Humans still much better at recognising faces than the best computer algorithms, study finds

People skilled in the forensic art of instantly recognising the faces of strangers can still easily outperform the best computer algorithms, a study has found.

Tests have demonstrated people rather than machines will continue to be needed for the foreseeable future to recognise faces in areas such as border control, crime and security, where speed is critical. A panel of 27 professional facial examiners, who had been specially trained to identify people from their faces as part of their job, proved far superior to either computers or untrained people at identifying the photographs of strangers, the study found.

“Computers at the moment can only outperform people in tasks when the images are of high quality, but often they are not. I don’t think there will ever be a time when humans are not needed in this process,” said David White, a psychologist at the University of New South Wales in Sydney, Australia, who led the study.

open image in gallery

The scientists also found the professionally trained examiners use fundamentally different techniques for recognising faces than is commonly used by people to identify the facial features of family members and friends.

While people normally use “holistic” methods that involve comparisons between faces, professional trained examiners use more analytical techniques that exploit differences between certain facial features, Dr White said. “Our results suggest that examiners are using a different strategy to identify faces in their casework than people use in their daily lives,” he said.

The study, published in the Proceedings of the Royal Society B, pitted the 27 professional forensic examiners against untrained people and some of the best computer algorithms for facial recognition. They were each presented with a pair of photographs and asked to make a judgement as to whether they were the same person – even when the photographs were upside down.

“The examiners’ superiority was greatest when they had a longer time to study the images, and they were also more accurate than others at matching faces when the faces were shown upside down. This is consistent with them tuning into the finer details in an image,” Dr White said.

Computer scientists have been working on facial recognition for years and their best algorithms only work well when the images are taken under good lighting conditions. Yet most ID situations involve poor images from CCTV or mobile phones.