Face Recognition

The problem of computerised personal identity verification has received considerable attention over the past decades and a large number of approaches have been proposed in the literature. Though the early strategy for face identification was based on geometrical features such as nose width and length, mouth position and chin shape, etc. Nearly all of the applicable approaches developed in recent years are based on holistic representations known as templates.

The most commonly used statistical representation for face recognition and verification is the Karhunen-Loeve (KL) expansion known also as the Principal Component Analysis (PCA). Its application to the face recognition problem has been pioneered by Sirovich and Kirby but the approach has been popularised by Turk and Pentland where the PCA bases are referred to as eigenfaces. Since then the eigenface method has been widely used by many researchers.

Although PCA is very effective for information compression, it does not guarantee most efficient capture of discriminatory information. More recently, linear discriminant analysis (LDA) has been adapted to face recognition by Belhumeur. The LDA representation referred to as "fisherfaces" have been demonstrated to outperform the PCA representation in most applications.

A further improvement that achieves superior performance to any other known method is Client-Specific fisher face representation or CS-LDA. Equally importantly, the superior performance is not the only benefit:

1. Simplicity of training: which for large user databases requires only a matrix multiplication of the client mean vector
2. Insulation of a client enrolment: from the enrolment of other clients. This enables the possibility to implement other than a centralised architecture for a personal identity or authentication system - i.e. architecture where client models are stored and updated centrally.
3. Smart card processing: becomes a reality without the need to restrict the representation framework and therefore the representational capacity of the system.
4. Speed of probe testing: being more than two orders of magnitude faster than conventional PCA and LDA methods because CS-LDA involves only a single fisher face per client.