Contrastive learning with pole regularization for deep anomaly detection

Abstract

In this paper we address the self-supervised learning based anomaly detection method. Recently, self-supervised learning outperforms other anomaly detection techniques in image domains. Despite its high performance in anomaly detection, there are not enough research on how to specialize self-supervised learning for anomaly detection or how to utilize learned representations more efficiently for anomaly scoring. In this paper, we regularized representations of contrastive learning to be concentrated near designated pole by utilizing the fact that representations from contrastive learning distributed along hypersphere. And through experiment, we verified improvement in performance than previous methods. For further improvement, we utilized the fact that there are large distributional shift if rotating transformation applied to input. We proposed modified version of pole regularization, which setting a pole for each cluster of rotating transformations. And we identified further improvement. Also, by using the fact that learned representation is concentrated near pole, we introduced a new scheme for anomaly score, which is more efficient and perform well.