Welcome to my research homepage! Here you can find out more about my current research and my previous activities. I am also planning to start a blog if I find time so keep an eye out. I currently work within the COMPASS (Computational Security Science) grouping at UCL, whose focus is on developing state-of-the-art computational methods for the exploitation of security imagery.
Below I give a quick motivation for my current interest, namely developing methods for the automated analysis of imagery for border security.
Throughout history humans have tried to optimise trade, be it the domestication of the camel to allow transport of spices and silk, or the introduction of the ISO standards on containerisation to simplify international trade. The benefit of trade is that a geopolitical region can swap locally abundant items for those that are locally scarce. Such items were, and still are, mainly natural resources. However, items may also include the products that can be engineered from those resources, or the experience and skill-set of the engineers that build them. In the last century technology has facilitated international trade by land, sea and air. It is now possible to trade items anywhere on earth within hours, or even instantly when services are traded on the Internet. People are even able to trade Britain, with its scarce solar resources, for places on the other side of the world, like Australia, which has a plentiful supply.
All traded goods and people have to pass through at least two borders; their country of origin and their destination. These border crossings give the local government a chance to check for illicit behaviors. In the country of origin they might check that the goods aren't stolen, that the money isn't being laundered, or the person isn't a wanted felon. Whilst at the destination they might check that goods aren't mislabeled to avoid tariffs, the goods aren't hiding drugs or illegal firearms, and that the person is legally entitled to passage. However, these security checks are often slow and annoying, and turn the dream of having optimal trade into a nightmare, unless compromises are made.
Over the last few decades, technologies have been developed that form an image by non-destructively interrogating baggage, people, or cargo. The image is inspected by a customs official who searches it for signs of threats or abnormalities. Either security or throughput are compromised. At airports all passengers and baggage are screened, thus security is strong but throughput is reduced. For cargo, only a very small fraction of high risk cargo is inspected, thus container throughput is high but security is weakened. This compromise originates from the human operator. Human operators are expensive, slow, error-prone, susceptible to bribery, and have fueled public concerns over privacy violation.
We are currently undergoing a revolution in Machine Learning and Vision. Over the last few years, algorithms have begun to outperform humans at single well-defined tasks such as object identification, and playing Go. And it is now becoming possible to train and run such systems on GPUs inside a desktop computer.
Machine vision algorithms can be used to understand the contents of images. For example, they can decide whether an image contains a threat. The specific cues the algorithm uses are usually machine learnt from a training dataset which contains examples of threats and non-threats under varying conditions. Until recently, the cues were chosen so that they were invariant to, and thus the algorithm robust to, varying conditions (e.g. illumination, orientation) relevant to the application. However, over the last few years, methods have been developed, such as Deep Learning, that learn the cues as well as their optimum combination scheme.
With these advances, it makes sense that, one day, such algorithms will outperform border officials at image search tasks in terms of both accuracy and speed. Such systems, won't be tempted by bribery, won't require holidays, and could potentially remedy concerns over privacy.