Tested masked-face system reaches about 96% on one database

The nose and mouth disappear behind a mask. For conventional facial-recognition systems, that missing part can be enough to reject an authorized user—a weakness exposed when masks were often compulsory during the pandemic. Researchers have now tested a system that works instead with the eyes and forehead, reaching up to about 96% accuracy on one database.

The method combines several views of the face. A deep-learning model extracts features from around the eyes, known as the periocular region. Two other techniques—local binary patterns and histograms of oriented gradients—analyze the forehead before the system fuses the results to identify or verify a person.

That choice also changes the hardware requirements. Unlike iris recognition, which can require specialized near-infrared imaging and controlled conditions, periocular recognition can operate with ordinary visible-light cameras. The approach therefore targets a part of the face that remains visible without demanding the same specialized setup described for iris-based systems.

The result is not equally strong everywhere. The about 96% figure comes from one database, while tests on other test beds produced lower accuracy. The authors consequently describe the system as more suitable for initial screening than as definitive proof of identity.

So what does that change in practice? A masked-face match could help a system decide who should proceed to another authentication step, such as entering a password, PIN or using a security token. The researchers also say the model could add ears, face shape or profile; those extra features should improve performance significantly, though the paper presents that as a possibility rather than a completed result.