A visitor uses facial recognition system for buying a cup of coffee. Photo:Chen Xi/GT

A team of researchers from Wuhan University claim they have developed facial recognition software which can verify the identities of people wearing face masks.

The team said the software boasts 90 percent accuracy rate.

Existing facial recognition devices cannot identify individuals with face masks, and the new software aims to solve what has become an urgent problem as masks have become standard during the coronavirus epidemic.

The research was carried out by Wang Zhongyuan, a professor of computer science at Wuhan University, together with a team of over 10 students working remotely from their homes.

Even if the mask covers part of the face, it can still achieve accurate recognition by giving priority to the features in the upper half of the face, which is exposed to the camera, Wang told the Global Times on Tuesday, in response to whether faces will be mistakenly identified as a result of wearing masks.

"In a large number of controlled application scenarios, such as attendance checks in work places, security checks at train stations and facial scan payments, people usually have their front face to the camera. The image you can get is high quality and thus the recognition is not so difficult," he said.

To improve the recognition accuracy, Wang and his team resorted to the public as well as self-built datasets and worked on the analysis of features in the exposed parts.

"By analyzing the key features in visible parts of the mask, such as the facial shape, details around the eye, eyebrow and forehead, we improve the recognition accuracy from the original 50 percent to 90 percent," Wang said. 

"The 10 percent error often resulted from an unclear image, not using the front posture or very identical faces."

According to Wang, face recognition technology can already be used to identify people wearing masks but it is still not very reliable compared with the regular facial recognition technology which already witnessed an accuracy of over 99 percent.

At present, not many companies are working in the field to apply facial recognition technology to people wearing masks, Wang said.

"Based on our survey, Sense Time Technology reported a pass rate of 85 percent when the person exposes 50 percent of the bridge of the nose while Hanvon also reports about 85 percent accuracy rates. The best results reported so far are from MINIVISION with the recognition rate of people wearing masks reaches 90 percent," he said.

According to the team it requires multiple face images of the same person with and without a face mask to build a sample set, which makes it difficult to conduct.

So far the team constructed the world's first public masked face dataset, with 5,000 pictures of 525 people wearing masks, and 90,000 images of the same 525 people without masks. The database is open to the public and free to use.

According to the team, their research is based on relevant algorithms and a huge database collected from public data pools. 

During research and development, the team collected 360,000 raw data samples and developed a semi-automatic tool to clean and label data to enhance efficiency. 

In order to expand the diversity of data, the team developed a software program to add masks to facial images in the public data pool, which has generated 500,000 images of people in masks for 10,000 subjects.

The team has planned to work with technology companies such as Wuhan Citms Technology Co, Wuhan Hongxin Technical Services Co, and Wuhan Bontor Co to further develop their software for scenarios such as students entering and exiting compounds, and attendance checks in some workplaces.