Facial recognition researchers look for a culprit in gender inequality, come up empty
A group of researchers has looked at possible reasons that biometric facial recognition accuracy is different for men and women, but they could not pin the blame on any of several proposed factors.
Researchers from the University of Notre Dame and the Florida Institute of Technology published a paper January 31 that found that net matching accuracies continued even after they trained an advanced deep-learning method from scratch with two data sets “explicitly balanced between male and female images and subjects.”
Factors examined were facial expression, head pose, forehead “occlusion,” use of makeup and the practice of researchers drawing images from a data set not balanced between men and women. None were found to be the culprit.
Facial expressions were looked at because women tend to be more expressive in photographs than men, something that could in theory make differentiating women’s images more difficult. Head pose examined the angle at which the face in an image was tilted, something also is more common in women’s photos.
Forehead occlusion means the area of a forehead that is obscured by hair or hats.
Likewise, makeup did not confuse facial-recognition systems, even though makeup can inadvertently make two people look alike.
The one area many data scientists have perhaps speculated about the most publicly is training dataset imbalance. Women’s images often are underrepresented in data sets, according to the paper’s authors. But when custom, equally balanced data sets were created, researchers found that imbalance is not likely the main factor, as the same distribution patterns were found.
The teams said their results could prompt further research into a cause that is “more intrinsic” to what make male images and female images different. Specifically, they suggest examining face morphology, which refers to the bone structure inherited from parents.
Physiological differences are suspected by some researchers of making it difficult to produce voice recognition technology which matches women’s voices as accurately as men’s.