Using Manifold Embedding for Automatic Threat Detection: An Alternative Machine Learning Approach

This paper presents an approach to automatic threat detection in X-ray imagery based on the combination of manifold embedding and a specialized classification strategy to identify a wide range of firearm threats in imagery captured in operational settings and in challenging simulated settings which closely reflected operational conditions. The results show that this approach is successful in reducing the amount of data needed to produce a flexible classification system by two orders of magnitude if compared to most state-of-the-art deep learning systems used for object recognition. The work also offers useful indications to the amount of data required for such an alternative machine learning approach. It also shows an enhanced algorithmic architecture for the fusion of very different classifiers, so that a generic and flexible system may be adapted to varied operational environments.