Smile: you’re on camera!

Font Size

Print

Feedback

Email

Discuss

: digital album. Imagine, also, the uses genuine face recognition would have in man-machine communication, let alone in analysing video from surveillance cameras for terrorists and criminals and comparing them against a rogues’ gallery.

For the security benefits alone, governments everywhere have perked up to the possibilities of face recognition. There exists a growing awareness that information about a person’s state of mind and intention can be gleaned from that person’s image. Computers scanning the images can then act immediately to warn or prevent acts of violence from occurring. But that’s a lot easier to imagine than implement. Photographic images contain thousands (sometimes hundreds of thousands) of more background patterns than face patterns. That can make detection either incredibly slow or unacceptably sloppy. A practical system would need to achieve at least a 90% success rate with no more than one in a million false positives (ie, a region of the image being declared a face when it was not).

The task is one of the most challenging in the whole field of pattern-recognition. That’s because faces are far more ‘elastic’ than almost all other objects needing to be recognised autonomously—such as furniture in a room for a domestic robot, or vehicles on a road for a self-driving car.

For one thing, faces have to be recognised when they are not necessarily head on. For another, they may be looking up or down, as well as in profile or posed at an angle. They may have beards, mustaches, spectacles, all of which can be different shapes and colours. They may be partly covered by other objects. Meanwhile, lighting conditions can alter appearances considerably. And, yes, the subjects may be smiling or expressing other emotions.

Humans can sort all this out in the blink of an eye. Machines struggle manfully to find the proverbial needle in a haystack. Most automated detection systems have to be ‘trained’ using collections of facial images, each with different poses and lighting conditions. So far, researchers have focused mainly on detecting faces amid a jumble of other patterns, and then tracking them as they move within a sequence of frames.