Search

CN-120766043-B - Construction method of image classification model based on semi-supervised learning

CN120766043BCN 120766043 BCN120766043 BCN 120766043BCN-120766043-B

Abstract

The invention discloses a construction method of an image classification model based on semi-supervised learning, which belongs to the technical field of image classification, wherein the method can evaluate unmarked images in a fine granularity way by judging risk coefficients of unmarked image samples, and further set higher risk coefficients for images which are difficult to distinguish by a feature information comparison fuzzy model so as to reduce the influence of the images in the model training process; the image classification model is guaranteed to be stable in performance based on the integrated learning method and the baseline model. Specifically, as uncertainty exists in the existing model training process of the unmarked image, a plurality of semi-supervised learning subsets are constructed from an original data set by utilizing a sampling technology, and a plurality of image classification models are trained on the basis of the subsets to serve as a base learner, so that a base learner pool is created, a final image classification model is generated from the base learner pool on the basis of integrated learning, and the model performance can be ensured to be not weaker than a baseline model trained from marked image samples.

Inventors

  • ZHANG TENG
  • CAO NAN

Assignees

  • 华中科技大学

Dates

Publication Date
20260505
Application Date
20250718

Claims (10)

  1. 1. The method for constructing the image classification model based on semi-supervised learning is characterized by comprising the following steps of: S1, calculating similarity information of unlabeled image samples in an unlabeled image sample set and positive type samples and negative type samples in a labeled image sample set; S2, calculating a positive optimal transmission matrix for transmitting the unlabeled image sample set to the positive type sample by using similarity information of the unlabeled image sample set and the positive type sample; S3, calculating risk coefficients of the unlabeled image sample set by utilizing the positive optimal transmission matrix and the negative optimal transmission matrix; S4, training a machine learning model by using a semi-supervised training subset selected each time and risk coefficients of an unlabeled image sample set selected in the semi-supervised training subset, wherein the semi-supervised training subset is formed by sampling a certain number of samples from the unlabeled image sample set and the labeled image sample set respectively based on a random sampling strategy; s5, constructing a base learner pool by utilizing a plurality of machine learning models obtained by multiple training; and S6, integrating all the models in the base learner pool to obtain an image classification model.
  2. 2. The method for constructing an image classification model based on semi-supervised learning as set forth in claim 1, wherein S4 comprises training a decision boundary based on semi-supervised optimal interval distribution learning machines with the semi-supervised training subsets and corresponding risk factors for each selection and using the decision boundary as the machine learning model.
  3. 3. The method for constructing an image classification model based on semi-supervised learning as set forth in claim 2, wherein the target expression based on the semi-supervised optimal interval distribution learning machine is: Wherein, the For the first model parameter set, the second model parameter set and the third model parameter set to be solved, For sample marking, i and j are both sample numbers, To balance the model parameters of the ith sample corresponding to the hyper-parameters, Representing the subset of semi-supervised training, For the total number of samples in the semi-supervised training subset, Is a soft-interval super-parameter which is a soft-interval super-parameter, To balance the hyper-parameters of interval mean and interval variance, In order to balance the super-parameters of the loss, For the adjacency matrix of the neighbor graphs constructed on the semi-supervised training subsets, Is that The normalized adjacency matrix is used to determine, For the risk factor of the i-th unlabeled image sample, To train a baseline model based on the marked image samples, For the ith sample, superscript T table transpose, Mapping for kernel functions.
  4. 4. The method for constructing an image classification model based on semi-supervised learning as set forth in claim 1, wherein S4 comprises training a neural network classifier based on semi-supervised learning with a selected subset of semi-supervised training as the machine learning model, and using the risk factors in a loss function or as a sample weight term to guide update iterations in the training process of the neural network classifier.
  5. 5. The method for constructing an image classification model based on semi-supervised learning as set forth in claim 1, wherein the S2 includes: taking similarity information between the unlabeled image sample set and the positive sample as a positive transmission loss matrix Solving a first optimization problem corresponding to forward transmission to obtain a forward optimal transmission matrix ; Taking similarity information between the unlabeled image sample set and the negative sample as a negative transmission loss matrix Solving a second optimization problem corresponding to negative transmission to obtain a negative optimal transmission matrix 。
  6. 6. The method for constructing an image classification model based on semi-supervised learning as set forth in claim 5, wherein the first optimization problem and the second optimization problem are respectively: Wherein, the For the unitary matrix, the superscript T denotes the transpose, The source domain is represented by a representation of the source domain, The domain of the object is represented by a representation, In order to balance the super-parameters of the two losses, The trace-out operation of the matrix is represented, Representing entropy regularization terms.
  7. 7. The method for constructing an image classification model based on semi-supervised learning as set forth in claim 5, wherein S3 comprises utilizing a formula Calculating a risk factor for each unlabeled image sample in the set of unlabeled image samples, For a marked negative-type sample distribution, A distribution of marked positive class samples; for the ith unlabeled image sample And marked image samples Entropy distance of (2); Wherein, the The marked sample is represented by a pattern of marks, Representing a marked positive class of samples, Representing marked negative class samples Representing unlabeled exemplar sets when ∈ Time of day Representing forward optimal transmission matrix When the element of the ith row and the jth column ∈ Time of day Representing a negative-going optimal transmission matrix The elements of row i and column j.
  8. 8. The method for constructing an image classification model based on semi-supervised learning of claim 1, wherein S6 comprises employing a convex combination of the machine learning models in the base learner pool that are optimal with respect to baseline model performance as the image classification model based on ensemble learning, wherein the baseline model is trained based on the set of labeled image samples.
  9. 9. The method for constructing an image classification model based on semi-supervised learning as recited in claim 8, wherein the constructing the convex combination of the machine learning models in the base learner pool that is optimal with respect to the baseline model performance as the image classification model includes constructing a third optimization problem that maximizes the image classification model performance improvement with respect to the baseline model, solving the optimization problem to obtain the convex combination of the machine learning models as the image classification model, wherein the third optimization problem is: Wherein, the The baseline model is represented as such, Representing an ith machine learning model in the base learner pool Is used for the weight of the (c), The combination weights are convex for the basis learner, For the set of convex combining weights, For the number of base learners, As a function of the loss of the hinge, Representing a 1-norm.
  10. 10. An image classification method based on semi-supervised learning is characterized by comprising the step of carrying out image classification by using the image classification model constructed according to any one of claims 1-9.

Description

Construction method of image classification model based on semi-supervised learning Technical Field The invention belongs to the technical field of image classification, and particularly relates to a method for constructing an image classification model based on semi-supervised learning. Background With the rapid development of artificial intelligence related research, the data requirements required for training a model are gradually increased, and in the traditional machine learning field, model training is generally classified into supervised learning, semi-supervised learning and unsupervised learning. The semi-supervised learning is a method for realizing model training under a small amount of labeling information, and is one of the important research directions in the artificial intelligence field. In semi-supervised learning studies, it is generally believed that indiscriminate use of unlabeled samples may certainly allow for improved generalization performance of the model. However, some existing studies have demonstrated that blind use of unlabeled samples may lead to misleading of the model by part of the samples, on the other hand, training a single model may create problems of uncertainty due to the large number of unlabeled samples in semi-supervised learning, i.e. multiple sub-optimal models may be generated during training, without good model selection methods, may lead to dramatic degradation of model performance, resulting in model performance even weaker than the baseline model, i.e. models trained using only labeled samples in semi-supervised data. In many special scenarios, for example, the number of label images available in reality such as medical image recognition is small, when a traditional semi-supervised model training image classifier is used, all unlabeled images are used for model training, and part of unlabeled images are blurred due to shooting, so that the semi-supervised model mistakes the unlabeled images, the model training process is obviously deviated from a correct track, and the final model performance is reduced, even weaker than a baseline model trained by only using a small number of labeled images. Therefore, how to provide a robust semi-supervised learning model training method, so that fine-grained evaluation and use can be performed on unlabeled images, and the influence of the unlabeled images on model training is reduced, which is a problem to be solved currently. Disclosure of Invention Aiming at the defects or improvement demands of the prior art, the invention provides a construction method of an image classification model based on semi-supervised learning, which aims to solve the technical problem that the model is not robust enough due to the fact that the existing semi-supervised learning model training method cannot evaluate and use unmarked images in a fine granularity. In order to achieve the above object, according to one aspect of the present invention, there is provided a method for constructing an image classification model based on semi-supervised learning, comprising: S1, calculating similarity information of unlabeled image samples in an unlabeled image sample set and positive type samples and negative type samples in a labeled image sample set; S2, calculating a positive optimal transmission matrix for transmitting the unlabeled image sample set to the positive type sample by using similarity information of the unlabeled image sample set and the positive type sample; S3, calculating risk coefficients of the unlabeled image sample set by utilizing the positive optimal transmission matrix and the negative optimal transmission matrix; S4, training a machine learning model by using a semi-supervised training subset selected each time and risk coefficients of an unlabeled image sample set selected in the semi-supervised training subset, wherein the semi-supervised training subset is formed by sampling a certain number of samples from the unlabeled image sample set and the labeled image sample set respectively based on a random sampling strategy; s5, constructing a base learner pool by utilizing a plurality of machine learning models obtained by multiple training; and S6, integrating all the models in the base learner pool to obtain an image classification model. Further, the step S4 comprises training a decision boundary based on the semi-supervised optimal interval distribution learning machine by utilizing the semi-supervised training subsets and the corresponding risk factors selected each time. Further, the objective formula based on the semi-supervised optimal interval distribution learning machine is as follows: Wherein, the As a model parameter to be solved for,For sample marking, i and j are both sample numbers,To balance the model parameters of the ith sample corresponding to the hyper-parameters,Representing the subset of semi-supervised training,For the total number of samples in the semi-supervised training subset,Is a soft-interval su