Revelations In Online Facial Recognition

Ground-breaking biometric research has shown that the freely available facial recognition search engines used by social networking sites such as Facebook and Picasa are as accurate as some specialist biometric systems sold to government agencies, such as police forces.

Michael Thieme, director of special projects for the International Biometric Group in the US, revealed the results of the research at the Biometrics 2012 conference in London in October. The study looked at detection rates and the accuracy of grouping and tagging of images on social networking engines.

To conduct the study, a large quantity of online facial images were retrieved from three major social networking sites that rely heavily on image recognition – Facebook, Face.com and Picasa – and measured how well they performed in terms of detection rates, grouping and tagging. The exercise was basically data processing, but in each case human adjudication was used to check the results. “This wasn’t just a case of us throwing data at Facebook and seeing what comes back,” says Thieme.

“We actually looked at every match that was returned and made sure it was correct. It was extremely labour intensive, but it was the only way for us to accurately measure how well these systems performed,” he adds.

The images used were often far outside the strict controls used, for example, by passport authorities to ensure their biometric engine was able to recognise a face – so the angle of the face and its position relative to the frame of the picture would vary quite widely.

“The results were really very surprising and were not based on super high quality images,” says Thieme. “In fact, in a lot of cases, we were looking at 21-30 pixels. These were small images, not something produced from an iPhone. It made us wonder what could be achieved if you put some really high quality images through these online services.”

The results of the tests showed that the social networking sites were extremely accurate at focusing on faces in pictures where a lot of other objects could have been picked up in error. The tests also measured how often the websites were able to correctly match two images of the same person as opposed to confusing the image with one of a relative for example.

In the case of Picasa, the study showed the accuracy rate was 99 per cent. “It was right in almost every case,” says Thieme. “Now that doesn’t mean it found every face. It did miss some altogether but when it made matches they were right most of the time. For us this was a stunning finding. That the sort of thing you get online with those quality of images was grouping those images with that degree of reliability.”
He says to his knowledge no one has measured this before and it goes against the traditional view that the free or cheap search engines available to everybody to use online are in a different league to expensive and sophisticated biometric search tools developed by specialist technology companies. “The concept that the only way to do reliable face matching is by using an enterprise technology with enterprise licences is not correct.”

So what does this research mean for the law enforcement and homeland security users of biometric facial recognition face engines? Thieme thinks that for agencies who only need to conduct searches using a relatively small watch list – under 10,000 images – these free, online services would be accurate enough for their requirements.

In some cases, Thieme even suggests that using online biometric search engines would be more reliable than commercial ones. This is because, he says, a lot of specialist engines are calibrated to work with strict controls in force in the image – such as a passport or identity card photo. “They will be able to find a higher proportion of images with faces than your normal commercial tools,” he says. “If you had a task in which you were given a random dataset and you needed to find as many faces as possible – you would be better to use the free stuff, because it would find a higher percentage of faces in those images. It finds them with a greater degree of reliability – it is not just finding garbage.”

Thieme emphasised his company has no commercial interest in showing that social networking sites employ reliable facial recognition search engines. “I don’t have a horse in the race and personally I don’t care whether Facebook is good at doing this or not,” he said. “We just measure things.”

For more information
www.ibgweb.com