Security staff inspect x-ray images by: threat detection, where they look for particular items (e.g. knives, detonators); and anomaly detection, where they look for deviations from normal. This project will automate anomaly detection, for which there are no current systems. For firearms, anomaly detection is particularly important for ISO containers and vehicles, where the fabric of the container or vehicle provides opportunities for concealment. Firearms so concealed may not be visible as such, hence not detectable by threat detection methods, but may still be noticed by security staff who spot a darkening out of place (e.g. in the roof of an ISO container) or a shape not quite right (e.g. the engine block of a car). The system we develop for anomaly detection, like experienced security staff, will 'know' what is normal so that it can spot such deviations.

This project will develop firearm detection algorithms using the most recent methods of Computer Vision. Specifically, we will use Convolutional Neural Networks (CNNs), with parameters Deep Learnt from training images. CNNs approximate the action and connections of neurons in the human brain. ‘Deep’ because of the many layers of the network of artificial neurons that they employ. ‘Learning’ because only the broad architecture of the network is engineered, its detailed parameters being learnt by exposing it to relevant images. Across the two phases of the project, we will develop: a single algorithm for detection of firearms in x-ray images; with performance validated as being at least at human-level; applicable to images from scanners of any modality and manufacturer, after a one-off automated tuning process for operation on a new scanner type, requiring only a sample dataset of benign images.

Since 9/11 commercial flights have been attacked by groups with political aims. Their strategic aim is to provoke fear, leading to pressure for political change. Reactive security measures, which visibly guard against a repeat attack, achieve the purpose of the attack since they constantly remind of the existence of the adversaries, who are confirmed as national enemies, not mere criminals. Understanding this, the terrorists readily vary their tactics. Each variation provokes a new measure layered on top of existing measures. The apparent potency of the adversaries is hugely magnified: we walk unshod, and have strangers touch our `junk' and taste our children's food. To exit this cycle a novel approach is needed that has the capability to detect the next attack, not the previous; and to do so invisibly, so that fear is not further magnified. In this project, we will apply the latest Computer Science to automate Anomaly Detection (noticing the suspicious or unusual), as used by experienced, trained security staff. Deep Learning algorithms, that mimic the operation of the human brain, makes this feasible: they recognize faces better than humans. If automated Anomaly Detection at human-levels of performance can be achieved, then computers calculating out of sight, can pore over every luggage and baggage scan and every airport CCTV feed, as if a thousand trained, experienced, unflagging security staff were constantly employed in every airport looking for oddities and passing them up to security staff when found.

In 2013 the Renzoni lab at UCL demonstrated, for the first time, the possibility of imaging using atomic magnetometers in the Magnetic Induction Tomography modality. This opened up a new realm for electromagnetic imaging, given the extreme sensitivity of atomic magnetometers at low frequency. Electromagnetic imaging has potential in many security domains where x-rays are not applicable, such as universal fast parcel screening. COMPASS is working with the Renzoni lab, applying machine learning methods to produce tomographic images from magnetometer measurements that depend non-linearly on the scene. The Machine Learning approach provides an alternative to the standard Inverse Problem approach requiring many Finite Element simulations to produce each image. By shifting the computational burden onto a training stage, we permit fast enough inversion to be applicable in security applications with high throughput.