Mutual Consistency-Ensured Bi-Directional GAN for Anomaly Detection

Abstract

Anomaly detection is a challenging and fundamental issue in computer vision tasks. In recent years, Generative Adversarial Networks (GAN) and Bi-directional GAN based anomaly detection methods have achieved remarkable results. However, the performance of these approaches for anomaly detection depends on the ability to reconstruct a given normal image and to predict its latent variables. Therefore, we design a novel Bi-directional GAN-based anomaly detection model to improve these abilities. Especially, in order to reconstruct the image, we introduce the consistency loss for ensuring mutual mappings in both image and latent space. Moreover, we propose introducing the projection discriminator as an alternative of concatenating discriminator in order to perform efficient conditioning in the Bi-directional GAN model. In experiments, we evaluate the effectiveness of our model in MNIST and MVTec Metal Nut. Our experiment showed that our model allows us to detect various real anomalies such as bent, scratch, color, and flip, and outperforms the conventional ones.