Automatic detection of narcotics

Why is it challenging and how can it be successfully achieved?

The war against illegal drugs is being fought across the globe by many organisations including customs, border control agencies, prison authorities and more recently courier services. Traffickers are exploiting the rapid growth in e-commerce by shipping narcotics via freight forwarding and mail delivery companies, hoping the illicit packages will be hidden in the sheer quantity of items moving around the world.

Automatic narcotics detection using X-ray scanners could certainly be a very effective weapon. However, the security screening industry generally agrees that there are some unique challenges to overcome for conventional X-ray and Computed Tomography (CT) to deliver the same exceptional performance we now expect from explosives detection or object recognition systems.

So why is detecting narcotics so different to detecting explosives or other dangerous items and how can it be successfully addressed?

Two algorithmic approaches

There are two ways to approach automatic detection using algorithms – object recognition and material discrimination. Which technique offers the greatest benefits depends largely on the items to be detected.

Object recognition (sometimes referred to as visual detection) relies on machine learning to support the development of algorithms which imitate the way the human brain processes data to identify patterns used in decision-making. Thousands of X-ray images are fed to the algorithm so it can learn to identify patterns in the shape and texture of objects. Material is also taken into consideration, but only at a high level by focusing on the colour and its intensity as represented in the X-ray image: blue indicating metal and orange organic material, for example.

There are several smart, adaptable, deep learning algorithms in common use – with more under development – for the automatic detection of dangerous, prohibited, and contraband goods. They achieve very high detection and very low false alarm rates for weapons and lithium batteries and provide invaluable support for security operators, customs officers, and other control authorities.

Algorithms can be taught to recognise anything with clearly identifiable characteristics, but object recognition is less useful for detecting substances which are inconsistent in shape or form. When items need to be detected by material discrimination rather than shape, conventional image processing segments and classifies them based on X-ray absorption characteristics – automatic explosives detection is an excellent example of this approach.

A compound problem…

Narcotics detection is more complex because neither algorithmic approach is a perfect fit.

Object recognition can be used to find pills and blister packs but the optical signature is very similar to benign tablets, leading to inevitable false alarms. Without a recognisable shape to ‘learn’ (e.g. powder or liquids), it is more difficult for these algorithms to detect drugs.

Narcotics coming direct from producers and relatively pure in composition can, however, be identified by material discrimination based on physical characteristics, density, or effective atomic number (Zeff) with a high degree of probability.

Drugs such as fentanyl, heroin and cocaine are however often cut with other substances to either stretch the product to increase profits or generate a different effect on the user. As this changes the physical characteristics, it is more difficult with material discrimination to detect smaller quantities which have already been cut and are destined for the street market. In these scenarios an algorithm can be trained to detect individual combinations coming from a specific source which uses the same cutting substance, however if the compound constantly varies, detection becomes more challenging.

…Needs a compound solution

Currently, the most effective option is to deploy an optimal combination of both object recognition and material discrimination depending on individual circumstances. Narcotics are a huge problem around the world and every gram detected is a success – this is a very different use case to explosives detection.

In addition, close cooperation with authorities and recording large numbers of images of substances to train the algorithms will support the development of increasingly effective solutions.

Going forward

Although object recognition and material discrimination can be used successfully together, it is likely that, in the longer term, a different technology will be applied: X-ray diffraction is known to provide highly accurate information about the physical composition of any material and would be best suited for narcotics detection. A system-of-systems concept, deploying X-ray or CT screening with diffraction used for secondary checks, could be a far superior solution when diffraction systems become available.