The detection and recognition of human faces in a cluttered background in real-time is being investigated. We propose a system which detects facial features from an edge map using a symmetry operator. The locations of key facial features are then used to normalize the facial image. Two approaches for recognition are being investigated. One method, similar to Pentland's, first employs the Karhunen-Loeve expansion of the normalized facial image. The resulting eigenvectors are used to search for a matching face in a database. The second approach constructs a feature vector from direct measurements of the face in the image. Ultimately it is hoped to track the face in real-time as it moves about the scene.
T. Jebara, M.D. Levine