Research On Semi-supervised Learning Method Based On Generative Adversarial Networks For Image Classification

In many big data application fields,data annotations require related experts spending plenty of their time and energy,so it is difficult to label all the data in those fileds.Semi-supervised learning,a machine learning method which can combine large amount of unlabeled data with small amount of labeled data,was born to solve this problem.Because it doesn't need much labeled data,semi-supervised learning has received increasing attention in recent years.The traditional semi-supervised learning methods base on the concept of‘pseudo label',which has following shortcoming:1)The pseudo label is the model's prediction of unlabeled data.It is not the real label and therefore has no guarantee of correctness.2)While training through incorrect pseudo labels,the model may learn wrong features,which degrades the model performance.In this paper,a new semi-supervised learning method based on generative adversarial networks is proposed,aiming at the problem that the quality of‘pseudo label'is unstable which leads to the poor training results.The main contributions are as follows:1.The ability to automatically generate new data with spectified labels.This paper proposes a new semi-supervised learning method which adds a generative adversarial network to semi-supervised learning,in order to avoid the 'pseudo label' problems above.In the training process,the generator takes labels as an input to generate more realistic images in specific category,so that the generator can learn not only the overall features of dataset but also the small features from different category.The number of training data can be increased through the data generated from generator.2.Redesign the discriminator of generative adversarial network by adding the classification function to it,so that our model can be used in semi-supervised learning tasks.After changing the discriminator's structure,it can judge both the authenticity and classification of the input data.The new discriminator can learn from the labeled training data,unlabeled training data and the data generated by the generator.Through the coordination game of discriminator and generator,the model can reach the Nash equilibrium and achieve good performance on semi-supervised learning.3.Migrate the model from 2D images to 3D images,further improve and encapsulation the model to make it suitable for more scenes.The algorithm subdivides the loss functions,adds the balance coefficient between them and removes the discriminator's label judgment for the synthetic data from the generator.In experiments,the classification accuracy of our proposed algorithm on public datasets is generally higher than the state-of-the-art semi-supervised learning method Mean Teacher.