GAN-based Abnormal Detection by Recognizing Ungeneratable Patterns

Soto Anno, Yuichi Sasaki

Abstract: One of the approaches of image anomaly detection is to build a model to generate normal images and distinguish ungeneratable images as abnormal. Recently multiple studies reported the effectiveness of applying generative adversarial networks (GANs) in the task. A GAN trained on normal images is unable to generate abnormal image which is not included in the manifold of normal images, hence the generated image shows a substantial difference from the original images. Most of the previous studies measure the difference by pixel-wise residual loss, where the essential difference between normal and abnormal is often smeared out by inevitable pixel-level reconstruction error of Generator. In this report, a new GAN-based semi-supervised anomaly detection model, AnnoGAN, is proposed, in which the pixel-wise residual loss is replaced by a classifier network that is trained to recognize the essential difference between normal and abnormal. Proposed model achieves state-of-the-art performance in image anomaly detection using public datasets.