- Input face image
- Face detection
- Normalization: rotation, scale, normalization
- e.g. normalize distance between eyes
- Face feature extraction
- Interesting feature compression techniques available
- Classifier: feature matching against database
- Decision maker
- Thresholding
Early face detection:
- Geometric (e.g. using eyes)
- Eyes and mouths are good features to detect
- Color distribution
- Need to segment background from face color
- Skin tones ended up accounting for only a tiny portion of the color space
- Wood color is similar to skin color, which makes for much fun
- Hence, choice of color space is important
- Need to segment background from face color
Surveillance cameras usually high up, which isn’t ideal for most face recognition algorithms.
Surveillance-based tracking: tracking people for the entire time they are in an area.
Normalization:
- e.g. normalize distance between eyes
Features:
- Eyebrow thickness, vertical position at eye center
- Eyebrow Arch
- Nose vertical position, width
- Mouth vertical position, width, lip height
- Chin shape (e.g. distance from keypoint at fixed angles)
- Bigonial breadth - face width at nose position
- Zygomatic breadth - face width halfway between nose tip and eyes
Neural networks:
- Use NN-based filter: use small filter window to scan portions of the image and detect if a face exists
- Merge overlapping detections to eliminate false positives
Static face recognition:
- Eigenface: reduce face to many eigenvalues, project to high-dimensional feature face and find closest match. Worked up to ~1000 faces
- Linear and Fisher discriminant analysis
- Fishface: finds linear trasnformation which maximizes inter-class scatter while minimizing intra-class scatter
- Now: CNNs/deep learning - can work well even with millions of faces
Video face recognition:
- Low quality, small images
- But also allows tracking of the face image and continuity: re-use classification information from high-quality images when processing low-quality images
- Motion structure: create 3D model of face to match against frontal views
- Non-rigid motion analysis