We have completed maintenance on DiscoverMagazine.com and action may be required on your account. Learn More

Algorithm Is First to Identify Faces Better Than Humans Can

By Carl Engelking
Apr 23, 2014 9:37 PMNov 20, 2019 3:14 AM


Sign up for our email newsletter for the latest science news

Let’s face it, from trivia to teaching, computers are doing things we thought were uniquely human — and doing them better. Now, facial recognition, a skill humans once dominated, has a new champion. Computer scientists have developed a facial recognition algorithm that, for the first time, outperforms humans' own abilities. Facial recognition systems are already better than humans at comparing two images to determine if they are the same person, but only when variables like lighting, pose and expression are controlled. When you shake things up, humans still correctly identify faces with over 97 percent accuracy; before now no other algorithm had come close that mark. However, scientists’ new algorithm, called GaussianFace, performed the feat with an accuracy of 98.52 percent. 

Training the Champ

The new algorithm condenses each face it “sees” into a small image to position the eyes, nose, and corners of the mouth in consistent locations. Then, it further divides the image into small, overlapping squares and mathematically charts each square’s unique characteristics, allowing it to compare two images. Scientists trained the algorithm to look for correct facial features by setting it loose on four massive datasets of face images. Once training was complete, it was time for the main event: Labeled Faces in the Wild (LFW).

LFW is a dataset of 13,000 face images, and is considered the gauntlet for face recognition algorithms. It challenges programs to determine whether a pair of pictures is the same person or different people. However, each photo switches up angles, lighting, makeup and expressions to really test an algorithm’s proficiency. The GaussianFace algorithm emerged from the test a champion, beating humans' 97.35 percent average performance on LFW. The algorithm could someday be used in myriad applications including security, image retrieval, and biometric credentials for our computers and mobile devices. The scientists published their findings earlier this month in the online journal arXiv.

Want to see how you'd do against the computer? Check out Medium's collection of facial images

and see if you can tell which are matches.

Photo credit: Dean Drobot/Shutterstock

1 free article left
Want More? Get unlimited access for as low as $1.99/month

Already a subscriber?

Register or Log In

1 free articleSubscribe
Discover Magazine Logo
Want more?

Keep reading for as low as $1.99!


Already a subscriber?

Register or Log In

More From Discover
Recommendations From Our Store
Shop Now
Stay Curious
Our List

Sign up for our weekly science updates.

To The Magazine

Save up to 40% off the cover price when you subscribe to Discover magazine.

Copyright © 2024 Kalmbach Media Co.