Novelty Detection Methods Based On Double Discrimination Adversarial Learning

Generative Adversarial Nets(GAN)and Adversarial Autoencoder(AAE)have been successfully applied to image generation.Moreover,adversarial nets can learn the data features contained in the given samples in an unsupervised manner.However,when the traditional adversarial nets are applied to novelty detection,the poor classification performance may be obtained.The reasons lie in two aspects.One is that GAN belongs the generative model,while the novelty detection model is often recognized as the discriminative model.The other is that the existing AAEs utilizing the latent vectors as the discriminative inputs for autoencoder,which makes the reconstruction results of the given data unsatisfying.To make adversarial nets fit for tackling the novelty detection problems,two novelty detection approaches based on double discrimination adversarial learning are proposed in this thesis,which are described in detail as follows.1.A novelty detection method using the global and local discrimination based adversarial autoencoder is proposed.The proposed model consists of one autoencoder and two discrimination models.In the training phase,learning the proposed model consists of minimizing the reconstruction error of autoencoder and two groups of adversarial procedures.The possible model collapse problem for the adversarial learning can thus be effectively avoided.In the testing phase,utilizing the autoencoder trained by double discriminators to reconstruct the testing samples can sufficiently improve the performance of identifying the novel data with the proposed model.Experimental results on MNIST,Fashion-MNIST and CIFAR10 demonstrate that the proposed method can efficiently avoid the problem of possible model collapse.Moreover,in comparison with its related methods,the proposed method obtains better recognition ability.2.D2 GAN based Novelty Detection method is proposed.In the proposed model,dual discriminator GAN is utilized to generate the samples with the same distribution as those few novel data within the given training set.Then,the original training set and the newly generated novel data are combined together to compose a new training set.Furthermore,abinary classifier can be trained by the obtained training set.Due to incorporating double discriminators,the proposed method can also efficiently avoid the possible model collapse caused by the adversarial learning.At last,the comparison between the proposed method and its related approaches are carried on MNIST and Fashion-MNIST.The effectiveness of the proposed method is validated.