Threat Image Projection (TIP) into X-ray images of cargo containers for training humans and machines

Abstract

We propose a framework for Threat Image Projection (TIP) in cargo transmission X-ray imagery. The method exploits the approximately multiplicative nature of X-ray imagery to extract a library of threat items. These items can then be projected into real cargo. We show using experimental data that there is no significant qualitative or quantitative difference between real threat images and TIP images. We also describe methods for adding realistic variation to TIP images in order to robustify Machine Learning (ML) based algorithms trained on TIP. These variations are derived from cargo X-ray image formation, and include: (i) translations; (ii) magnification; (iii) rotations; (iv) noise; (v) illumination; (vi) volume and density; and (vii) obscuration. These methods are particularly relevant for representation learning, since it allows the system to learn features that are invariant to these variations. The framework also allows efficient addition of new or emerging threats to a detection system, which is important if time is critical. We have applied the framework to training ML-based cargo algorithms for (i) detection of loads (empty verification), (ii) detection of concealed cars (ii) detection of Small Metallic Threats (SMTs). TIP also enables algorithm testing under controlled conditions, allowing one to gain a deeper understanding of performance. Whilst we have focused on robustifying ML-based threat detectors, our TIP method can also be used to train and robustify human threat detectors as is done in cabin baggage screening.