Glossary
What is facial recognition
In short
Facial recognition is a computer vision technology that finds a face in a photo or video and compares it with samples to identify a person or confirm who they are.
Security and law · face recognition
MMOOZZGG.com editorsChecked: 1 min read
For a child
You know your friend in the yard even in a new hat: you remember his eyes, nose and smile. A computer can also learn to know faces if you show it many photos. This is handy, for example, so that a phone opens only for its owner. But you cannot change your face like a password, so take care of your photos and do not give them to just anyone.
In plain words
First the system finds a face in the image, then aligns it and turns it into a set of numbers, a kind of digital fingerprint. This fingerprint is compared with stored ones. To confirm that the phone owner is in front of the camera, it is compared with one sample. To find a person in a crowd, it is compared with a whole database. The more similar the sets of numbers, the more likely a match, and a threshold decides when two faces count as the same. Errors happen: poor light, a mask, age and too few similar faces in the training data lower accuracy.
For business
Companies use the technology for staff entry passes, payment and entry by face, checking a customer who signs up for a service remotely, and counting visitors. Under 152-FZ, Russia’s personal data law, features of appearance used to identify a person are biometric personal data. They may be processed only with the person’s written consent, except in cases set by the law. This leads to rules on protecting the database, on where it is stored and on an alternative for people who did not consent. The risks: leaks of face templates, which cannot be changed like a password, errors on some groups of people, and fooling the camera with a photo or a mask. Before launch, test accuracy with your lighting and cameras, and the protection against spoofing. Plan the storage of biometrics together with the data security of the whole project.
For a specialist
Facial recognition (face recognition) includes face detection, alignment by key points, extraction of an embedding with a convolutional network, and comparison by cosine similarity or Euclidean distance. There are two modes: verification 1:1 (is this the same person) and identification 1:N (who is this in the database). FaceNet (Schroff and co-authors, 2015) trained a network on triplets of images with a triplet loss function and reached 99.63% on the LFW dataset, describing a face with only 128 bytes. ArcFace (Deng and co-authors, 2018) added an additive angular margin to the loss function, which separates classes more strongly. Quality is set by the false match rate and the false rejection rate at a chosen threshold. Presentation attack detection rejects photos, videos and masks. Article 11 of Russia’s personal data law 152-FZ requires written consent to process biometrics. Article 5 of EU Regulation 2024/1689 bans untargeted scraping of face images for recognition databases and, with exceptions, real-time remote identification in public places for law enforcement.
Example
A business center moves staff entry passes to entry by face. Each employee signs written consent, people who do not agree keep their cards, and face templates are stored on a server in the building, not with a cloud provider.
Related terms
Questions and answers
How accurate is such a system?
Sources
- Schroff, Kalenichenko, Philbin, FaceNet: A Unified Embedding for Face Recognition and Clustering, 2015arxiv.org
- Deng et al., ArcFace: Additive Angular Margin Loss for Deep Face Recognition, 2018arxiv.org
- Federal Law No. 152-FZ of 27 July 2006 "On Personal Data", Article 11, ConsultantPlusconsultant.ru
- Regulation (EU) 2024/1689, Article 5, AI Act Service Desk of the European Commissionai-act-service-desk.ec.europa.eu
Not sure whether your company needs this?
AI solutions map