Abstract

We describe a system for face recognition that has achieved 97% accuracy (with 20% classified as unknown) for 16 people under a range of lighting and viewing conditions. Recognition occurs in < 1 s on a Sun 4 computer. Our system characterizes the variation in face images from a small initial training set by performing an eigenvector decomposition similar to that performed by a Kohonen neural network. In the current system seven imagelike face basis vectors (termed eigenfaces) are extracted and converted into a distributed spatial-frequency representation. This conversion is accomplished by a windowed Fourier transform equivalent to a Gabor-filter representation. To locate and recognize people in a scene, the input image is converted into the same Gabor-like representation and is matched against the learned face basis vectors. Areas of the image that are facelike produce characteristic matches, indicating the presence of a person. The precise values of the match are then used to determine the identities of the people being viewed or to learn to recognize an unfamiliar face.

© 1990 Optical Society of America