Face recognition: Algorithms geared towards protective masks often fail

Many manufacturers of technology for biometric face recognition have recently claimed that they have adapted their algorithms to nasal and mouth protective masks and that they are now able to identify people who are partially hidden. However, the latest results of a study by the US standards authority NIST show that the masks also significantly increase the error rates of new systems.

In one on Tuesday published update on previous investigations The Institute for Standards and Technology examined 41 algorithms for automatic face recognition that developers submitted after mid-March during the Covid-19 diseases that were spreading rapidly in Western countries. Many of them are said to have been specifically designed or redesigned with facial masks in mind. A first test by NIST with algorithms that were not adapted to masks had shown that even the 89 best of these had error rates between five and 50 percent.

According to the new findings, even the adapted systems with partially covered faces have difficulties. Sometimes they let themselves be completely confused with masks, the false detection rate was increased by up to 99 percent. Rank One, for example, whose systems are used in cities like Detroit, had an error rate of 0.6 percent for uncovered faces. This increased to 34.5 percent as soon as the researchers digitally attached masks.

Don’t miss any news! With our daily newsletter you will receive all the news from heise online for the past 24 hours every morning.

  • Subscribe to the newsletter now

In May the company brought the new service “Periocular Recognition“on the market, which is supposed to be able to identify people directly by the eyes and parts of the nose. Rank-0ne managing director Brendan Klare told the US magazine CNet but that it was not yet possible for the company to transmit the algorithm to the NIST, as this only allows one entry per organization. The updated study therefore does not yet take into account the latest technology.

The result looks very similar with the manufacturer TrueFace, whose products are used in the USA, for example, in schools or Air Force barracks. The false detection rate rose from 0.9 to 34.8 percent with masks attached. The company management had told CNN in mid-August that an algorithm for an even better consideration of the covers is currently in work.

The Chinese company Dahua performed better with its technology, in which the error rate increased from 0.3 to six percent. The Amsterdam-based provider VisionLabs achieved a comparably low false detection rate of 3.5 percent despite masks, compared to 0.28 percent for uncovered faces.

NIST wants to continue its test series. To do this, she uses a database with around six million portraits. If the monochrome masks were not first added in the computer, the error rates would be even higher: Real fabrics have different patterns, textures and tones that could also negatively affect the algorithms.


To home page