Since the outbreak of the Coronavirus disease in 2019, medical professionals have advised people to put on facemasks. Facemasks are considered a way of reducing the spread of the COVID-19 infection, by preventing the spread of droplets that could contain the virus.
As it stands now, facemasks do not only prevent the spread of COVID-19, it also prevents facial recognition algorithms from doing their job accurately.
Now that everyone is expected to put on a facemask, especially every time they are in public places, or whenever they’re not alone, the problem facial recognition algorithms are trying to solve, have just gotten tougher.
The National Institute of Standard and Technology (NIST), under the US Department of Commerce, decided to launch an investigation into how well facial recognition algorithms are faring, now that people’s faces don’t look the way they used to. The NIST found out that, while the algorithms are doing their jobs, they are struggling to get it done.
The results of the research was published yesterday as an NIST Interagency Report – NISTIR 8311. This is the first in a planned series from NIST’s Face Recognition Vendor Test (FRVT) program. The program is bothered on the performance of face recognition algorithms on faces partially covered by protective masks.
In the report, NIST found out that while facemasks have been successful in the fight against COVID-19, it has also been successful in increasing the error rate of facial recognition algorithm from 5% to about 50%, depending on the capability of the algorithm.
The NIST tested 89 facial recognition algorithm with the photo of a person on a digitally applied facemask, against the photo of the same person, without facemasks.
The masked images causes the facial recognition algorithms tested to encounter the “Fail to enroll or template” (FTE) error in most cases. The FTE error is a situation when the algorithm could not extract a face’s features well enough to make an effective comparison in the first place.
This is because facial recognition algorithms typically work by measuring a face’s features — their size and distance from one facial feature, to another.
The research also tested facial recognition algorithms against the amount of the nose covered. High, Medium, and Low levels were used. Studies found that the greater the amount of the nose covered, the lesser the accuracy of the algorithm.
Also, the result of NIST’s research shows that the colour and shape of the mask, affects the accuracy of the algorithm. Black masks reduced the accuracy of the algorithm more than blue coloured masks do. Round masks, however, have lower effects on facial recognition algorithms, compared to other shapes.
Mei Ngan, a computer scientist at NIST, and an author of the report, commented that:
“We can draw a few broad conclusions from the results, but there are caveats. None of these algorithms were designed to handle facemasks, and the masks we used are digital creations, not the real thing.”
In reaction to the research, Shaun Moore, CEO of TrueFace, whose algorithm was also tested, said:
“I do think that this is a solvable problem, and that it will require continued research and development efforts to close the accuracy gap. The more (mask) data that we are able to train our algorithms on, the better the performance will be.”
While the use of facemasks in a large scale might be strange to facial algorithm, it is expected that with more data, the accuracy of facial recognition algorithms will increase, and this will result in stronger and more effective facial recognition software.