GTOE: Anomaly Detection Geometric Transformation Network with Outlier Exposure for Real Industry Data

Anomaly detection is a technology that is actively applied to many industries. Supervised learning-based anomaly detection requires a large amount of the normal and the abnormal data. However, it is difficult to get the abnormal data compared to the normal data. Therefore, in this paper, we propose a network for anomaly detection using only the normal data based on semi-supervised learning. It is a method that uses only the normal data to learn to fit 72 geometric transformations after using 72 geometric transformations. Furthermore, we improve performance by applying outlier exposure methods, which are methods to learn together by exposing the abnormal data. The abnormal data is not obtained additionally, but uses some from the normal data. The abnormal data are trained using a uniform distribution to avoid affecting the normal data. By using the abnormal data, we generalize anomaly detectors to improve performance on anomaly detection.