Sign in to save

Bookmark this page so you can find it later.

Sign in to save

Bookmark this page so you can find it later.

Facial recognition is a computer vision method that identifies or verifies a person by analyzing patterns in a face image. It matters because it is used in phone unlocking, airport security, photo organization, access control, and missing person searches. The basic idea is to turn a face into numerical data that a computer can compare with stored examples.

Good systems must handle changes in lighting, angle, expression, age, camera quality, and background.

Key Facts

  • A common pipeline is capture image, detect face, locate landmarks, align face, extract features, compare embeddings, then decide match or no match.
  • Face detection finds where faces are in an image, often by predicting a bounding box around each face.
  • Facial landmarks are key points such as eye corners, nose tip, mouth corners, and jaw points used to align and normalize the face.
  • An embedding is a vector of numbers that represents facial features, for example v = [0.12, -0.45, 0.88, ...].
  • Euclidean distance between embeddings can be computed as d = sqrt((x1 - y1)^2 + (x2 - y2)^2 + ... + (xn - yn)^2).
  • A decision threshold controls the tradeoff between false accepts and false rejects: match if distance < threshold.

Vocabulary

Face detection
Face detection is the step that finds the location of one or more faces in an image.
Facial landmark
A facial landmark is a specific point on the face, such as an eye corner or the tip of the nose, used to measure face geometry.
Embedding
An embedding is a compact numerical vector produced by a model to represent important facial features.
Similarity score
A similarity score is a number that tells how closely two face representations match.
Threshold
A threshold is the cutoff value used to decide whether a comparison counts as a match.

Common Mistakes to Avoid

  • Thinking facial recognition stores a normal photograph as the main identity record is wrong because many systems compare numerical embeddings rather than raw images.
  • Ignoring lighting and camera angle is a mistake because shadows, blur, and head rotation can change the input and reduce accuracy.
  • Using one fixed threshold for every situation can be wrong because high security systems usually need a stricter threshold than low risk applications.
  • Assuming a high accuracy number means equal performance for everyone is wrong because datasets, lighting, age, skin tone, and camera conditions can affect error rates differently.

Practice Questions

  1. 1 A system detects 92 faces correctly out of 100 face images. What is its detection accuracy as a percent?
  2. 2 Two face embeddings are A = (0.2, 0.7) and B = (0.5, 0.3). Calculate the Euclidean distance between them.
  3. 3 A phone unlock system uses a very strict threshold and starts rejecting its real owner more often. Explain why changing the threshold affects both security and convenience.