“Eigenfaces for Recognition,” by Matthew Turk and Alex Pentland of the MIT Media Laboratory, appeared in the Journal of Cognitive Neuroscience in 1991 (volume 3, issue 1, pages 71 to 86). It described a near-real-time computer system that could locate and track a person’s head and then recognize the person by comparing facial features to those of known individuals. It became one of the most influential early approaches to automatic face recognition.
The method’s core idea is to treat a face image as a single high-dimensional point - one number per pixel - and then find the directions of greatest variation across a collection of faces. Those directions, the principal components (eigenvectors) of the set of face images, are themselves face-like images, which Turk and Pentland called “eigenfaces.” Any face can be approximated as a weighted sum of a modest number of eigenfaces, so a whole face collapses to a short list of weights. Recognition then becomes a matter of comparing those weights against stored ones.
This was a sharp departure from the feature-measuring approach of the 1960s. Rather than asking a human to mark eye corners and mouth width, eigenfaces learned a compact representation directly from pixel data, motivated, the authors wrote, by “information theory” as much as by physiology. The technique is a direct application of principal component analysis to images.
The eigenfaces approach dominated face recognition research through the 1990s and remains a standard teaching example of dimensionality reduction. It set the stage for the FERET evaluations and, much later, for the deep-learning systems that replaced hand-built features entirely.