Kate Smith-Miles, Deakin University in Australia
Listen Now Download as mp3 from the show Gauging Age, Virtual Life, Reading Emotions and Cyber-Forensics
Chris - Recently there have been several news reports on the dangers that young people can face when they get onto the internet. For example, there is some content that young people shouldn't be allowed to access in the first place. Now, how can a computer actually know the age of the person who is using it? Well at the moment it can't, but that could be about to change.
Kate - I'm Kate Smith-Miles, I'm head of the school of Engineering and IT at Deakin University in Australia. We've been looking at trying to get computers to be able to guess a person's age based on their face. This is a part of machine intelligence and there are lots of obvious applications.
Chris - Is it very difficult to do that Kate 'cause, I would have thought that it's relatively easy to just build a big database of what faces look like at different ages and then compare them?
Kate - It is, the typical approach that others have taken is to have a database of faces where you know their ages and say here's a face, this is the answer, here's another face, this is the answer and get a computer model to learn the relationships. We've taken a slightly different approach and I think it's the reason we're getting better performance than the existing methods. So we're taking photos of people at different ages in their life and we're getting the computer to learn what happens to the mathematical properties of the face over time.
Chris - But surely that's rather difficult because you've got to do a lifetime of tracking of an individual to see how they are going to change over say 70 years, aren't you? So how do you squash 70 years into the few years that you've actually been working on this?
Kate - Well we're very lucky that there are some existing databases of about 600 people with maybe 5 photos each throughout their life. We've got about 3000 faces so we're able to learn for a particular person what happens to the facial features over time.
Chris - So how does the program actually get to grips with a face? How does it take a face apart and work out what bits are what and how age affects those different bits of the face?
Kate - First of all, we're working with face images that are tightly cropped, so we're not looking at hair or ears or anything like that, just the actual face region. We then have a sort of semi-manual process identifying landmark features around the face, so the corners of the eyes and the tip of the nose and the bottom of the chin. We've got 68 points around the face that we landmark and then that, of course, is a series of numbers. We then tell the computer here is a set of numbers for a certain face and this is how old the person is and we get a mathematical model to learn the relationship between those inputs and the outputs.
Chris - So in other words, how those different landmarks change relative to each other over time to then map out what changes. How do you see this being used? What sorts of applications would it have in say the workplace or on the internet?
Kate - Well one of the things that is attracting most of the attention at the moment with our work is the applications in protecting children from adult material on the internet. For instance, I've got a 5 year old daughter. I walked into the study one day and she was spinning around and asking me what our password was for an online dating agency, and I thought, 'how did she get to that?' I hit the back button on Google and she'd gone into Google and typed, 'I want a friend,' and up had popped a whole series of adult related sites. Now the webcam was on it would have been great if we'd have had some software on the computer that had of recognized that she was just a child and when a child says, 'I want a friend,' they mean a completely different thing to when an adult says it. That's one application we're having a lot of conversations at the moment with ISP's. In Australia here's been some new legislation that was introduced last month that requires ISP's to be able to verify the age of computer user prior to making access available to adult material. At the moment the way they do it is with a popup question, 'are you over 18?' So I think there are a lot of applications of improved technology in helping to protect children from these sorts of things. Then, of course, there are the applications with vending machines and cigarettes and things like that.
Chris - Well you very kindly analyzed some of the faces of us from the Naked Scientists. I've actually got them in a sealed envelope so I haven't seen these. I'm going to have a look and see what you found when you looked at our mug shots that we sent to you.
Kate - I hope we haven't insulted anybody.
Chris - Well let's find out. You reckoned for Ben: 27, so that's not far off. Chris, that's me: 30. Now that's good because I'm actually 33 so you
thought I was a little bit younger! Dave: 35. Now Dave's actually in his late twenties so how do you explain the fact that Dave has been aged by so many years? He's gained 18% more years than he really has I think.
Kate - I'm not sure. I'm looking at the photo now myself. Compared to the 3000 facial images that the computer model has been trained on there's something in his face that he has some similarity with a 35 year-old. I don't know, maybe he didn't get much sleep the night before. Sorry, really not sure, we don't want to insult anybody.
Chris - Well you can be my best friend because you actually thought I was younger than I am so thank you very much for that.
Kate - Well you must get that a lot, people telling you that you look younger.
Chris - I do actually and it's really interesting that I think, 'why is that, why do people think I look younger than I am?' Do you have any insights into why some people actually do look their age and others don't?
Kate - No, I don't. I mean obviously humans have the advantage of looking at hair and body build and gesture and other things and voice as well. This computer was just looking at the face region alone but I wouldn't be at all surprised if you're constantly hearing that you look younger than you are and I don't know why. Is it skin tone? You know, there are a lot of things that our computer algorithm isn't considering as well. Humans, I should say, have a really hard time with this task. We rely very heavily on those additional factors. If you ask a human - if you give them a tightly cropped facial image and ask them to estimate how old somebody is they're not very good at it. In fact our algorithm is better than the human subjects we tested. So we can predict the age with our computer algorithm to within about 4 years for adults and 1 year for children, and humans the error is usually more like 6 years for adults.
Chris - So how long will it be before I can buy this and have this installed on my computer to stop my daughter finding a friend that she shouldn't find online?
Kate - I'd say we're a couple of years off having a commercial product. There's a number of challenges that we still face, first of which is of course even identifying where the face is in the webcam shot and then there's the landmarking process identifying all the facial features is a semi automated process, but we need to fully automate that. There's a number of challenges we're working towards at the moment.
Chris - I particularly like the way Dave turned out to be age 35.
Kat - Ha, I think it's the beard actually.
Chris - Dave hasn't got a beard.
Kat - Well he's a bit beardy.
Chris - Do you think it's the stubble?
Kat - Yeah. You're a baby face anyway.
Chris - You're very polite. Anyway, that was Professor Kate Smith-Miles. She's at Australia's Deakin University where she's developing a system to enable a computer to work out the age of the user based on what they look like.