning method in any of the above embodiments, and has the beneficial effects of the corresponding method embodiment, which will not be repeated here.

[0250] Based on the same inventive concept, corresponding to any of the above-mentioned embodiments and methods, an embodiment of the present disclosure also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the semi-supervised learning method as described in any of the above embodiments is implemented.

[0251] Figure 3 A more specific schematic diagram of the hardware structure of an electronic device provided in this embodiment is shown, and the device may include: a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040 are connected to each other through the bus 1050 in the device.

[0252] The processor 1010 can be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present disclosure.

[0253] The memory 1020 may be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 may store an operating system and other application programs. When the technical solution provided by the embodiment of the present disclosure is implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0254] The input / output interface 1030 is used to connect the input / output module to realize information input and output. The input / output module can be configured in the device as a component (not shown in the figure), or it can be externally connected to the device to provide corresponding functions. The input device may include a keyboard, a mouse, a touch screen, a microphone, various sensors, etc., and the output device may include a display, a speaker, a vibrator, an indicator light, etc.

[0255] The communication interface 1040 is used to connect a communication module (not shown) to realize communication interaction between the device and other devices. The communication module can realize communication through a wired mode (such as USB, network cable, etc.) or a wireless mode (such as mobile network, WIFI, Bluetooth, etc.).

[0256] The bus 1050 includes a path that transmits information between the various components of the device (eg, the processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040).

[0257] It should be noted that, although the above device only shows the processor 1010, the memory 1020, the input / output interface 1030, the communication interface 1040 and the bus 1050, in the specific implementation process, the device may also include other components necessary for normal operation. In addition, it can be understood by those skilled in the art that the above device may also only include the components necessary for implementing the embodiments of the present disclosure, and does not necessarily include all the components shown in the figure.

[0258] The device of the above embodiment is used to implement the corresponding semi-supervised learning method in any of the above embodiments, and has the beneficial effects of the corresponding method embodiment, which will not be repeated here.

[0259] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present disclosure (including the claims) is limited to these examples. Based on the concept of the present disclosure, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of different aspects of the embodiments of the present disclosure as described above, which are not provided in detail for the sake of simplicity.

[0260] In addition, to simplify the description and discussion, and in order not to make the embodiments of the present disclosure difficult to understand, the known power / ground connections to the integrated circuit (IC) chips and other components may or may not be shown in the provided figures. In addition, the device can be shown in the form of a block diagram to avoid making the embodiments of the present disclosure difficult to understand, and this also takes into account the fact that the details of the implementation of these block diagram devices are highly dependent on the platform on which the embodiments of the present disclosure will be implemented (that is, these details should be fully within the scope of understanding of those skilled in the art). Where specific details (e.g., circuits) are set forth to describe exemplary embodiments of the present disclosure, it is apparent to those skilled in the art that the embodiments of the present disclosure can be implemented without these specific details or with changes in these specific details. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0261] Although the present disclosure has been described in conjunction with specific embodiments of the present disclosure, many replacements, modifications and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may use the embodiments discussed.

[0262] The embodiments of the present disclosure are intended to cover all such substitutions, modifications and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present disclosure should be included in the scope of protection of the present disclosure.

Claims

1. A semi-supervised learning method, in which: include: The image data is formed into an image data matrix, and normalized, and the Euclidean distance of the normalized image data is obtained, and the Euclidean distance is formed into a distance matrix, and input into a single-layer neural network; The single-layer neural network generates an affinity matrix using the received distance matrix according to a preset activation function, and optimizes the affinity matrix using iteration of a preset cost function; Combine the image data matrix with the affinity matrix, obtain a label matrix output by the graph convolution layer according to a preset convolution function, and optimize the label matrix by iterating the preset cost function; Sending the label matrix and the image data matrix output by the convolutional layer to an extreme learning neural network, and optimizing the extreme learning neural network parameter matrix by iteration of the preset cost function, so as to classify the image data; The step of optimizing the affinity matrix by iterating a preset cost function includes: The label matrix and the extreme learning neural network parameter matrix are fixed, and a gradient descent algorithm is used to iterate the first combination of the cost function to optimize the affinity matrix; The optimizing the label matrix by iterating the preset cost function includes: The affinity matrix and the extreme learning neural network parameter matrix are fixed, and a gradient descent algorithm is used to iterate the second combination of the cost function to optimize the label matrix; The method of optimizing the extreme learning neural network parameter matrix by iterating the preset cost function includes: Fixing the affinity matrix and the label matrix, and optimizing the extreme learning neural network parameter matrix by using the third combination of the cost functions; and In response to determining that the cost function does not appear to converge, continuing optimization of the extreme learning neural network parameter matrix, the affinity matrix, and the label matrix; In response to determining that the cost function exhibits convergence, stopping the optimization of the extreme learning neural network parameter matrix, the affinity matrix, and the label matrix, and retaining the optimization results; The cost function is expressed as: in, represents the cost function, F represents the label matrix, S represents the affinity matrix, M represents the distance matrix, O represents the output of the hidden layer neurons in the extreme learning neural network, X represents the image data matrix, L represents the Laplace matrix, and is a preset hyperparameter, and y represents the label of the image data; wherein the affinity matrix is ​​optimized by using the first combination of the second function term, the third function term and the fourth function term in the cost function; Adopting a second combination consisting of a first function term, a second function term and a fifth function term in the cost function to optimize the label matrix; Adopting a third combination of fifth function terms in the cost function to optimize the extreme learning neural network parameter matrix; The fifth function term is used in the cost function to optimize the extreme learning neural network parameter matrix, including: Formula Take the partial derivative of W in the equation and set the derivative result equal to zero to get the value of W. Wherein, W is the extreme learning neural network parameter matrix, F represents the label matrix, and O represents the output of the hidden layer neurons in the extreme learning neural network; The adopting a gradient descent algorithm to the cost function to optimize the label matrix includes: Using the image data, a label matrix is ​​established as an initialization value for optimizing the label matrix; Wherein, the label data in the image data is encoded using unique hot encoding as the label of the label data; A vector that sets the unlabeled data in the image data to zero; The same data label represents the same category of the image data, and each category of the image data has no less than one data sample; Generate an affinity matrix based on the preset activation function, including: Using formula in, represents the affinity matrix, a represents the weight vector in a single-layer neural network, x i and x j Represents the elements in the image data matrix; As an activation function of the single-layer neural network, mapping the data matrix to the affinity matrix; The step of obtaining the label matrix output by the graph convolution layer according to a preset convolution function includes: In the graph convolution layer, the formula is used And combining the rule of the affinity matrix, mapping the image data matrix to the graph convolution layer output matrix; Wherein, H represents the output of the convolutional layer, S represents the affinity matrix, l represents the order of the convolutional layer, and Z represents the graph convolutional layer parameter matrix; The output values ​​of the output matrix of the graph convolution layer are mapped into probability values ​​using a softmax classifier, and used as the label matrix; Wherein, at least one convolutional layer is set in the graph convolutional layer neural network.

2. A semi-supervised learning system, comprising: Image data preprocessing layer, single-layer neural network, graph convolutional layer neural network and extreme learning neural network; The image data preprocessing layer is configured to: form an image data matrix from the image data, normalize it, obtain the normalized Euclidean distance of the image data, form a distance matrix with the Euclidean distance, and input it into a single-layer neural network; The single-layer neural network is configured to: generate an affinity matrix using the received distance matrix according to a preset activation function, and optimize the affinity matrix using iteration of a preset cost function; The graph convolutional layer neural network is configured to: combine the image data matrix with the affinity matrix, obtain a label matrix output by the graph convolutional layer according to a preset convolution function, and optimize the label matrix by iterating the preset cost function; The extreme learning neural network is configured to: send the label matrix and the image data matrix output by the convolution layer to the extreme learning neural network, and optimize the extreme learning neural network parameter matrix by iteration of the preset cost function, so as to classify the image data; The step of optimizing the affinity matrix by iterating a preset cost function includes: The label matrix and the extreme learning neural network parameter matrix are fixed, and a gradient descent algorithm is used to iterate the first combination of the cost function to optimize the affinity matrix; The optimizing the label matrix by iterating the preset cost function includes: The affinity matrix and the extreme learning neural network parameter matrix are fixed, and a gradient descent algorithm is used to iterate the second combination of the cost function to optimize the label matrix; The method of optimizing the extreme learning neural network parameter matrix by iterating the preset cost function includes: Fixing the affinity matrix and the label matrix, and optimizing the extreme learning neural network parameter matrix by using the third combination of the cost functions; and In response to determining that the cost function does not appear to converge, continuing optimization of the extreme learning neural network parameter matrix, the affinity matrix, and the label matrix; In response to determining that the cost function exhibits convergence, stopping the optimization of the extreme learning neural network parameter matrix, the affinity matrix, and the label matrix, and retaining the optimization results; The cost function is expressed as: in, represents the cost function, F represents the label matrix, S represents the affinity matrix, M represents the distance matrix, O represents the output of the hidden layer neurons in the extreme learning neural network, X represents the image data matrix, L represents the Laplace matrix, and is a preset hyperparameter, and y represents the label of the image data; wherein the affinity matrix is ​​optimized by using the first combination of the second function term, the third function term and the fourth function term in the cost function; Adopting a second combination consisting of a first function term, a second function term and a fifth function term in the cost function to optimize the label matrix; Adopting a third combination of fifth function terms in the cost function to optimize the extreme learning neural network parameter matrix; The fifth function term is used in the cost function to optimize the extreme learning neural network parameter matrix, including: Formula Take the partial derivative of W in the equation and set the derivative result equal to zero to get the value of W. Wherein, W is the extreme learning neural network parameter matrix, F represents the label matrix, and O represents the output of the hidden layer neurons in the extreme learning neural network; The adopting a gradient descent algorithm to the cost function to optimize the label matrix includes: Using the image data, a label matrix is ​​established as an initialization value for optimizing the label matrix; Wherein, the label data in the image data is encoded using one-hot encoding as the label of the label data; A vector that sets the unlabeled data in the image data to zero; The same data label represents the same category of the image data, and each category of the image data has no less than one data sample; Generate an affinity matrix based on the preset activation function, including: Using formula in, represents the affinity matrix, a represents the weight vector in a single-layer neural network, x i and x j Represents the elements in the image data matrix; As an activation function of the single-layer neural network, mapping the data matrix to the affinity matrix; The step of obtaining the label matrix output by the graph convolution layer according to a preset convolution function includes: In the graph convolution layer, the formula is used And combining the rule of the affinity matrix, mapping the image data matrix to the graph convolution layer output matrix; Wherein, H represents the output of the convolutional layer, S represents the affinity matrix, l represents the order of the convolutional layer, and Z represents the graph convolutional layer parameter matrix; The output values ​​of the output matrix of the graph convolution layer are mapped into probability values ​​using a softmax classifier, and used as the label matrix; Wherein, at least one convolutional layer is set in the graph convolutional layer neural network.

3. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable by the processor, characterized in that: When the processor executes the computer program, the method according to claim 1 is implemented.