Computer trained to detect micro-expressions better than humans

Computer software that can detect the micro-expressions that humans often fail to pick up on is being developed at Oxford and Oulu Universities.

Micro-expressions last between a 25th and a third of a second and are widely noted as a way of recognising intentions that someone is trying to hide -- for example, as a method of detecting suspicious behaviour at an airport. As we reported last year, computer scientists at the University of Tsukuba in Japan are seeking to build datasets of micro-expressions associated with suspicious behaviour to help combat terrorism.

Tomas Pfister of Oxford University's Department of Engineering Science is heading up this latest project, which is aimed at producing software that will detect micro-expressions, and will be better than humans at doing so.

The team faced two hurdles: Firstly, says Pfister, "How can we get human training data for our algorithm when the expressions are involuntary? We cannot rely on actors as they cannot act out involuntary expressions." To tackle this problem, the team asked volunteers to suppress their emotions while they watched 16 "emotion-eliciting film clips". Says Pfister: "They were told that experimenters are watching their face and that if their facial expression leaks and the experimenter guesses the clip they are watching correctly, they will be asked to fill in a dull 500-question survey. This induced 77 micro-expressions in 21 subjects."

The second issue, he says, is actually capturing the micro-expression on film. "[Micro-expressions] occur for only a fraction of a second: this means that, with normal speed cameras, they will only appear in a very limited number of frames, leaving only a small amount of data for a computer to go on." To counter this, the team deployed a "temporal interpolation method", which filled in gaps in the data with existing data.

The initial findings suggest that the automated detection method is more effective than human detection. Says Pfister: "The human detection accuracies reported in literature are significantly lower than our 79 percent accuracy. We are currently running human micro-expression recognition experiments on our data to get a directly comparable human accuracy."

Pfister adds, however, that further experimentation is needed before micro-expressions could be deemed "conclusive evidence of deception". As he says: "Our initial experiments do indicate that our approach can distinguish deceptive from truthful micro-expressions, but we will need to conduct further experiments to confirm this."

This article was originally published by WIRED UK