Self-supervised Algorithms for Anomaly Detection on X-Rays

Abstract

In this paper, we consider the problem of prohibited objects detection on X-Ray images obtained by personal inspection scanners. Such scanners are often used on objects that require increased security control. The available data has a number of problems, which are described and addressed in the text. In this paper we consider only self-supervised anomaly detection algorithms. We are using several architectures of autoencoders and comparing them with the state-of-the-art algorithm Patch SVDD, which could be designed and trained on our data from scratch. Unlike supervised learning algorithms, which are often used for such problems, these models do not require a large amount of labeled data for training.