A new paper from researchers at Sony argues the importance of including red and yellow skin hues when examining skin tones through AI.
William Thong and Alice Xiang from Sony AI and Przemyslaw Joniak from The University of Tokyo found AI systems currently ignore a range of factors relating to the addition of yellow and red hues (the graduation of the skin colour). These are associated with warm and cool undertones that make skin tone diverse.
One current standard to examine skin tone with software utilizes a sliding scale that examines skin tone from light to dark. In 2o18, groundbreaking research from Joy Buolamwin and Timnit Gebru, found such facial recognition methods were biased towards women with darker skin tones, leading companies like Google to implement changes.
However, Sony’s research states systems need to go beyond just examining the lightness or darkness of skin tones. In a blog post, the researchers said this could be problematic for “East Asians, South Asians, Hispanics, Middle Eastern individuals, and others who might not neatly fit along the light-to-dark spectrum.”
One real-world example Sony included in its paper details how image datasets over-represent those with “light skin tones and red skin hues.” However, those with darker, yellow skin tones are underrepresented.
“Expanding the skin tone scale into two dimensions: light vs. dark and warm (yellow-leaning) vs. cool (red-leaning), can help to identify far more biases,” the blog post states/
The researchers point towards the CIELAB colour standard as a possible solution. When the researchers used this preexisting measure in its tests, they found photos of people with different skin tones varied in tone and hue.
According to Wired, Google and Amazon are among the players to welcome the report. Both companies have told the publication they’re reviewing the research.
Image credit: SonyAI
MobileSyrup may earn a commission from purchases made via our links, which helps fund the journalism we provide free on our website. These links do not influence our editorial content. Support us here.
