Rajeev Sharma has a new computer system in his lab at Pennsylvania State University that really separates the boys from the girls.
The computer uses pattern recognition software to decide whether an image is that of a male or female.
"It's a learning technique," says Sharma, an associate professor of computer science and engineering. By trial and error, the software has "learned" how to separate images by gender.
The system was fed 1,755 thumbnail images of human faces from a standard database and told to determine the gender of each by "looking at the geometry of the facial structures," Sharma says.
Not Just Another Pretty Face
The images included only eyes, nose and mouth — eliminating such things as beards and hair — and the system initially did fairly well, learning which gender to assign each image just on the basis of visual cues.
But it wasn't good enough. So Sharma and his team of researchers took it a step further and added an audio signal consisting of tiny snippets of male and female voices a fraction of a second long.
And that, he says, puts the system way over the top. It gets it right nearly 100 percent of the time, he says, and that's significantly better that humans can do. We only get it right a little more than 90 percent of the time, given the same cues, he says.
So in one sense, at least, scientists have come up with a computer that's better at sex than we are.
One of the reasons the system succeeds, he says, is there is considerable difference in the geometrical structure of male and female facial features.
"Fortunately, there is enough difference" for the system to get it right, just on the basis of facial cues, most of the time, he adds.
Throwing Out Bad Clues
But sometimes, the cues are not all that clear, and that's where his system really shines.
"There are situations where the video is not good, or the audio is not good," and the system receives conflicting signals, Sharma says. Men who have watched a beautiful head of hair moving down the street, and then were startled to see a beard when the "woman" turned around, can identify with that.
So Sharma's system has learned how to do what humans do routinely when getting such conflicting signals.
It decides which signals to throw out.
"It decides automatically which one to ignore, and which one not to ignore," he says. That's possible because the decision making process, and the image and audio recognition systems, are "fused" into a computer "manager" that makes the decision based on weaknesses in the face or voice cues.
Toward Safer Skies and Better Market Research?
Of course, the purpose of the system is not just to tell the boys from the girls. It's part of a much broader research program, sponsored partly by the National Science Foundation, to develop all sorts of image recognition systems. Sharma's lab is also developing systems that can classify images by age group, and even recognize specific individuals.
There are many potential applications for such systems. Near the top of the list these days are computers that can screen faces to be sure that the person trying to enter a secured area — like an airport operations center, for example — is really who he or she claims to be.