8.3. Facial Pattern Classification Techniques

As discussed in Section 2.2.6, the pattern classifier and feature extractor are often a complementary pair in the biometric identification paradigm. A powerful classifier needs to be adopted to separate the features from a simple extractor. Neural network techniques are widely applied and suitable for face recognition algorithms. Instead of recognizing a face by following a set of human-designed rules, neural networks learn the underlying rules from a given collection of representative examples. This ability to automatically learn from examples makes neural network approaches attractive and exciting. Moreover, it is well known that neural networks are very robust and adaptive. Therefore, for applications with many variation factors, neural networks seem to be a good remedy.

Constructing a neural model is very much dependent on the intended application and is crucial for successful recognition. For example, the model for gender classification [115] is different than that used for facial expression classification [293]. For face detection, multi-layer perceptrons [343] and convolutional neural networks [323] have been applied. For face verification, the dynamic link architecture [194] is applied, using Gabor wavelets as features. Another example is the Cresceptron [376], which is a multiresolution pyramid structure similar to Fukushima's Neocognitron. Support vector machines [132] have been applied to recognize faces that are rotated as much as 40 degrees.

This chapter focuses on probabilistic decision-based neural network (PDBNN) [209, 210] and its applications to face recognition. PDBNN enjoys the merits of both neural network and statistical approaches and it inherits the modular structure from its predecessor, the decision-based neural network (DBNN) [192]. For each person to be recognized, a PDBNN devotes one of its subnets to the representation of that particular person. Such a structure makes system maintenance easy. The updating of a PDBNN-based security system for any personnel change in an organization is relatively straightforward—just a simple addition or deletion of the subnets (using a localized training process). A centralized system, in contrast, would require global updating. PDBNN's modular structure also facilitates portable ID devices; it is simple to implement a PDBNN-based biometric ID on a wallet-size smart card system. The smart card records only the parameters of the subnet in the PDBNN that represent the cardholder. In terms of statistical merits, PDBNN adopts the form of probability density as its discriminant function, which yields far fewer false acceptance cases even if there are not many "negative examples" in the training phase (see Section 7.5.2). This characteristic is highly desirable in systems that are frequently under intruder attack. The remaining sections of this chapter will demonstrate PDBNN's ability to handle face detection, eye localization, and face recognition all at once.