Consistency Ensured Bi-directional GAN for Anomaly Detection

EasyChair Preprint 2583

Abstract

Anomaly detection is a challenging and fundamental issue in computer vision tasks. In recent years, GAN (Generative Adversarial Networks) based anomaly detection methods have achieved remarkable results. But the instability of training of GAN could be considered that decreases the anomaly detection score. In par-ticular, Bi-directional GAN has the following two causes that make the training difficult: the lack of consistency of the mutual mapping between the image space and the latent space, and the difficulty in conditioning by the latent variables of the image. Here we propose a novel GAN-based anomaly detection model. In our model, we introduce the consistency loss for ensuring mutual mappings. Fur-ther, we propose introducing the projection discriminator as an alternative of con-catenating discriminator in order to perform efficient conditioning in the Bi-directional GAN model. In experiments, we evaluate the effectiveness of our model in a simple dataset and real-world setting dataset and confirmed that our model outperforms the conventional anomaly detection methods.

Keyphrases: Generative Adversarial Networks, Projection Discriminator, anomaly detection

@booklet{EasyChair:2583,
  author    = {Kyosuke Komoto and Hiroaki Aizawa and Kunihito Kato},
  title     = {Consistency Ensured Bi-directional GAN for Anomaly Detection},
  howpublished = {EasyChair Preprint 2583},
  year      = {EasyChair, 2020}}