Real-time methods for face recognition.

Identification of individuals on the basis of facial features is the most natural method of distinguishing one individual from another. Automating such a process, based upon quantifiable measures, is of great interest in a variety of applications, such as passport identification and automatic teller machine verification. The most crucial aspect of such applications is their tolerance with respect to variations in facial expressions and the noise introduced by the operating environment. In this thesis, various face recognition methods are evaluated under conditions of real-time response, varying operating factors, and implementation feasibility. The approaches are based on histogram mapping, wavelet transform, Karhunen and Loeve transform, and optical correlation techniques. A brief review of the basic concepts in optics is first presented. This is followed by a detailed review of optical methods in pattern recognition. A comprehensive background of algorithmic approaches for face recognition is described. A detailed analysis of the photobook system, which is based on the Karhunen and Loeve transform (KLT), is presented. It is argued that, even though the KLT possesses many useful attributes in image processing applications, the performance of KLT face recognition systems is based entirely upon the initial training set. A method for choosing the proper training set is presented. Novel statistical methods that exploit the stationary behaviour of the operating environment are introduced. It is shown that under the condition that control may be exercised on the operating environment, these methods provide a satisfactory result in real-time. The application of histogram, moment, and 2-D discrete wavelet transforms in statistical methods is described. A novel optical correlation based system is presented. It is shown that such a system tolerates changes in facial expressions and can operate under real time constraints