Algorithm better than humans at recognizing micro facial expressions
Researchers at the Chinese Academy of Sciences have developed an algorithm that is as good as or better than humans at recognizing microexpressions. Those are small, involuntary facial expressions that indicate emotions.
Training the algorithm was challenging for the research team because the emotions must be real and their expression must be suppressed by the observer. To achieve this, they had a group of subjects watch a movie that evokes certain emotions, while their faces were filmed with a high-speed camera. They are instructed not to show their emotions. If they do, she was told for motivation, then they have to fill in a large and boring form for each emotion shown.
The film gave the researchers a good idea of which emotion is at stake at any given moment. In this way, the researchers could teach the algorithm which twitching muscles in the face belong to the different feelings. The starting point for the algorithm is a frame from the video that does not contain any emotion.
The biggest challenge in building the algorithm is the format of the expressions. They are so small that the computer has difficulty recognizing them. As a solution, the researchers have the algorithm digitally enlarge the facial expression. For example, very slightly raised eyebrows are transformed into an expression of extreme surprise. That makes it easier for the computer to name the emotion. The image is also captured in other color spaces to make the emotions more visible. In addition to rgb, CIELuv and CIELab are also used and a new color space that the researchers call tensor independent color space, or tics. This four-dimensional array makes the work of the algorithm as easy as possible.
Image: MIT Technology Review
When the trained algorithm has to compete with the human ability to name micro-expressions, it goes at least more or less evenly. When the algorithm and 15 human opponents are tasked with both signaling and naming micro-expressions in a video recording mentioned above, the two sides score roughly equal. When frames are taken that are pre-determined to show a micro-expression, the algorithm outperforms the human opponents.
The researchers believe the algorithm could be applicable to things like lie detection, mental health, and law enforcement. The research appeared in IEEE Transactions on Image Processing.