A Method for Constructing an Image Classification Model Based on Semi-Supervised Learning

Abstract

This invention discloses a method for constructing an image classification model based on semi-supervised learning, belonging to the field of image classification technology. The method assesses the risk coefficient of unlabeled image samples, enabling fine-grained evaluation of unlabeled images. Furthermore, it sets a higher risk coefficient for images with ambiguous feature information that are difficult for the model to distinguish, thus reducing their impact on model training. The method also ensures robust performance of the image classification model by using ensemble learning and a baseline model. Specifically, due to the uncertainty introduced during model training by the presence of unlabeled images, sampling techniques are used to construct multiple semi-supervised learning subsets from the original dataset. Multiple image classification models are then trained based on these subsets as base learners, creating a base learner pool. Ensemble learning is then used to generate the final image classification model from this pool, ensuring that the model's performance is no weaker than that of the baseline model trained from labeled image samples.

Description

Technical Field

[0001] This invention belongs to the field of image classification technology, and more specifically, relates to a method for constructing an image classification model based on semi-supervised learning. Background Technology

[0002] With the rapid development of artificial intelligence research, the data requirements for training models are gradually increasing. In the traditional field of machine learning, model training is usually divided into supervised learning, semi-supervised learning, and unsupervised learning. Among them, semi-supervised learning is a method used to train models with a small amount of labeled information, and it is also one of the key research directions in the field of artificial intelligence.

[0003] In semi-supervised learning research, it is generally believed that indiscriminate use of unlabeled samples will definitely improve the generalization performance of the model. However, some existing studies have confirmed that blindly using unlabeled samples may lead to the model being misled by some samples. On the other hand, due to the large number of unlabeled samples in semi-supervised learning, training a single model may introduce uncertainty, that is, multiple suboptimal models may be generated during training. Without a good model selection method, this may lead to a sharp decline in model performance, resulting in the model's performance being even weaker than the baseline model trained only using labeled samples from semi-supervised data.

[0004] In many special scenarios, such as medical image recognition, where the number of labeled images available in reality is limited, when using traditional semi-supervised models to train image classifiers, all unlabeled images are used for model training. Some of these unlabeled images are blurry, causing the semi-supervised model to misclassify them. This obviously deviates the model training process from the correct path, resulting in reduced model performance, or even weaker than the baseline model trained using only a small number of labeled images.

[0005] Therefore, how to provide a robust semi-supervised learning model training method that enables fine-grained evaluation and use of unlabeled images and reduces the impact of unlabeled images on model training is an urgent problem to be solved. Summary of the Invention

[0006] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a method for constructing an image classification model based on semi-supervised learning. Its purpose is to solve the technical problem that the existing semi-supervised learning model training methods cannot perform fine-grained evaluation and use of unlabeled images, resulting in the model being less robust.

[0007] To achieve the above objectives, according to one aspect of the present invention, a method for constructing an image classification model based on semi-supervised learning is provided, comprising:

[0008] S1: Calculate the similarity information between unlabeled image samples in the unlabeled image sample set and positive and negative samples in the labeled image sample set;

[0009] S2: Calculate the optimal positive transmission matrix from the unlabeled image sample set to the positive sample using the similarity information with the positive sample; calculate the optimal negative transmission matrix from the unlabeled image sample set to the negative sample using the similarity information with the negative sample;

[0010] S3: Calculate the risk coefficient of the unlabeled image sample set using the positive optimal transfer matrix and the negative optimal transfer matrix;

[0011] S4: Train a machine learning model using the risk coefficients of the semi-supervised training subset selected each time and the selected unlabeled image sample set; the semi-supervised training subset is composed of a certain number of samples sampled from the unlabeled image sample set and the labeled image sample set respectively based on a random sampling strategy;

[0012] S5: Construct a base learner pool using multiple machine learning models obtained through multiple training sessions;

[0013] S6: Integrate all models in the base learner pool to obtain an image classification model.

[0014] Further, S4 includes: training a decision boundary based on a semi-supervised optimal interval distribution learning machine using the semi-supervised training subset selected each time and the corresponding risk coefficient.

[0015] Furthermore, the objective function of the semi-supervised optimal interval distribution learning machine is:

[0016]

[0017]

[0018]

[0019] in, The model parameters to be solved are... The sample labels are i and j, where i and j are both sample numbers. To balance the model parameters for the i-th sample corresponding to the hyperparameters, This represents the semi-supervised training subset. The total number of samples in the semi-supervised training subset. For soft-interval hyperparameters, To balance the hyperparameters of interval mean and interval variance, To balance the hyperparameters of the loss, Let be the adjacency matrix of the nearest neighbor graph constructed on the semi-supervised training subset. for Normalized adjacency matrix, Let be the risk coefficient of the i-th unlabeled image sample. This is a baseline model trained based on labeled image samples. For the i-th sample, the superscript T represents the transpose. This is a kernel function mapping.

[0020] Furthermore, S4 includes: training a semi-supervised learning-based neural network classifier using multiple selected semi-supervised training subsets; and using the risk coefficient in the loss function or designing it as a sample weight term to guide the update iteration during the training process of the neural network classifier.

[0021] Further, S1 includes: using the similarity information between the unlabeled image sample set and the positive class samples as the positive transmission loss matrix. Unlabeled image samples are transmitted to the positive class samples; the optimal forward transmission matrix is ​​obtained by solving the first optimization problem corresponding to the forward transmission. ;

[0022] The similarity information between the unlabeled image sample set and the negative class samples is used as the negative transmission loss matrix. The unlabeled image samples are transmitted to the negative class samples; the second optimization problem corresponding to the negative transmission is solved to obtain the negative optimal transmission matrix. .

[0023] Furthermore, the first optimization problem and the second optimization problem are:

[0024]

[0025]

[0026] in, It is an identity matrix, with the superscript T indicating transpose. Represents the source domain. Indicates the target domain. To balance the hyperparameters of the two losses, This represents the trace operation of a matrix. This represents the entropy regularization term.

[0027] Further, S3 includes: using the formula

[0028]

[0029] Calculate the risk coefficient of each unlabeled image sample in the unlabeled image sample set. The distribution of labeled negative class samples, The distribution of labeled positive class samples; For the i-th unlabeled image sample With labeled image samples Entropy distance;

[0030]

[0031] in, Indicates labeled samples, where This indicates that the positive class samples have been labeled. Indicates that negative class samples have been labeled. Represents the set of unlabeled samples. This represents the element in the i-th row and j-th column of the i-th element in the optimal transfer matrix.

[0032] Further, S6 includes: based on ensemble learning, taking a convex combination of the machine learning models that have the best performance relative to the baseline model in the base learner pool as the image classification model; wherein the baseline model is trained based on the labeled image sample set.

[0033] Further, the step of using the convex combination of the machine learning models that have the best performance relative to the baseline model in the base learner pool as the image classification model includes: constructing a third optimization problem to maximize the performance improvement of the image classification model relative to the baseline model; solving the optimization problem to obtain the convex combination of the machine learning models, and using it as the image classification model; the third optimization problem is:

[0034]

[0035] in, This represents the baseline model. This represents the i-th machine learning model in the base learner pool. The weight, For the convex combination weights of the base learner, For a convex combination weight set, As the number of base learners, Let the hinge loss function be... It represents the 1-norm.

[0036] According to another aspect of the present invention, a semi-supervised learning-based image classification method is provided, comprising: performing image classification using the image classification model.

[0037] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0038] (1) This invention provides a method for constructing an image classification model based on semi-supervised learning. By judging the risk coefficient of unlabeled image samples, it is possible to perform fine-grained evaluation of unlabeled images, thereby reducing their impact on model training and improving the performance robustness of the image classification model. Due to the existence of unlabeled images, there will be uncertainty in the model training process, and the model obtained in each training may not be the optimal model. By using sampling techniques to construct multiple semi-supervised learning subsets from the original dataset, and training multiple image classification models based on the subsets as base learners, a base learner pool is created. Based on ensemble learning, the final image classification model is generated from the base learner pool, which can ensure that the model performance is not weaker than the baseline model trained from labeled image samples.

[0039] (2) In semi-supervised learning, unlabeled images are not always beneficial to the model training process. Since semi-supervised learning methods usually require labeling and assigning labels to unlabeled images, the model is prone to giving incorrect label information for some images with relatively vague information and high recognition difficulty, which leads to the model training process being misled. This scheme performs a preliminary analysis of unlabeled samples and calculates the similarity between unlabeled samples and positively and negatively labeled samples respectively, in preparation for the subsequent risk coefficient judgment. Then, by using the optimal transmission strategy, the unlabeled images are transmitted to known positively and negatively labeled images respectively to obtain the corresponding transmission matrices. By analyzing the information contained in the matrices, the risk coefficient of the unlabeled images is judged and assigned the corresponding risk coefficient, so that the model can be trained with unlabeled images in a differentiated manner during the training process, which can reduce the uncertainty of unlabeled images.

[0040] (3) Methods for constructing a base learner pool by training multiple high-quality, low-density decision boundaries based on semi-supervised optimal margin distribution learning machines, or by training multiple high-quality neural network classifiers based on margin distribution loss, include: training a semi-supervised image classifier as a base learner based on traditional semi-supervised learning model training methods, such as semi-supervised models based on SVM, semi-supervised models based on ODM, or semi-supervised models based on deep neural networks, etc., with the training data coming from a subset formed by random sampling from the original dataset. Compared to the traditional method of training only a single semi-supervised model, this method can better reduce the uncertainty impact of unlabeled image samples on the model during training. Each semi-supervised classifier performs well on its respective data distribution, and the final integrated model can combine the advantages of each model to improve the overall generalization performance.

[0041] (4) This scheme combines the models in the base learner pool in a convex manner so that the final model can achieve good classification performance across the entire distribution of semi-supervised data. The model integration weights are determined by maximizing the performance improvement of the model relative to the baseline model trained with labeled images, ensuring that the performance of the final semi-supervised image classifier is not weaker than that of the baseline model.

[0042] (5) This scheme collects all semi-supervised base learner models, and then designs a performance difference objective based on hinge loss among the baseline model trained on labeled image samples, the model of convex combination of semi-supervised base learners, and the potential optimal model. Specifically, it assumes that the optimal classifier is This means that for this optimal classifier, compared to the baseline classifier... That is, a classifier trained using supervised information has the greatest potential for improvement. Attached Figure Description

[0043] Figure 1 This is a flowchart of the method for constructing an image classification model based on semi-supervised learning provided in Embodiment 1 of the present invention;

[0044] Figure 2 This is a schematic diagram of the construction of the base learner pool based on the semi-supervised learning model provided in Embodiment 1 of the present invention;

[0045] Figure 3 This is a schematic diagram of an image classification model based on a semi-supervised base learner pool, provided in Embodiment 1 of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0047] This invention provides explanations for some terms and nouns.

[0048] Semi-supervised learning: This refers to a semi-supervised learning model whose performance is not worse than that of a model trained solely on labeled data in a semi-supervised learning environment. In traditional semi-supervised learning models, unlabeled data is typically used indiscriminately for model training. However, unlabeled data does not always improve model performance. In fact, some difficult-to-identify unlabeled data may mislead the model training process, leading to a sharp drop in the final model performance, even below that of a model trained solely on labeled data.

[0049] Optimal transport: a mathematical tool for calculating the difference between two distributions. The advantage of optimal transport lies in its ability to handle discrete, continuous, or both measures within the same framework. A measure is a quantitative concept used to describe size or size relationships in space; intuitively, it can be represented as mass, area, and volume, and abstractly as a discrete histogram or probability density function in a function space. Optimal transport between probability measures can be understood as establishing the lowest-cost transport scheme between masses in different distribution spaces. The cost here typically refers to the distance between the transported subsets, thus establishing a mapping or transfer between probability measures.

[0050] Ensemble learning is a machine learning method that combines multiple base learners according to certain rules to obtain models with stronger performance and generalization ability. Ensemble learning has been proven to significantly improve model performance not only in supervised learning, but also in unsupervised and semi-supervised learning, where it can significantly improve the predictive performance and robustness of models.

[0051] Example 1

[0052] like Figure 1 As shown, this embodiment provides a method for constructing an image classification model based on semi-supervised learning, including: S1-S6. S1: Obtain similarity information between labeled image sample sets and unlabeled image sample sets; S2: Calculate the optimal transmission matrix based on the optimal transmission loss theory using the similarity information; S3: Calculate the risk coefficient of the unlabeled image sample set using the optimal transmission matrix; S4: Sample a certain number of samples from both the unlabeled and labeled image sample sets using a random sampling strategy to form a semi-supervised training subset; Train a machine learning model using each selected semi-supervised training subset and its corresponding risk coefficient; S5: Construct a base learner pool using multiple machine learning models obtained from multiple training iterations; S6: Integrate all models in the base learner pool to obtain an image classification model.

[0053] in, Figure 2 This demonstrates the processing flow for constructing the base learning pool. Step 1: Divide the semi-supervised learning data into unlabeled, positively labeled, and negatively labeled datasets. Step 2: Randomly sample from each of the three datasets to construct multiple semi-supervised data subsets for subsequent base learner pool construction. Step 3: Based on the semi-supervised learning algorithm, train multiple base learners from these subsets to construct the base learner pool. Furthermore, during training, the risk coefficient of the unlabeled image samples is introduced into the training process to refine the model update process.

[0054] Furthermore, S4 includes: training a decision boundary based on a semi-supervised optimal margin distribution learning machine using each selected semi-supervised training subset and the corresponding risk coefficient.

[0055] Furthermore, the objective function based on the semi-supervised optimal margin distribution learning machine is:

[0056]

[0057]

[0058]

[0059] in, The model parameters to be solved are... The sample labels are i and j, where i and j are both sample numbers. To balance the model parameters for the i-th sample corresponding to the hyperparameters, This represents a subset of semi-supervised training. This represents the total number of samples in the semi-supervised training subset. For soft-interval hyperparameters, To balance the hyperparameters of interval mean and interval variance, To balance the hyperparameters of the loss, Let be the adjacency matrix of the nearest neighbor graph constructed on the semi-supervised training subset. for Normalized adjacency matrix, Let be the risk coefficient of the i-th unlabeled image sample. This is a baseline model trained based on labeled image samples. For the i-th sample, the superscript T represents the transpose. This is a kernel function mapping.

[0060] Specifically, taking a semi-supervised optimal margin distribution learning machine as the base learner as an example, in order to extract the correlation between the sample label space and the feature space, a model is constructed on the dataset. Nearest neighbor graph ,in Represents the dataset The sample points in This represents the edges formed between sample points, and the weights of the edges represent the similarity between the samples. Based on a graph... Set the adjacency matrix , where if the sample and samples If an edge exists, then ,otherwise Furthermore, let Let represent the normalized adjacency matrix, where , It can be used to reflect the correlation information between samples. To further improve the safety of the base learner and reduce the impact of high-risk samples on the model, the following regularization term is proposed:

[0061]

[0062] in This represents the optimal margin distribution learning machine for supervised learning, which is based on a set of labeled image samples. It is derived from training. Combining the regularization term and the adjacency matrix term with the optimal margin distribution learning machine yields the following objective expression:

[0063]

[0064]

[0065]

[0066] in, , and To balance the hyperparameters, the empirical loss is used to adjust the labeled and unlabeled image samples. It is the risk assessment value of the unlabeled image sample obtained through optimal transmission.

[0067] Furthermore, S4 includes: training a neural network classifier based on semi-supervised basis learning using multiple selected semi-supervised training subsets; and incorporating the risk coefficient into the loss function or designing it as a sample weight term to guide the update iteration during the training process of the neural network classifier. During training, the model update process is corrected by introducing the risk coefficient of unlabeled image samples into the training.

[0068] Further, S1 includes: dividing the labeled image samples into positive class samples and negative class samples; calculating the similarity information between the unlabeled image samples in the unlabeled image sample set and the positive class samples, and the similarity information between the unlabeled image samples and the negative class samples. S2 includes: using the similarity information with the positive class samples as the forward transmission loss matrix. Unlabeled image samples are transmitted to positive class samples; the optimal forward transmission matrix is ​​obtained by solving the first optimization problem corresponding to the forward transmission. The similarity information with the negative class samples is used as the negative transmission loss matrix. The unlabeled image samples are transferred to the negative class samples; the optimal negative transfer matrix is ​​obtained by solving the second optimization problem corresponding to the negative transfer. .

[0069] Furthermore, the first optimization problem and the second optimization problem are:

[0070]

[0071]

[0072] in, It is an identity matrix, with the superscript T indicating transpose. Represents the source domain. Indicates the target domain. To balance the hyperparameters of the two losses, This represents the trace operation of a matrix. This represents the entropy regularization term. Further, using the formula...

[0073]

[0074] Calculate the risk coefficient of each unlabeled image sample in the unlabeled image sample set. The distribution of labeled negative class samples, The distribution of labeled positive class samples; For the i-th unlabeled image sample With labeled image samples Entropy distance:

[0075]

[0076] in, Indicates labeled samples, where This indicates that the positive class samples have been labeled. This indicates that the negative class samples have been labeled. Represents the set of unlabeled image samples. This represents the element in the i-th row and j-th column of the i-th element in the optimal transfer matrix.

[0077] Specifically, labeled image samples ( ) and unlabeled image samples ( Each of these is considered as a positive target distribution. and negative target distribution Unlabeled image samples ( Then it is considered as the source distribution. The same importance score is assigned to samples from each source distribution. The transmission loss matrix is ​​defined as follows: ,in , This is the bandwidth parameter. Then distribute the source... Distributed to the target respectively and To perform the transmission, we need to solve the following objective function:

[0078]

[0079] in, For entropy regularization, and To balance the hyperparameters of the two terms, we solve the above objective equation based on the positive and negative target distributions respectively, which yields two transmission strategies. and These reflect the association information between unlabeled image samples and known positive-labeled image samples, and the association information between known negative-labeled image samples, respectively.

[0080] Calculate the entropy distance based on two transmission strategies:

[0081]

[0082] Based on this distance, the following risk score assessment formula can be defined:

[0083]

[0084] Clearly, the risk score of unlabeled image samples can reflect the uncertainty of the sample's labeling information. For images where the labels are difficult to identify, it should have a small impact on the model's training process.

[0085] Furthermore, S6 includes: based on ensemble learning, using a convex combination of machine learning models that have the best performance relative to the baseline model in the base learner pool as an image classification model; wherein the baseline model is trained based on a set of labeled image samples.

[0086] Further, the step of using the convex combination of the machine learning models that have the best performance relative to the baseline model in the base learner pool as the image classification model includes: constructing a third optimization problem to maximize the performance improvement of the image classification model relative to the baseline model; solving the optimization problem to obtain the convex combination of machine learning models, and using it as the image classification model; the third optimization problem is:

[0087]

[0088] in, Represents the baseline model. This represents the i-th machine learning model in the base learner pool. The weight, For the convex combination weights of the base learner, For a convex combination weight set, As the number of base learners, Let the hinge loss function be... It represents the 1-norm.

[0089] Figure 3The execution flow for S6 is demonstrated. Step 1: Collect labeled data from the semi-supervised dataset and train a supervised learning model as a baseline. Step 2: Employ hinge loss and solve for the convex combination weights based on maximizing the performance improvement of the convex combination model of base learners relative to the baseline model. Specifically, it is assumed that the optimal classifier can be represented by a convex combination of multiple base learners, i.e. At this point, the objective function can be written as:

[0090]

[0091] Choosing the objective function to optimize for the worst-case scenario to ensure safety yields the following optimized expression:

[0092]

[0093] In scenarios based on hinge loss, the above equation can be further rewritten as:

[0094]

[0095] because Then there is Therefore, the above equation can ultimately be solved as the following linear programming problem:

[0096]

[0097] Step 3: Generate the final robust semi-supervised image classifier from the base learner pool based on convex combination weights.

[0098] Example 2

[0099] This embodiment provides an image classification method based on semi-supervised learning, including: classifying images using an image classification model constructed using the aforementioned method for constructing a semi-supervised learning-based image classification model. Specifically, inputting the current image into the image classification model outputs the corresponding classification result.

[0100] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing an image classification model based on semi-supervised learning, characterized in that, include: S1: Calculate the similarity information between unlabeled image samples in the unlabeled image sample set and positive and negative samples in the labeled image sample set; S2: Calculate the optimal positive transmission matrix from the unlabeled image sample set to the positive sample using the similarity information with the positive sample; calculate the optimal negative transmission matrix from the unlabeled image sample set to the negative sample using the similarity information with the negative sample; S3: Calculate the risk coefficient of the unlabeled image sample set using the positive optimal transfer matrix and the negative optimal transfer matrix; S4: Train a machine learning model using the risk coefficients of the semi-supervised training subset selected each time and the selected unlabeled image sample set; the semi-supervised training subset is composed of a certain number of samples sampled from the unlabeled image sample set and the labeled image sample set respectively based on a random sampling strategy; S5: Construct a base learner pool using multiple machine learning models obtained through multiple training sessions; S6: Integrate all models in the base learner pool to obtain an image classification model.

2. The method for constructing an image classification model based on semi-supervised learning as described in claim 1, characterized in that, S4 includes: using the semi-supervised training subset selected each time and the corresponding risk coefficient to train a decision boundary based on a semi-supervised optimal interval distribution learning machine, and using it as the machine learning model.

3. The method for constructing an image classification model based on semi-supervised learning as described in claim 2, characterized in that, The objective function of the semi-supervised optimal interval distribution learning machine is: in, Let the first set of model parameters, the second set of model parameters, and the third set of model parameters be the ones to be solved. The sample labels are i and j, where i and j are both sample numbers. To balance the model parameters for the i-th sample corresponding to the hyperparameters, This represents the semi-supervised training subset. The total number of samples in the semi-supervised training subset. For soft-interval hyperparameters, To balance the hyperparameters of interval mean and interval variance, To balance the hyperparameters of the loss, Let be the adjacency matrix of the nearest neighbor graph constructed on the semi-supervised training subset. for Normalized adjacency matrix, Let be the risk coefficient of the i-th unlabeled image sample. This is a baseline model trained based on labeled image samples. For the i-th sample, the superscript T represents the transpose. This is a kernel function mapping.

4. The method for constructing an image classification model based on semi-supervised learning as described in claim 1, characterized in that, S4 includes: training a neural network classifier based on semi-supervised base learning using multiple selected semi-supervised training subsets, and using it as the machine learning model; and using the risk coefficient in the loss function or designing it as a sample weight term to guide the update iteration during the training process of the neural network classifier.

5. The method for constructing an image classification model based on semi-supervised learning as described in claim 1, characterized in that, S2 includes: The forward transfer loss matrix is ​​based on the similarity information between the unlabeled image sample set and the positive class samples. Unlabeled image samples are transmitted to the positive class samples; the optimal forward transmission matrix is ​​obtained by solving the first optimization problem corresponding to the forward transmission. ; The negative transfer loss matrix is ​​formed by using the similarity information between the unlabeled image sample set and the negative class samples. The unlabeled image samples are transmitted to the negative class samples; the second optimization problem corresponding to the negative transmission is solved to obtain the negative optimal transmission matrix. .

6. The method for constructing an image classification model based on semi-supervised learning as described in claim 5, characterized in that, The first optimization problem and the second optimization problem are as follows: in, It is an identity matrix, with the superscript T indicating transpose. Represents the source domain. Indicates the target domain. To balance the hyperparameters of the two losses, This represents the trace operation of a matrix. This represents the entropy regularization term.

7. The method for constructing an image classification model based on semi-supervised learning as described in claim 5, characterized in that, S3 includes: using the formula Calculate the risk coefficient of each unlabeled image sample in the unlabeled image sample set. The distribution of labeled negative class samples, The distribution of labeled positive class samples; For the i-th unlabeled image sample With labeled image samples Entropy distance; in, This indicates that the sample has been labeled. This indicates that the positive class samples have been labeled. Indicates that negative class samples have been labeled. Represents the set of unlabeled samples; when ∈ hour Represents the forward optimal transfer matrix The element in the i-th row and j-th column; when ∈ hour Represents the negative optimal transfer matrix The element in the i-th row and j-th column.

8. The method for constructing an image classification model based on semi-supervised learning as described in claim 1, characterized in that, S6 includes: using a convex combination of the machine learning models that have the best performance relative to the baseline model in the base learner pool as the image classification model based on ensemble learning; wherein the baseline model is trained based on the labeled image sample set.

9. The method for constructing an image classification model based on semi-supervised learning as described in claim 8, characterized in that, The step of using the convex combination of the machine learning models that have the best performance relative to the baseline model in the base learner pool as the image classification model includes: constructing a third optimization problem to maximize the performance improvement of the image classification model relative to the baseline model; solving the optimization problem to obtain the convex combination of the machine learning models, and using it as the image classification model; the third optimization problem is: in, This represents the baseline model. This represents the i-th machine learning model in the base learner pool. The weight, For the convex combination weights of the base learner, For a convex combination weight set, As the number of base learners, Let the hinge loss function be... It represents the 1-norm.

10. An image classification method based on semi-supervised learning, characterized in that, include: Image classification is performed using the image classification model constructed according to any one of claims 1-9.

Images