[ PCO Home Page ] [ Table of Contents of NEWSLETTER # 68 ]

........ published in NEWSLETTER # 68

By Professor H.Wechsler, George Mason University, Fairfax VA (U.S.A.)

A NATO Advanced Study Institute (ASI) on Face Recognition was held in Stirling, UK, in the summer of 1997. Face Recognition (FR), a complex and difficult problem, is important for surveillance and security, telecommunications and digital libraries, human-computer intelligent interactions, and medicine. FR solutions presented have been shown to require synergetic efforts from areas such as signal and image processing, pattern recognition, machine learning, neural networks and evolutionary computation, psychophysics of human perception and neurosciences, and systems engineering. The ASI brought together leading researchers from academia, governments, and industry from around the world to present an all-encompassing view on FR, and identify trends for future developments and the means for implementing robust FR systems.

Most of the FR methods presented (NATO ASI SERIES F163) implement some variation on either the Principal Component Analysis (PCA) approach, also known as the eigenfaces approach, or the Dynamic Link Architectures (DLA), where elastic graph matching is attempted between locally derived forensic landmark grids, possibly encoded using Gabor wavelets. The PCA method records 2nd order statistics across the face and can be enhanced using spatiotemporal constraints encoded as manifolds ('trajectories') corresponding to the views obtained as the face rotates in 3D space. Further enhancements on the basic FR methods consider (i) active vision and (ii) modular forensic systems consisting of similar or different modalities, as would be the case when human authentication or surveillance is done by fusing visual and audio information.

The meeting considered both cognitive sciences and neurosciences with the aim of determining which is most important for face recognition. For example, color information should help with face detection, while sequential classifiers should perform better than flat classifiers on face recognition. Another concept discussed was that of using consensus networks or multiple decision trees for handling the inherent variability of the image formation process and the uncertainty involved in modeling the overall FR system. The discussion held also emphasized the ever-increasing role video processing will play in face recognition as additional frames of facial imagery and motion information contribute to increased confidence in face recognition.

An important topic regarding FR is the development of appropriate means for performance evaluation. Towards that end it has become clear that one has to develop standard data bases of faces images, such as FERET, in order to assess and compare between competing FR systems. Decision theory and ROC curves provide the tools needed to quantify the level of performance displayed by specific FR systems. Statistical learning theory further provides the means to estimate the guaranteed risk, when testing on future and unseen facial imagery, in terms of the empirical risk experienced during training and the complexity of the classification model underlying the FR system.
Reference books: E22, F30, F44, F45, F136, F163

[ PCO Home Page ]