Improved Semi Supervised Image Classification Method Based On Triple-Gan

Supervised classification has been widely used in spam classification,gender classification,text classification and image classification.In the field of image classification,there have been a lot of achievements in supervised classification based on a large number of labeled data samples,but in practice,supervised classification tasks need adequate labeled samples.This is sometimes very difficult to be applied in practice.Because most of the data samples that are easy to obtain are unlabeled or lack of cluster tags,how to use a large number of unlabeled data samples to improve the classification ability of the model has become a hot and difficult topic in the field of image classification.Semi-supervised learning algorithm can improve the generalization ability of model by using the implicit supervised information in unlabeled data samples.This paper focuses on the generative method based on the generative countermeasure network.Generative adversarial net has been successfully applied to solve the task of semi-supervised image classification.As a hotspot in the research of semi-supervised image classification,the GAN model constructed by zero-sum game idea of generative confrontation network improves the generating ability and classification ability of its own model through continuous game optimization between the judge and the generator.The output of the network is the probability value of the distribution of the real data.The generator learns from the real data and generates the samples with the distribution of the real data.The network formed by the neural network judges the loss based on the probability value of the judgment of the accuracy of the real data and the generated data and updates the parameters by gradient descent.The network is also constructed by neural network and updated by the gradient descent of the probability value returned by the judge.The judge and the generator are trained alternately and finally reach Nash equilibrium.When the sample size of the labeled image data is very limited,GAN's traditional algorithm can not make full use of the unlabeled sample data to enhance the classifier.Based on the triple-gan semi-supervised classification network model and the well-known clustering hypothesis,a new cluster consistency network model is proposed to improve the classification ability.The former can maintain the cluster consistency of each unlabeled image to provide implicit monitoring information to enhance the classifier,while the latter urges the generator to generate the countermeasure image from the low density distribution region of the real image to improve the recognition ability of the classifier and suppress the collapse of the pattern.Experimental results on MNIST,CIFAR10,SVHN and insulator datasets show that the improved GAN model can achieve high accuracy in semi-supervised image classification tasks.