Face recognition is a type of artificial intelligence that helps a computer identify or compare faces in images and videos. It matters because similar ideas are used in phone security, photo organization, accessibility tools, and safety systems. The goal is not for the computer to see a face like a human does, but to measure patterns in the image and make a statistical decision.
Good face recognition systems must also be tested carefully for accuracy, privacy, and fairness.
Key Facts
- A digital image is a grid of pixels, and each pixel stores color or brightness values.
- Face detection finds where a face is located before recognition tries to identify whose face it is.
- Feature extraction converts a face image into numbers called an embedding.
- Similarity can be measured with distance, such as d = sqrt((x2 - x1)^2 + (y2 - y1)^2).
- A classifier or threshold decides whether two face embeddings are close enough to count as a match.
- Accuracy = correct predictions / total predictions.
Vocabulary
- Pixel
- A pixel is one tiny square of a digital image that stores color or brightness information.
- Face detection
- Face detection is the step where a computer locates a face in an image before trying to recognize it.
- Feature
- A feature is a measurable pattern in data, such as the shape of the eyes, spacing between facial points, or texture in an image.
- Embedding
- An embedding is a list of numbers that represents important information about a face for comparison.
- Training data
- Training data is the collection of examples used to teach a machine learning model how to make predictions.
Common Mistakes to Avoid
- Thinking face recognition memorizes a whole face image. It usually stores and compares numerical patterns, not a perfect copy of every photo.
- Skipping face detection before recognition. The system first needs to find and align the face so the later comparison is meaningful.
- Assuming a match is always certain. Face recognition uses probabilities or similarity scores, so lighting, pose, image quality, and thresholds can change the result.
- Ignoring bias in training data. If a model is trained on an unbalanced dataset, it may perform worse for groups that are underrepresented.
Practice Questions
- 1 A model correctly identifies 92 faces out of 100 test images. What is its accuracy as a decimal and as a percent?
- 2 Two simplified face embeddings are A = (2, 3) and B = (5, 7). Use d = sqrt((x2 - x1)^2 + (y2 - y1)^2) to find the distance between them.
- 3 A school wants to use face recognition for attendance. Explain two benefits and two risks they should consider before using it.