Facial Recognition Now Works In The Dark

Isopix / REX Shutterstock

Facial recognition tech has been eating its carrots: new developments mean that it can now tell who’s who in the dark.

The new technique, developed by researchers at the Karlsruhe Institute of Technology in Germany, uses the heat from your face to tell who you are. It uses thermal imaging to create an infrared picture of your face, which can then be matched to a photograph taken in the light.

You’ll need a computer to do the matching, though. As you can see, the thermal images don’t look so much like people as demons sent from a hell dimension to bring about armageddon:

Saquib Sarfraz and Rainer Stiefelhagen/Karlsruhe Institute of Technology

This is because the way your face gives off infrared light is very different to the way it reflects it in daylight. The infrared process is very variable, depending on the temperature of your skin. So your facial map will look different if you have a fever, for example, or if you’ve just been for a run.

Another problem the researchers had to deal with was that the resolution of the two images is madly different; photos taken with infrared cameras as very low resolution, while normal photos tend to be much better quality.

So, because of all this, they had to turn to deep neural networks to decode the images. These networks are computer programmes that imitate the way the human brain works, so feed them enough data and they can make complex connections between wildly different things just as we do.

Of course, you can’t guarantee that both the infrared image and the regular photo will feature the same facial expressions or other variables, so researchers used people photographed with different expressions, in different lighting set ups and lots of different photos of the same person taken over a period of time.

And it worked - most of the time. Well, some of the time. When the system had a lot of visible light images to compare to the thermal image, 80% of matches were right. But when it only had one visible image to play with, its hit rate was only 55%.

So there’s still a way to go before this is a perfect identification system - but it’s likely to improve very quickly as more images and data are added to the database. Cat burglars beware.