There's a New Technique for Facial Recognition That Works in the Dark

Illustration for article titled There's a New Technique for Facial Recognition That Works in the Dark

Researchers have developed a technique for recognizing faces using their thermal signature, opening the door for facial recognition even in the dark. Now computers can tell who you are, even at night.


In a new study by researchers at the Institute of Anthropomatics & Robotics and Karlsruhe Institute of Technology developed a new method for matching a regular visible light photo of a face with an infrared image of that face. This is a vexing problem because there is very little one-to-one correlation between the two types of images, as illustrated in this image from the study:

Illustration for article titled There's a New Technique for Facial Recognition That Works in the Dark

The researchers employed a “deep neural network” to find the correlation between the two types of images. As Gizmodo’s Jamie Condliffe explained recently, deep neural networks are computer programs that are designed to mimic the way that human brains think. If they’re trained with a large enough data set they can make connections based on a complex series of factors. You know, much like a human does.

In this case, there are only few existing datasets of visible light images with corresponding infrared images. The University of Notre Dame set that the researchers used includes batches of images which include subjects shot with different facial expressions, in different lighting conditions, as well as multiple images of the same subject over a period of time.

Using the neural network to find correlations significantly improved this kind of face matching, although, it’s got a ways to go before it’s what you would call definitely reliable. When the system had a lot of visible light images to compare to the thermal image, the network made the correct match 80 percent of the time. With only one visible image, the number of correct matches drops to 55 percent.

So what does this mean? The researchers are upfront about the fact that this stuff is mostly useful for covert surveillance. Using existing visible image databases—mugshots, driver’s license photos, etc—the new technology basically means that government agencies will be able to identify you even when they can’t really see your face. This means at night, as well as in situations where you’ve taken pains to conceal your features. Just another piece of cool science pointing toward applications that make most people squeamish.


[Technology Review]


Keeping in mind that this seems to be a test using stored photos and comparing to thermal imaging, this technology is similar to what the Kinect does to identify a person sitting in front of the camera, even in the dark. Additionally this looks the same as the upcoming Windows 10 Hello authentication system.