Semi-Supervised Image Classification Based on Auxiliary Branching Generative Adversarial Networks

Multi-label Image classification is a challenging task owing to lack of sufficient labeled data for training in many practical applications. In this paper, we design a novel Semisupervised image classification model based on a generative adversarial networks with a auxiliary branch. In our method, an end-to-end training procedure continuously improves the quality of pseudo-label generation, and the increasing accuracy of pseudolabels in turn enhances the classification ability of the discriminator. In addition, to improve the model generalization ability and stability, different levels of data enhancement operations are performed on the unlabeled data. An enhanced feature matching schema is also used to combine the density penalty term with the feature matching method to help the discriminator better learn the category boundary and effectively solve the pattern collapse problem. Experimental results on SVHN and CIFAR-10 image datasets show that compared with other semi-supervised methods, the classification accuracy of the proposed method improves by 0.3 % and 0.7 %, respectively, when using 1000 labeled data.