Exploring Deep Anomaly Detection Methods Based on Capsule Net

In this paper, we develop and explore deep anomaly detection techniques based on capsule network (named AnoCapsNet) for image data. Being able to encode intrinsic spatial relationship between parts and a whole, CapsNet has been applied as both a classifier and deep autoencoder. This inspires us to design three normality score functions: prediction-probability-based (PP-based), reconstruction-error-based (RE-based), and combination of both (PP+RE-based) for evaluating the “outlierness” of unseen images. Our results on four datasets demonstrate that PP-based and RE-based methods outperform the principled benchmark methods in many cases and the pp-based method performs consistently well, while the RE-based approach is relatively sensitive to the similarity between labeled and unlabeled images. The PP+RE-based approach effectively takes advantages of both methods and achieves state-of-the-art results.