Face recognition
Face recognition is a biometric technology that identifies or verifies a person by measuring the geometry of their face and comparing it against stored templates. It converts a face image into a numerical vector rather than storing the photograph, then scores how closely that vector matches an enrolled one.
How does face recognition work?
- Detection: the system locates faces within an image or video frame, which is a separate problem from recognising them.
- Alignment: the detected face is rotated and scaled to a standard position so that pose variation does not dominate the comparison.
- Template extraction: a neural network converts the aligned face into a fixed-length numerical vector, often called an embedding or template.
- Matching: the template is compared against one enrolled template (verification) or many (identification), producing a similarity score.
- Decision: the score is compared against a configured threshold, and the system reports a match, no match, or a candidate list for human review.
The stored template is not a photograph and cannot be viewed as one. It is a list of numbers describing measurements the network learned to make. That distinction matters for data protection: a compromised template database does not hand an attacker a set of usable face images.
What is the difference between face detection and face recognition?
Face detection answers where a face is in an image. Face recognition answers whose face it is. Detection is the cheaper, earlier step and is used on its own for autofocus, people counting and blurring; recognition requires an enrolled gallery to compare against and carries far heavier legal obligations.
The two are also measured separately, and conflating them is how a datasheet figure stops describing the system. A detection rate says how many faces were found in a frame; a recognition rate says how many of those were matched to the right person. A pipeline can be excellent at the first and unusable at the second, and only the second is what an identification deployment depends on.
Where is face recognition deployed?
- Border control and aviation, matching a traveller against the document they present.
- Building access, replacing a card at a door, gate or turnstile.
- Watchlist alerting in transport hubs and stadiums, where a small gallery is checked against a live camera.
- Device unlocking and payment confirmation, which is verification against a single template held on the device.
- Retrospective investigation, searching recorded footage for a person after an incident.
The deployment mode determines the legal analysis more than the algorithm does. Verification against a template a person supplied for that purpose is treated very differently from identification of passers-by who never enrolled, and several jurisdictions permit the first while restricting or prohibiting the second.
Frequently asked questions
- Is face recognition the same as facial recognition?
- Yes. Face recognition and facial recognition refer to the same technology and the terms are used interchangeably in industry, standards documents and regulation. NIST uses face recognition in its evaluation programme names, which is why that form appears more often in technical literature.
- Does face recognition store photographs?
- Well-designed systems do not. A face image is converted into a numerical template and the original image is typically discarded after enrolment. The template describes measurements rather than appearance and cannot be rendered back into a recognisable photograph, which limits the damage a template database breach can do.
- How accurate is face recognition?
- Accuracy depends on the dataset, matching threshold, gallery size, image quality and demographic group, so a single figure is not meaningful. NIST publishes comparative evaluations on sequestered data, though NIST does not endorse vendors. Ayonix publishes no accuracy benchmarks and points buyers to those reports and to pilots on their own cameras.
- Can face recognition run without an internet connection?
- Yes, on platforms built for it. On-premise systems such as Ayonix ATLAS AIBOX perform detection, template extraction and matching on local hardware, so recognition continues during an internet outage. Cloud face recognition APIs require connectivity because the matching happens in the vendor's data centre.
- What cameras does face recognition need?
- Most enterprise systems accept standard ONVIF or RTSP streams from existing IP cameras. What matters more than the brand is the pixels across the face, the mounting angle and the lighting. Identification generally needs more resolution and a more frontal view than detection does.
- Is face recognition legal?
- It depends on the jurisdiction and on the deployment. Verification with consent for access control is broadly permitted; identification of the public without consent is restricted in the EU under the AI Act and prohibited outright by several US city ordinances. Legal review belongs before procurement, not after.