Barriers to facial recognition technology are tumbling down

Tuesday, 15 May 2018

Kiwis appear divided over the revelation that supermarkets are using security cameras and facial recognition technology to pick out and keep watch on known shoplifters when they visit their stores.

About 60 per cent of people who responded to a Stuff poll said that is just what they would expect, while 40 per cent found that use of the technology 'creepy'.

**READ MORE

* Big brother is watching: Do you care?

* Police eyeing up newer, smarter CCTV facial recognition technology**

HOW DOES FACIAL RECOGNITION WORK?

By using a camera to capture an image of a face and using software to express the person's facial features, such as the relative size and position of their eyes, nose and mouth, in mathematical terms.

Software algorithms will then check that data against a database of previously stored patterns to try to find a match.

HOW ACCURATE IS FACIAL RECOGNITION TECHNOLOGY?

There is no single measure of accuracy, because there are two completely different ways in which facial recognition can be 'inaccurate'.

First, it could you tell that someone is, say, Lord Lucan when they are not. This is called a 'false positive'.

Secondly, it could tell you that someone was not Lord Lucan, when they really were. This is called a 'false negative'.

WHY IS THAT IMPORTANT?

It would be easy to set facial recognition software so it was 100 per cent accurate in terms of never giving false negatives. You would just twiddle the dial so it said anyone with two eyes was Lord Lucan.

It would be equally possible to ensure facial recognition technology never gave a false negative by turning the dial the other way, so the software was so sensitive it was never confident of a match.

There will always be a trade-off between 'false positives' and 'false negatives'; reduce one and you increase the other.

SO JUST SET THE DIAL HALF-WAY?

There are some applications where the user may really want facial recognition technology to avoid false positives.

For example, Customs does not want its Smartgate border control system to return a false positive, because it really wants to be sure the person coming into the country is the same person holding the passport.

But if it was using a security camera to try to pick out a known terrorist at an airport, it would probably be false negatives that were of most concern.

The airport would really not want it to miss the terrorist, even at the cost of a lot of 'false alarms' (false positives).

WHO USES IT?

Until recently, facial recognition has mainly been used by the likes of the Police and border control agencies for purposes such as crime prevention.

China is estimated to have more than 100 million security cameras on its streets.

Its use of facial recognition technology made headlines in April when police claimed to have used it to pick a suspect out of a crowd of 60,000 people attending a music concert.

But the technology is becoming a lot more accessible to businesses because of cloud computing.

Facial recognition can now be bought as a service in the cloud through Amazon Web Services (AWS) for example, with the processing carried out at its data centre in Sydney.

AWS advertises its 'Amazon Rekognition' cloud service as being able to perform 'real-time search' against tens of millions of faces.

IS THE TECHNOLOGY GETTING MORE ACCURATE?

Yes, and rapidly. There are three things driving this, better camera technology, better software algorithms and cheaper computing power which allows for more and faster processing of visual data.

AWS last month announced an update to Rekognition which it said made it 80 per cent more accurate in distinguishing between people who looked very similar to each other.

It also said it was about 30 per cent less likely to be tricked by people changing their hair colour, growing a beard or putting on glasses.

For the reasons above those numbers could benefit from a deeper delve, but they are indicative of the speed at which facial recognition is improving.

Facial recognition software can use the feedback it gets on whether it was right about its matches to tweak its algorithms through a process known as 'machine learning' – meaning it can automatically get smarter the more it is used.

Machine learning is faster in the cloud because it involves learning from lots of users with lots of data.

SO BIG BROTHER IS UPON US?

More like he is cruising down the road and looking like he might pull into the driveway.

Identifying people in a crowd from a distance is still difficult and a lot relies on the quality of the picture against which they are being matched.

The BBC reported on Tuesday that British police had trialled facial recognition cameras at football 'matches, festivals and parades' but reported a claim from a lobby group that the technology was highly inaccurate.

South Wales Police said its technology had made 2685 'matches' between May 2017 and March 2018 but that 2451 were false alarms.

That means it was generating a lot of false positives, which may not tell the full story. But certainly there are numerous reports of facial recognition trials for public safety being abandoned because of a lack of matches being made.

ANY GAME-CHANGING NEW TECHNOLOGY ON THE HORIZON?

Always. Japan's NEC is bringing to market new security camera technology that can automatically tell if multiple security cameras are pointing at the same person.

With multiple still images, or moving video, it becomes possible to draw up a 3D model of someone's face from a distance, which potentially greatly increases the accuracy of facial recognition technology in security-related applications.

If camera technology ever got to the point where it could reliably pick up skin prints in a crowd, facial recognition could become like finger-printing from a distance.

But for now it is a case of the steady advances in computing power and algorithms – and cloud computing – making more sophisticated facial recognition more available to more types of organisation for more uses.