Medical mask no longer saves face recognition

If you thought that a medical mask would deceive face recognition cameras, then there are two bad news for you. Firstly, the researchers were able to significantly improve machine vision systems, so that recognition is now quite reliably performed in half of the face or in the area of ​​the eyes (half in the face, the level of successful recognition is 90% ).

The second bad news is that the outbreak of the coronavirus prompted Chinese manufacturers SenseTime, FaceGo, and Minivision to introduce partially closed face recognition technologies into commercial camcorder models . Due to the outbreak of Covid-19, many citizens began to take to the streets in masks - therefore, it is necessary to modernize video surveillance systems.

New forms of face recognition can now recognize not only masked people covering their mouths, but also people in scarves or with fake beards. One of the first scientific papers on this topic was published back in 2017, this article is “ Disguised Face Identification (DFI) with Facial KeyPoints using Spatial Fusion Convolutional Network ; arXiv : 1708.09317v1 ).


Samples from the neural network training dataset

As you know, face recognition works by identifying several key points on a person’s face - and their connection, resulting in a unique “graphic” signature. These key points are usually found around the eyes, nose, and lips. So that the system can work with the lower half of the face closed, the researchers placed more key points around the eyes and nose.


The structure of the convolutional neural network in the DFI system The

neural network in the DFI system finds 14 key points in the face photograph, but the accuracy decreases depending on the level of masking and the complexity of the background behind the person.

However, since 2017, more research has been done on this topic, and now it is obvious that the technology has great commercial value. SenseTime, the Chinese leader in AI development, was the first to adapt its face recognition system, which the company announced last week.

A SenseTime press release says its algorithm is "designed to read 240 key points of the face around the eyes, mouth and nose." He can find a match using only those parts of the face that are visible. In other words, key points even around the eyes may be enough to create a unique imprint, albeit a partial imprint of the face.


SenseTime System

Researchers from Bradford University under the direction of Professor Hassan Ugail in May 2019 reportedabout the improved model of face recognition, having achieved recognition accuracy of 90% in half of the face and 100% in three quarters of the face. The scientific article “Deep face recognition using imperfect facial data” was published in Future Generation Computer Systems (doi: 10.1016 / j.future.2019.04.025 ).

Another Chinese facial recognition company, Minivision, claims that their software is now also able to recognize masked people. Faced with a flash of Covid-19 and a massive exit to the street of masked people, Minivision launched an emergency campaign to collect data to further train the model. “Management urgently mobilized employees and relatives to collect a limited set of data in two days. The key information that the system recorded on masked faces was the eyes, ” writes Abacus.

The rush is caused by China’s tough response to the epidemic. In many residential areas most affected by the virus, entry is restricted only to residents of the area. Minivision has implemented a new algorithm in its face recognition systems to block gates in communities in Nanjing to quickly recognize residents without having to take off their masks.



SenseTime and FaceGo programs are used primarily for recognizing company employees (for accounting working hours).

When the sample is limited to residents of one district or company, the task of the face recognition system is simplified by an order of magnitude. Expanding this system to a wider group of people will be difficult. When the sample reaches a certain scale, the system is more likely to come across people with similar eyes. In this case, the risk of false positives increases.

However, biometric systems are developing rapidly. Perhaps someday cameras will be able to read even iris and fingerprints from a distance. Remote sensors for heartbeat, body temperature, and gait identification systems are being developed. In addition, people often carry smartphones and other electronic devices with which they can be discreetly identified.

---

For more information about PKI solutions for enterprises , contact GlobalSign managers +7 (499) 678 2210, sales-ru@globalsign.com.