Wearing face masks is making it more difficult for the algorithms to identify people.
Face masks worn to slow the spread of COVID-19 are doing more than reducing exposure levels as they are also messing with facial recognition software accuracy.
A new report showed that algorithms have a tough time identifying people wearing the masks.
The report was issued by the National Institute of Standards and Technology (NIST), a US Commerce Department branch. Among NIST’s tasks is to measure facial recognition software algorithm accuracy. Its new report showed that when a picture of a person was tested alongside an image of that same person with a digitally added face mask, the algorithms failed to make the match 5 percent to 50 percent of the time.
Typically speaking, the failure rates for these algorithms usually land somewhere between 20 percent and 50 percent, according to NIST computer scientist, Mei Ngan, one of the report’s authors.
It’s not a mystery to the scientists why the masks reduce facial recognition software accuracy.
The algorithms usually function through a comparison of measurements among various facial features from one image to the next. When a portion of the face is covered, the software has less information to use for the comparisons in order to make its match.
This underscores a new challenge the tech industry has already been seeking to overcome, as the pandemic rages on, and more people are wearing face masks to help protect those around them. While this tech remains highly controversial, it remains used in a spectrum of products and services beyond law enforcement which has made the most headlines. For instance, this type of technology is used for passing through security checkpoints but is also a common method used for unlocking smartphones.
The researchers used nine different black and pale blue mask shapes to conduct the research outlined in the report. These mask colors and shapes were meant to represent the most common varieties worn in real life to cover their faces in the way they would be in typical real-life circumstances.
In total, 89 facial recognition software algorithms were tested on over 6 million images featuring a million different individuals. The photos were sourced from US immigration benefits applications (the unmasked images) and pictures of travelers entering the US across a border (which had digital masks added).