AVAE: Adversarial Neural Network for Self-representation in One Class Classification

To address the bottleneck of machine learning algorithms in one class classification(OCC), we present an end-to-end deep learning model for OCC. The basic idea is that our model is divided into two parts, reconstructor and discriminator, which are trained by a minmax game to learn the visual features of target class efficiently. Then the target class is enhanced and reconstructed, so that the discriminator can better distinguish between target class and non-target class. Furthermore, we propose an effective training method for RGB image to further improve the accuracy in OCC. The experiment results on CIFAR-10 show that the accuracy of our method is 6% higher than that of the baseline method which called ALOCC in CVPR201S.