Anomaly Detection with VAEGAN and Attention

Abstract

Visual anomaly detection is common in several applications including medical screening and production quality check. There are various methods that are using the reconstructed image from a generative model which is trained only normal patterns. However, the previous method used a generative adversarial network has a problem of dropping a local minimum and robustness of noise. In this paper, we propose an anomaly detection method using the reconstructed image from a conditional generative model which is called VAEGAN. We also propose introducing an attention mechanism by using the trained model. We conducted experiments on multiple datasets contained noise, we verified the proposed method over-performed the previous methods.