Method for constructing image classification model based on semi-supervised learning

Abstract

The invention discloses a method for constructing an image classification model based on semi-supervised learning, and belongs to the technical field of image classification. According to the method, the risk coefficient of an unmarked image sample is judged, fine-grained evaluation can be carried out on an unmarked image, and the image classification accuracy is improved. A relatively high risk coefficient is set for the image of which the feature information is difficult to distinguish compared with the fuzzy model, so that the influence of the image in the model training process is reduced; and the performance robustness of the image classification model is ensured based on the integrated learning method and the baseline model. Specifically, due to the fact that uncertainty exists in the model training process of unmarked images, a plurality of semi-supervised learning subsets are constructed from an original data set through the sampling technology, a plurality of image classification models are trained based on the subsets to serve as base learners, and therefore a base learner pool is created; the final image classification model is generated from the base learner pool based on ensemble learning, and it can be guaranteed that the model performance is not weaker than that of a baseline model trained from marked image samples.

Description

Technical Field

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

[0002] With the rapid development of AI research, the data requirements for model training are also increasing. In traditional machine learning, model training is generally divided into supervised learning, semi-supervised learning, and unsupervised learning. Semi-supervised learning, a method used to train models with minimal labeled information, is a key research direction in the field of AI.

[0003] In semi-supervised learning research, it is generally believed that the indiscriminate use of unlabeled samples will inevitably improve the model's generalization performance. However, some existing studies have confirmed that blindly using unlabeled samples may cause the model to be 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 lead to uncertainty. In other words, multiple suboptimal models may be generated during the training process. Without a good model selection method, this may lead to a sharp decline in model performance, even making the model performance weaker than the baseline model, that is, the model trained only using labeled samples from the semi-supervised data.

[0004] In many special scenarios, such as medical image recognition, the number of labeled images available in reality is relatively small. When using traditional semi-supervised models to train image classifiers, all unlabeled images are used for model training. Some of these unlabeled images are blurred, causing the semi-supervised model to misclassify them. This will obviously cause the model training process to deviate from the correct track, resulting in reduced performance of the final model, which may even be 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 can perform fine-grained evaluation and use of unlabeled images and reduce the impact of unlabeled images on model training is an urgent problem that needs to be solved. Summary of the Invention

[0006] In response to 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, which aims to solve the technical problem that the existing semi-supervised learning model training method is unable to perform fine-grained evaluation and use of unlabeled images, resulting in the model being not robust enough.

[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: S1: Calculate the similarity information between the unlabeled image samples in the unlabeled image sample set and the positive samples and negative samples in the labeled image sample set; S2: Calculating a forward optimal transmission matrix for transmitting the unlabeled image sample set to the positive class samples by using the similarity information with the positive class samples; Calculating a negative optimal transmission matrix for transmitting the unlabeled image sample set to the negative class samples by using the similarity information with the negative class samples; S3: Calculating the risk coefficient of the unlabeled image sample set using the positive optimal transfer matrix and the negative optimal transfer matrix; S4: training a machine learning model using the risk coefficient of the semi-supervised training subset selected each time and the unlabeled image sample set selected therein; wherein 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 based on a random sampling strategy; S5: Constructing a base learner pool using the multiple machine learning models obtained through multiple trainings; S6: Integrate all models in the base learner pool to obtain an image classification model.

[0008] Furthermore, the 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 margin distribution learning machine.

[0009] Furthermore, the objective formula based on the semi-supervised optimal margin distribution learning machine is: in, are the model parameters to be solved, is the sample label, i and j are the sample serial numbers, is the model parameter of the ith sample corresponding to the balanced hyperparameter, represents the semi-supervised training subset, is the total number of samples in the semi-supervised training subset, is the soft margin hyperparameter, is a hyperparameter that balances the interval mean and interval variance, is the hyperparameter for balancing loss, is the adjacency matrix of the neighbor graph constructed on the semi-supervised training subset, for Normalized adjacency matrix, is the risk coefficient of the i-th unlabeled image sample, is a baseline model trained based on labeled image samples, is the i-th sample, the superscript T indicates transposition, is the kernel function mapping.

[0010] Furthermore, the S4 includes: using multiple selected semi-supervised training subsets to train a neural network classifier based on semi-supervised learning respectively; 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.

[0011] Furthermore, the S1 includes: using the similarity information between the unlabeled image sample set and the positive class sample as the forward transmission loss matrix , transmit the unlabeled image samples to the positive samples; solve the first optimization problem corresponding to the forward transmission to obtain the forward optimal transmission matrix ; The similarity information between the unlabeled image sample set and the negative class sample is used as the negative transmission loss matrix , transfer the unlabeled image samples to the negative class samples; solve the second optimization problem corresponding to the negative transmission to obtain the negative optimal transmission matrix .

[0012] Furthermore, the first optimization problem and the second optimization problem are: in, is the identity matrix, the superscript T indicates transpose, represents the source domain, represents the target domain, To balance the hyperparameters of the two losses, represents the matrix trace operation, represents the entropy regularization term.

[0013] Furthermore, the S3 includes: using the formula Calculate the risk coefficient of each unlabeled image sample in the unlabeled image sample set, is the distribution of labeled negative samples, is the distribution of labeled positive samples; is the i-th unlabeled image sample With labeled image samples Entropy distance; in, represents the labeled samples, where represents the labeled positive samples, Indicates labeled negative samples represents the set of unlabeled samples, Represents the element in the i-th row and j-th column of the i-th matrix in the optimal transmission matrix.

[0014] Furthermore, S6 includes: based on ensemble learning, using the convex combination of the machine learning models with 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.

[0015] Furthermore, the method of using the convex combination of the machine learning models with the best performance relative to the baseline model in the base learner pool as the image classification model includes: constructing a third optimization problem for maximizing 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, represents the baseline model, Represents the i-th machine learning model in the base learner pool The weight of is the convex combination weight of the base learner, is a convex combination weight set, is the number of base learners, is the hinge loss function, represents the 1-norm.

[0016] According to another aspect of the present invention, a method for image classification based on semi-supervised learning is provided, comprising: performing image classification using the image classification model.

[0017] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art: (1) The present invention provides a method for constructing an image classification model based on semi-supervised learning. By judging the risk coefficient of unlabeled image samples, the unlabeled images can be evaluated in a fine-grained manner to reduce their influence during the model training process and improve the 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 from each training is not necessarily the optimal model. Sampling technology is used to construct multiple semi-supervised learning subsets from the original data set, and multiple image classification models are trained based on the subsets as base learners, thereby creating a base learner pool. The final image classification model is generated from the base learner pool based on ensemble learning, which can ensure that the model performance is not inferior to the baseline model trained from labeled image samples.

[0018] (2) In semi-supervised learning, unlabeled images are not always beneficial to the model training process. Since semi-supervised learning methods usually require label assignment for unlabeled images, the model is very likely to give incorrect label information for some images with relatively vague information and difficult to recognize, resulting in the model training process being misled. This scheme conducts a preliminary analysis of the unlabeled samples and calculates the similarity between the unlabeled samples and the positive labeled samples and negative labeled samples respectively, in preparation for the subsequent risk coefficient judgment. Then, by utilizing the optimal transmission strategy, the unlabeled images are transmitted to the known positive labeled and negative labeled images respectively, and the corresponding transmission matrix is ​​obtained. By analyzing the information contained in the matrix, the risk coefficient of the unlabeled image is judged and assigned a corresponding risk coefficient, thereby enabling the unlabeled images to be used for training in a differentiated manner during the model training process, which can reduce the uncertainty of the unlabeled images.

[0019] (3) The method of training multiple high-quality low-density decision boundaries based on a semi-supervised optimal margin distribution learning machine to construct a base learner pool, or training multiple high-quality neural network classifiers based on margin distribution loss to construct a base learner pool includes: based on traditional semi-supervised learning model training methods, such as a semi-supervised model based on SVM, a semi-supervised model based on ODM, or a semi-supervised model based on a deep neural network, training a semi-supervised image classifier as a base learner, and its training data comes from a subset constructed from the original data set through a random sampling strategy. Compared with the traditional method of training only a single semi-supervised model, this method can better reduce the uncertainty effect of unlabeled image samples on the model during the training process. Each semi-supervised classifier has good performance on its own data distribution, and the final integrated model can integrate the advantages of each model to improve the overall generalization performance.

[0020] (4) This scheme convexly combines the models in the base learner pool so that the final model can achieve good classification performance over the entire distribution of semi-supervised data. The model integration weight is determined by maximizing the performance improvement of the model relative to the baseline model trained using labeled images, ensuring that the performance of the final semi-supervised image classifier is not inferior to the baseline model.

[0021] (5) This scheme collects all semi-supervised base learner models, and then designs the performance difference objective between the baseline model trained on labeled image samples, the model of the convex combination of semi-supervised base learners, and the potential optimal model based on the hinge loss. Specifically, assuming that the optimal classifier is Indicates that for this optimal classifier, it is better than the benchmark classifier , that is, the classifier trained using supervised information has the greatest benefit improvement. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1is a flowchart of a method for constructing an image classification model based on semi-supervised learning provided in Embodiment 1 of the present application; Figure 2 is a schematic diagram of constructing a base learner pool based on a semi-supervised learning model provided in Embodiment 1 of the present application; Figure 3 is a schematic diagram of constructing a robust image classification model based on a semi-supervised base learner pool provided in Embodiment 1 of the present application. DETAILED DESCRIPTION

[0023] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application 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 only used to explain the present application and should not be used to limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0024] The present application explains some nouns.

[0025] Semi-supervised learning: refers to the performance of the semi-supervised learning model, which will not be weaker than the performance of the model trained only by the labeled data in the semi-supervised data. In the traditional semi-supervised learning model, unlabeled data is usually used for model training without distinction, but unlabeled data does not always improve the performance of the model. On the contrary, some difficult-to-identify unlabeled data may mislead the training process of the model, resulting in a sharp decline in the performance of the final model, and even lower than the model trained only by the labeled data.

[0026] Optimal transport: a mathematical tool for calculating the difference between two distributions. The advantage of optimal transport is that it can handle discrete, continuous or both measures in the same framework. The measure refers to the quantitative concept used to describe the size or size relationship in space, which can be intuitively represented as mass, area and volume, and abstractly represented as discrete histogram or probability density function in function space. The optimal transport between probability measures can be understood as establishing the lowest total transport scheme between masses in different distribution spaces. The cost here usually refers to the distance between the subsets, thereby establishing the mapping or transfer between probability measures.

[0027] Ensemble learning: a machine learning method based on multiple base learners, which combines them in a certain way to obtain a model 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, and can significantly improve the prediction performance and robustness of the model.

[0028] Embodiment 1 As Figure 1As shown, this embodiment provides a method for constructing an image classification model based on semi-supervised learning, including: S1-S6. S1: obtaining similarity information of a labeled image sample set and an unlabeled image sample set; S2: using the similarity information to calculate the optimal transmission matrix based on the optimal transmission loss theory; S3: using the optimal transmission matrix to calculate the risk coefficient of the unlabeled image sample set; S4: based on a random sampling strategy, a certain number of samples are sampled from the unlabeled image sample set and the labeled image sample set to form a semi-supervised training subset; using the semi-supervised training subset selected each time and its corresponding risk coefficient to train a machine learning model; S5: using multiple machine learning models obtained from multiple trainings to construct a base learner pool; S6: integrating all models in the base learner pool to obtain an image classification model.

[0029] in, Figure 2 The process for building a base learner pool is presented. Step 1: Divide the semi-supervised learning data into an unlabeled dataset, a positively labeled dataset, and a negatively labeled dataset. 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, multiple base learners are trained from the multiple subsets to construct the base learner pool. In addition, during the training process, the risk factor of the unlabeled image samples is introduced to modify the model update process.

[0030] Furthermore, 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 margin distribution learning machine.

[0031] Furthermore, the objective formula based on the semi-supervised optimal margin distribution learning machine is: in, are the model parameters to be solved, is the sample label, i and j are the sample serial numbers, is the model parameter of the ith sample corresponding to the balanced hyperparameter, represents the semi-supervised training subset, is the total number of samples in the semi-supervised training subset, is the soft margin hyperparameter, is a hyperparameter that balances the interval mean and interval variance, is the hyperparameter for balancing loss, is the adjacency matrix of the neighbor graph constructed on the semi-supervised training subset, for Normalized adjacency matrix, is the risk coefficient of the i-th unlabeled image sample, is a baseline model trained based on labeled image samples, is the i-th sample, the superscript T indicates transposition, is the kernel function mapping.

[0032] Specifically, taking the semi-supervised optimal margin distribution learning machine as the base learner as an example, in order to extract the association between the sample label space and the feature space, a Neighborhood Graph ,in Representation dataset The sample points in Represents the edges between sample points, and the weight of the edge represents the similarity between samples. Setting up the adjacency matrix , where if the sample and samples If there is an edge, then ,otherwise In addition, let represents the normalized adjacency matrix, where , It can be used to reflect the correlation information between samples. In order to further improve the security of the base learner and reduce the impact of high-risk samples on the model, the following regularization terms are proposed: in Represents the optimal margin distribution learning machine for supervised learning, which is composed of a set of labeled image samples Combining the regularization term and the adjacency matrix term with the optimal margin distribution learning machine, we can obtain the following objective formula: in, , and To balance the hyperparameters, it is used to adjust the empirical loss of labeled image samples and unlabeled image samples. is the risk assessment value of the unlabeled image sample obtained through optimal transmission.

[0033] Furthermore, S4 includes: using the multiple selected semi-supervised training subsets to train a neural network classifier based on semi-supervised learning; 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. During the training process, the risk coefficient of the unlabeled image samples is introduced into the training to correct the model update process.

[0034] Furthermore, S1 includes: dividing the labeled image samples into positive samples and negative samples; respectively calculating the similarity information of the unlabeled image samples in the unlabeled image sample set with the positive samples and the similarity information of the unlabeled image samples with the negative samples. S2 includes: using the similarity information with the positive samples as the forward transmission loss matrix , transfer the unlabeled image samples to the positive samples; solve the first optimization problem corresponding to the forward transmission to obtain the forward optimal transmission matrix ; The similarity information with the negative class sample is used as the negative transmission loss matrix , transfer the unlabeled image samples to the negative class samples; solve the second optimization problem corresponding to the negative transmission to obtain the negative optimal transmission matrix .

[0035] Furthermore, the first optimization problem and the second optimization problem are: in, is the identity matrix, the superscript T indicates transpose, represents the source domain, represents the target domain, To balance the hyperparameters of the two losses, represents the matrix trace operation, represents the entropy regularization term. Further, using the formula Calculate the risk coefficient of each unlabeled image sample in the unlabeled image sample set, is the distribution of labeled negative samples, is the distribution of labeled positive samples; is the i-th unlabeled image sample With labeled image samples Entropy distance: in, represents the labeled samples, where represents the labeled positive samples, represents the labeled negative samples, represents a set of unlabeled image samples, Represents the element in the i-th row and j-th column of the i-th matrix in the optimal transmission matrix.

[0036] Specifically, the labeled image samples ( ) and unlabeled image samples ( ) are regarded as positive target distributions and negative target distribution , while the unlabeled image samples ( ) is considered as source distribution Assign the same importance score to each sample in the source distribution. The transmission loss matrix is ​​defined as ,in , is the bandwidth parameter. Then the source distribution Distribute to the target and To perform the transmission, we need to solve the following objective formula: in, is the entropy regularization term, and To balance the two hyperparameters. Solving the above target equation based on the positive target distribution and the negative target distribution, we can get two transmission strategies. and , respectively reflecting the association information between unlabeled image samples and known positive labeled image samples, and the association information between known negative labeled image samples.

[0037] Based on two transmission strategies, the entropy distance is calculated: Based on this distance, the following risk score assessment formula can be defined: Obviously, the risk score of unlabeled image samples can reflect the uncertainty of the labeling information of the samples. For images whose labels are difficult to identify, it should have a smaller impact on the model training process.

[0038] Furthermore, S6 includes: based on ensemble learning, using a convex combination of the machine learning models with 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 labeled image sample set.

[0039] Further, the convex combination of the machine learning model in the base learner pool that is optimal relative to the baseline model performance as the image classification model includes: constructing a third optimization problem that maximizes the performance improvement of the image classification model relative to the baseline model; solving the optimization problem to obtain a convex combination of machine learning models as an image classification model; and the third optimization problem is: wherein, represents the baseline model, represents the weight of the i-th machine learning model in the base learner pool, is the weight of the base learner convex combination, is the set of convex combination weights, is the number of base learners, is the hinge loss function, represents the 1-norm.

[0040] The execution flow for S6 is shown. Step 1: Collect the labeled data of the semi-supervised dataset, and train a supervised learning model as a baseline. Step 2: Use the hinge loss, and solve the convex combination weights based on maximizing the performance improvement of the base learner convex combination model relative to the baseline model. Specifically, assuming that the optimal classifier can be represented by the convex combination of multiple base learners, i.e. Figure 3 At this time, the target formula can be written as: The target formula in the worst case is selected for optimization to ensure safety, i.e. the following optimization formula can be obtained: In the scenario based on the hinge loss, the above formula can be further rewritten as: Since , there is Therefore, the above formula can finally be written as the following linear programming problem for solving: Step 3: Based on the convex combination weights, generate the final robust semi-supervised image classifier from the base learner pool. Embodiment 2

[0041] The embodiment provides an image classification method based on semi-supervised learning, including: using the image classification model constructed by the construction method of the image classification model based on semi-supervised learning to perform image classification. Specifically, inputting the current image into the image classification model can output the corresponding classification result.

[0042] ​It will be easily understood by those skilled in the art 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 in 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 the unlabeled image samples in the unlabeled image sample set and the positive samples and negative samples in the labeled image sample set; S2: Calculating a forward optimal transmission matrix for transmitting the unlabeled image sample set to the positive class samples by using the similarity information with the positive class samples; Calculating a negative optimal transmission matrix for transmitting the unlabeled image sample set to the negative class samples by using the similarity information with the negative class samples; S3: Calculating the risk coefficient of the unlabeled image sample set using the positive optimal transfer matrix and the negative optimal transfer matrix; S4: training a machine learning model using the risk coefficient of the semi-supervised training subset selected each time and the unlabeled image sample set selected therein; wherein 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 based on a random sampling strategy; S5: Constructing a base learner pool using the multiple machine learning models obtained through multiple trainings; 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 according to claim 1, wherein: The 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 use it as the machine learning model.

3. The method for constructing an image classification model based on semi-supervised learning according to claim 2, wherein: The objective formula based on the semi-supervised optimal margin distribution learning machine is: in, are the first model parameter, the second model parameter set and the third model parameter set to be solved, is the sample label, i and j are the sample serial numbers, is the model parameter of the ith sample corresponding to the balanced hyperparameter, represents the semi-supervised training subset, is the total number of samples in the semi-supervised training subset, is the soft margin hyperparameter, is a hyperparameter that balances the interval mean and interval variance, is the hyperparameter for balancing loss, is the adjacency matrix of the neighbor graph constructed on the semi-supervised training subset, for Normalized adjacency matrix, is the risk coefficient of the i-th unlabeled image sample, is a baseline model trained based on labeled image samples, is the i-th sample, the superscript T indicates transposition, is the kernel function mapping.

4. The method for constructing an image classification model based on semi-supervised learning according to claim 1, wherein: The S4 includes: using multiple selected semi-supervised training subsets to train a neural network classifier based on semi-supervised basis learning, and using it as the machine learning model; 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 according to claim 1, wherein: The S2 includes: The similarity information between the unlabeled image sample set and the positive class sample is used as the forward transmission loss matrix , transmit the unlabeled image samples to the positive samples; solve the first optimization problem corresponding to the forward transmission to obtain the forward optimal transmission matrix ; The similarity information between the unlabeled image sample set and the negative class samples is used as the negative transmission loss matrix , transfer the unlabeled image samples to the negative class samples; solve the second optimization problem corresponding to the negative transmission to obtain the negative optimal transmission matrix .

6. The method for constructing an image classification model based on semi-supervised learning according to claim 5, wherein: The first optimization problem and the second optimization problem are respectively: in, is the identity matrix, the superscript T indicates transpose, represents the source domain, represents the target domain, To balance the hyperparameters of the two losses, represents the matrix trace operation, represents the entropy regularization term.

7. The method for constructing an image classification model based on semi-supervised learning according to claim 5, wherein: The S3 includes: using the formula Calculate the risk coefficient of each unlabeled image sample in the unlabeled image sample set, is the distribution of labeled negative samples, is the distribution of labeled positive samples; is the i-th unlabeled image sample With labeled image samples Entropy distance; in, represents the labeled samples, where represents the labeled positive samples, Indicates labeled negative samples represents the unlabeled sample set; when ∈ hour represents the positive optimal transmission matrix The element in row i and column j of ; when ∈ hour represents the positive optimal transmission matrix The element at row i and column j in .

8. The method for constructing an image classification model based on semi-supervised learning according to claim 1, wherein: The S6 includes: based on ensemble learning, taking the convex combination of the machine learning models with 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.

9. The method for constructing an image classification model based on semi-supervised learning according to claim 8, wherein: The method of using the convex combination of the machine learning models with the best performance relative to the baseline model in the base learner pool as the image classification model includes: constructing a third optimization problem for maximizing 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, represents the baseline model, Represents the i-th machine learning model in the base learner pool The weight of is the convex combination weight of the base learner, is a convex combination weight set, is the number of base learners, is the hinge loss function, 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 to 9.

Images