Top-Level World Ranking for Fujitsu's Mask-Wearing Face Recognition Technology in latest U.S. NIST Evaluation

Fujitsu is a leading contributor to this research, and in the latest National Institute of Standards and Technology’s (NIST) Face Recognition Vendor Test, we achieved the top ranking in the Japanese vendor and third place overall in the world for the mask-wearing category.
As we all become more used to mask wearing continuing in the future, this is likely to become an ever more important face recognition tool.

Background to the Face Recognition Vendor Test (FRVT)

NIST’s Face Recognition Vendor Test (FRVT) provides a fair benchmarking evaluation of authentication accuracy across the many existing face recognition technologies. FRVT uses the faces of more than 1 million people of different races and ages to evaluate large-scale facial recognition engines that can be used in a wide range of applications, including passport and visa authentication. More than 100 organizations, both in Japan and abroad, typically participate in the FRVT process.

In August 2021, NIST held an accuracy evaluation using facial images wearing a mask, with Fujitsu’s participation involving using facial images for identity verification. Our technology ranked third overall in the world, securing an impressive first place among all the Japanese vendors.

Increasing face recognition accuracy with our natural mask synthesis technology!

As proved in the FRVT, our current technology delivers highly accurate face recognition even when someone is wearing a mask. When people wear a mask, most areas of their face are covered, and this in turn changes their overall appearance. With conventional existing face recognition technology, a face covered by a mask results in incorrect recognition. As a result of this problem, we have developed a new technology that generates and learns each face image, based on combining the facial image with and without a mask. As demonstrated, this technology has already realized highly accurate face recognition even when wearing a mask, absorbing the differences in someone’s appearance as a result of the mask’s presence or absence. Our advanced technology means that highly accurate face recognition can now take place in indoor public places, without the need to remove the mask.

Mask synthesis technique applied to face image.

■Developer Comments

・Tomoaki Matsunami, Senior Researcher, Advanced Converging Technologies Laboratories.

With the demand for facial recognition when wearing masks growing, I am personally delighted that we have received such a strong result in this important fair benchmark evaluation test, involving face recognition vendors from around the world.
Our team at Fujitsu R & D Center Co., Ltd. (FRDC) was responsible for overseeing the process of learning face recognition models, and I was involved in promoting the face recognition algorithm development for the system as a whole. All of the team members worked closely together on the research and we achieved some impressive results. We will continue to cooperate closely, and aim for even higher rankings in future.

・Zhang Meng, Research Fellow, Fujitsu Research & Development Center Co., Ltd. (FRDC)

Despite the success of deep learning models for general facial recognition scenarios, the deep features still show inferior performance in wearing a mask, as not all of the facial image can be accessed for the description building. We proposed a low-cost and accurate mask shifting method that synthesizes the mask onto the non-masked face while maintaining pose and illumination consistency. This means that it can add more natural-looking masks when compared to traditional data augmentation methods. We then proposed an attention-aware scheme for masked face recognition, which puts more emphasis on key facial features while ignoring the ones that are corrupted by the face mask.
I think our success is mainly down to the efficient cooperation between our team members, creating a real collaboration and pooling our knowledge to improve our techniques.

・Dr. Guo Song, Researcher, Fujitsu Research and Development Center Co., Ltd. (FRDC)

To improve the accuracy of face recognition technology using deep learning, we constructed an extensive face data set that includes both masks and non mask-wearing images. We also developed a new learning method for this data set.
I believe that these technologies contributed significantly to our good results at FRVT.
We have a really good team that is working closely together, and have made significant progress.

Looking Ahead

Our focus is on developing a new face posture-based robust face authentication technology that enables authentication without the need for the person to face the camera while wearing a mask. We also plan to expand our technology’s range as a biometric authentication technology in the new normal era, by applying it next to cashless settlement applications that can be used while wearing a mask.

Related Links

・Fujitsu Delivers Cashless, Contactless Retail Experience for Masked Shoppers with Hygienic, Multi-Factor Biometric Authentication Technology[Fujitsu Global Site] (January 21, 2021 Press Releases)

For more information on this topic

fj-act-sd-face-frvt@dl.jp.fujitsu.com

Please note that we would like to ask the people who reside in EEA (European Economic Area) to contact us at the following address.
Ask Fujitsu
Tel: +44-12-3579-7711
http://www.fujitsu.com/uk/contact/index.html

Fujitsu, London Office
Address :22 Baker Street
London United Kingdom
W1U 3BW