SA scientist develops facial recognition technology identifying faces through face masks
Johannesburg - A South African scientist, Ishmael Msiza, has developed the world’s first facial recognition technology prototype able to recognise human faces through face masks.
Msiza, 35 from Kameelrivier in KwaNdebele, Mpumalanga, told The Star on Wednesday that within biometric recognition there were different modalities such as fingerprint recognition, facial recognition or iris recognition, however fingerprint recognition was the most accurate and dominant.
“When we were hit by this pandemic, we were told we need to be careful when we touch surfaces, and fingerprint recognition exposes you to that because you have to place your finger on a scanner, and the scanner is used by a number of people,” he said.
Realising a gap in the market, the scientist and his team started to look for a contact-less modality such as facial recognition, but there was a Catch-22 because of masks.
Conventionally, facial recognition systems authenticate a person based on their unique facial features from the area around the forehead, all the way to the area around the chin.
“What makes this dynamic of masked faces interesting and – in most cases – difficult, is the fact that a facial recognition system should attempt to accurately identify a human face with less than half of the facial landmarks exposed. This is because a Covid-19 mask obscures the nose, both lips, the jaws, and the chin. It only exposes the forehead, the eyebrows, and the eyes,” Msiza said.
Undeterred by the challenge, Msiza and his team had to create a solution that blended conventional image processing techniques and artificial intelligence (AI).
“In the past, practitioners could use conventional techniques or AI, but now when you send a facial image that has been obscured by a Covid-19 mask to that artificial neural network, you are making its learning task complicated. One way of simplifying its task is to pre-process the data using conventional image processing techniques,” the scientist said.
Msiza added that pre-processing the data could include converting an image from colour to grayscale using conventional techniques, or enhancing the contrast of the grayscale image before adding the data to the AI model for it to learn the pattern.
The real applications of this technology range from banking, criminal suspect identification, access control, attendance monitoring, time monitoring, border control, electronic visitor management, office security, home security, estate security, and device security.
Lunga Masuku, managing director of Lungsta Investments and financier of the project, said it was high time for South Africa to have its seat at the proverbial innovation table.