Studies On Improved Semi-supervised Generative Adversarial Networks For Image Classification

Semi-supervised learning is a hot point in deep learning field,which is an image classification method based on a big number of unlabeled samples and less labeled data.Traditional semi-supervised learning methods include collaborative training,semisupervised support vector machine,graph theory semi-supervised learning and so on.With continuous development of deep learning in the field of image recognition,more and more experts and scholars have adopted deep learning method to solve the problem of semisupervised learning,among which the most representative is the semi-supervised generative adversarial network.As a novel deep learning method,the generative adversiarial network has shown good performance in graphics generation,graphics repair and image classification,however,its pattern collapse problem is still exist in the semisupervised classification problem,and the shortage of classification accuracy also exist.In view of these shortcomings,this paper proposes two image classification methods based on improved semi-supervised generative adversarial networks.The specific research contents are as follows:Firstly,because of the semi-supervised generative adversarial networks have a weak distribution matching ability,when the data complexity increases,gererated samples inevitably produced outside the manifold,so a method based on manifold regularization is proposed.Manifold regularization encourages the classifier to keep the local perturbation of generator parameters unchanged,that is,to assign similar labels to the close points in the data manifold,so as to improve the generalization ability of the model.At the same time,for labeled samples,the loss function of scalable SVM was used to replace Softmax to improve the classification accuracy of labeled samples.Secondly,global manifold regularization is not sensitive enough to the perturbation of local data manifold,so a local manifold regularization method is proposed to prevent the model from falling into local collapse by building local generators on the data manifold,so as to solve the problem of overtraining of discriminator.The core of this method is to propose a new loss function with local manifold regularization term,which forces the model to keep invariance to the local disturbance of data manifold,so as to make the model have better robustness and improve the accuracy of image classification.Experiments on Cifar-10,SVHN and cifar-100 data sets show that our proposed models have better classification performance on semi-supervised classification problem compared with the traditional semi-supervised generative adversarial networks.