Learn AI /Computer Vision and Language Technology /Facial recognition
Facial recognition
- Starter
- Computer Vision and Language Technology
- Human rights practitioners
- Policy officers
Also called: face recognition, biometric identification.
Back to the milestone you came from
In five seconds
What it is
A computer vision method that compares face images or representations to estimate whether they show the same person.
In plain words
It takes a face image, turns it into numbers that represent visible patterns, and compares those numbers with stored ones to estimate a match.
An everyday example
Imagine comparing a new photograph with pictures in a family album. You might first look for one particular person, which resembles one-to-one verification. Or you might scan the whole album for the closest-looking face, which resembles one-to-many identification. A blurry image or an unusual angle can make the comparison less reliable. The standard for deciding that two faces look similar enough plays the role of the system’s threshold.
A human staff member may use context, ask questions or notice unusual situations. A facial recognition system does not do that by itself. It compares patterns in pixel data and can produce a score even when the image is poor or the person is not in the database. Also, modern systems do not usually rely on a few hand-picked features in the simple way the story suggests; many use learned numerical representations that are harder for people to inspect directly.
Remember thisFacial recognition compares numerical representations of faces. Its output is a threshold-based estimate of a match, not certain proof of identity.
This term is contested
The basic idea of comparing face images is not contested, but the boundaries between verification, identification, analysis and broader 'biometric identification' are described differently across technical, legal and policy contexts.
Understand it
Understand it
Why it matters
Facial recognition matters because it can affect access, identity checks, policing, service delivery and surveillance. A non-specialist may be asked to buy it, regulate it, approve it, audit it or respond to its errors. To do that well, you need to know that it does not produce certainty, that verification and identification are different, and that performance can vary across settings and populations. You also need to recognise the governance questions around consent, necessity, proportionality, retention, oversight and remedies when mistakes happen.
How facial recognition turns an image into a match score
After viewing this, a reader should be able to describe the main stages from face image input to comparison result, and explain that the output is usually a score rather than certainty.
01
1. Input image
A photo or video frame arrives. It may contain one face, several faces or none.
Draw a rectangle representing a camera image. Include simple silhouettes of two people, with one face highlighted faintly to suggest the target region.
02
2. Find the face
The system locates the face area in the image so it knows what part to analyse.
Show the same image with a clear bounding box around one face. Add a small caption: 'face location'.
03
3. Convert to numbers
The face area is processed into a numerical representation, often called an embedding, that captures visual patterns useful for comparison.
Show the cropped face leading to a small grid or row of numbered dots labelled 'embedding'.
04
4. Compare with stored records
The embedding is compared with stored face embeddings from one person or from many people in a database.
Show one vector on the left and three stored vectors on the right, with lines connecting them and small labels such as 'similarity 0.91', '0.62', '0.40'.
05
5. Return a result
The system outputs a similarity score, a yes-no result based on a threshold, or a ranked list of possible matches. A human may then review it.
Show a result panel with 'top match', 'score', and a threshold line. Include a small human-review icon after the result.
A left-to-right flow diagram with five boxes linked by arrows. It starts with a photo, then shows face detection, feature extraction, comparison with stored templates, and a final result box with scores and possible actions. Read from left to right to see how one image becomes a match estimate.
On hover or tap, each stage reveals one sentence explaining what changes at that stage; a threshold slider updates the final match/no-match label for the example scores.
Describe this without the diagram
This diagram explains facial recognition as a five-stage process. First, a photo or video frame is the input. Second, the system finds the face location in that image. Third, it converts the face area into a numerical representation called an embedding. Fourth, it compares that representation with stored face representations, either for one claimed identity or across many records. Fifth, it returns a result such as a similarity score, a ranked list of candidates, or a match decision based on a threshold. The key message is that the system compares patterns and produces an estimate, not certainty.
With reduced motion: A static five-box flowchart shows the same process without animation: input image, face location, numerical representation, comparison with stored records, and result with score or threshold-based decision.
Facial recognition uses images or video that contain a human face as input. The system first tries to find where the face is in the picture. It then measures visible patterns such as the distances and shapes around the eyes, nose, mouth and jaw, or uses a trained model to turn the face into a compact numerical representation called an embedding.
Next, it compares that representation with other stored face representations. In verification, it checks a claimed identity, such as “Is this the same person as in this passport photo?” In identification, it searches many records and returns the closest matches. These are different tasks, and they have different risks.
The output is usually not a certain answer. It is often a similarity score, a ranked list of possible matches, or a yes-or-no result based on a threshold. A threshold is a cut-off chosen by people or organisations: above it counts as a match, below it does not. Changing the threshold changes the balance between false matches and missed matches.
People often confuse facial recognition with simply detecting that a face is present. Face detection only finds faces in an image. Facial recognition goes further by comparing one face with another or with a database. Measured performance can vary with image quality, lighting, age, pose, camera angle and the populations and conditions used for testing.
How “Sorting family photographs” maps onto the real thing
| In the story | In the AI system |
|---|---|
| The arriving visitor’s face | The input image or video frame containing a face |
| The staff member focusing on visible facial features | The system extracting numerical face features or an embedding |
| The photo board of registered attendees | The stored database of enrolled face representations |
| Saying a photo looks close enough | Using a similarity score and threshold to decide a match |
A worked example
A phone uses one-to-one facial verification to decide whether the person holding it matches the enrolled owner.
Starting from: A face image from the front camera and the owner’s enrolled face template.
- The camera captures an image, and the system detects and aligns the face.
- A trained model converts the detected face into an embedding, a compact numerical representation.
- The system compares this embedding with the owner’s enrolled template and calculates a similarity score.
- It applies the configured decision threshold. A score on the matching side of the threshold is accepted as a likely match.
- The result can still be affected by lighting, camera angle, image quality and the chosen threshold.
Result: The phone accepts or rejects the attempt based on the similarity score. Acceptance means the images were sufficiently similar under the system’s rule, not that identity was established with certainty.
In a United Nations workplace
During a staff town hall at an inter-agency secretariat, an internal auditor is asked whether a proposed lobby system can “confirm” every visitor’s identity. The front office wants a note for the file before close of business, ahead of the vendor demo.
The note should distinguish one-to-one verification from a search across enrolled records. It should also make clear that the system compares face embeddings against a threshold, producing a similarity-based result rather than proof of identity. Testing must cover error rates, image conditions and performance across relevant groups.
Before you try thisStaff and visitor face images are sensitive personal data. They must not be uploaded to a public or otherwise unapproved tool, including for a quick vendor test. No deployment should proceed without the required privacy, human rights, security, legal and procurement reviews, clear purpose limits, retention controls and a process for challenging incorrect matches.
Where you may hear it
- A vendor demo during a procurement meeting
- A staff town hall on new building-access arrangements
- An audit closing meeting on access controls
- A privacy review before a visitor system is cleared
Avoid confusion
Avoid confusion
Often confused with
Computer vision
Computer vision enables software to analyse images or video and produce useful outputs such as labels, locations, measurements or text.
Object detection
A computer vision task that finds instances of chosen object types in an image or video and marks where each one appears.
Image classification
A system assigns one or more predefined labels to an image based on patterns in its pixels.
Common misconceptions
Common misconception: If the system finds the right person most of the time, it is reliable enough for any use.
Correction: A high overall accuracy figure can hide serious problems. The real question is how the system performs in the specific setting, on the relevant population, at the chosen threshold, and what harm follows from false matches or missed matches. A tool acceptable for unlocking one person’s phone may be unacceptable for searching a large population or making high-stakes decisions.
Common misconception: Facial recognition and face detection are the same thing.
Correction: They are related but different. Face detection only locates a face in an image. Facial recognition compares faces to estimate whether they belong to the same person or to find likely matches in stored records.
Common misconception: A match means the computer has proved who the person is.
Correction: Usually it means the system found a face pattern that is similar enough under its settings. Similarity is not proof of identity. Image quality, ageing, pose, lighting, masks and database composition can all affect the result, so human review and context may still be necessary.
Common misconception: The system works the same way and equally well for everyone everywhere.
Correction: Performance can vary with camera quality, environment and the data used to build and test the system. If some groups or conditions were less well represented, error rates may differ. That is why independent testing in the actual context of use matters.
Quick check
Which description best fits facial recognition?
Facial recognition is about comparing faces for verification or identification. Detecting that a face exists is a different task. Inferring emotion is a separate and contested task. Writing image descriptions is image captioning, not facial recognition.
Go deeper
Go deeper
The simple version
It takes a face image, turns it into numbers that represent visible patterns, and compares those numbers with stored ones to estimate a match.
The technical version
A typical pipeline first detects and aligns a face in an image, then uses a trained model to generate a feature vector, often called an embedding. Similarity between embeddings is measured with a distance or similarity function such as cosine similarity. In one-to-one verification, the system compares the probe image with one enrolled template for a claimed identity. In one-to-many identification, it searches a gallery of enrolled templates and returns top candidates. A decision threshold converts the similarity score into match or non-match, which creates a trade-off between false acceptances and false rejections. Real-world performance depends on image capture conditions, enrolment quality, demographic and environmental variation, and how the system was evaluated.
Reviewed external resources
EncyclopaediaFacial recognition (opens in a new tab)Wikipedia
This link gives a broad introductory overview of facial recognition and suits readers who want a general starting point.
Intermediatelink checked 2026-07-31
ArticleMachine Learning Glossary (opens in a new tab)Google
This link provides short glossary-style definitions of machine learning terms and suits readers who want quick clarifications while reading.
Intermediatelink checked 2026-07-31
CourseMachine Learning Crash Course (opens in a new tab)Google
This link offers beginner-friendly machine learning lessons that help readers place facial recognition within the wider field and suits newcomers.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Facial recognition (opens in a new tab)YouTube search
This link is a YouTube search results page for video explanations on facial recognition and suits readers who prefer to compare several video styles before choosing one.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Facial recognition (opens in a new tab)Reddit search
This link is a Reddit search results page for community discussion about facial recognition and suits readers who want to see common questions, opinions and debates with caution about reliability.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
Facial recognition compares numerical representations of faces. Its output is a threshold-based estimate of a match, not certain proof of identity.