Assessing skin hue to skin tone makes measuring biometric bias easier — research
A team of Sony and Tokyo University researchers say they have created a better way to measure apparent skin color in computer vision, a task at the heart of worries that facial recognition algorithms might never grow beyond harmful biases.
In their pre-print paper posted on Cornell University’s arXiv server, the scientists say their proposal is a “simple, yet effective, first step towards a multidimensional skin color score.”
They have added a skin hue axis – from red to yellow — to simple skin tone for measuring. The result, they say, is a multidimensional color scale in an x-y chart combining tone and hue for assessing fairness in algorithms.
Conventionally, industry and government use Fitzpatrick skin classification to diagnose skin-color bias in computer vision. But Fitzpatrick is only a measure of skin tone, from light to dark.
Bias can be bad enough when software misidentifies people, but it can be injurious to people’s health, too. The team (two from Sony and one from Tokyo University) cite two medical examples related to skin classification.
It can render less accurate diagnoses of skin lesions or erroneously record heart rates.
The combination of tone and hue helps better read biometric data from, or in the case of synthetic models, attributed to people. White skin darkens with ages but also reddens. Asian skin also darkens but also grows yellower.