Selection of optimal dimensionality reduction methods for face recognition using genetic algorithms

Abstract

A new approach for optimal selection of dimensionality reduction methods for individual classifiers within a multiple classifier system is introduced for the face recognition problem. Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA) and Independent Component Analysis (ICA) are used as the appearance-based statistical methods for dimensionality reduction. A face is partitioned into five segments and each segment is processed by a particular dimensionality reduction method. This results in a low-complexity divide-and-conquer approach, implemented as a multiple-classifier system where distance-based individual classifiers are built using appearance-based statistical methods. The decisions of individual classifiers are unified by an appropriate combination method. Genetic Algorithms (GAs) are used to select the optimal dimensionality reduction method for each individual classifier. Experiments are conducted to show that the proposed approach outperforms the holistic methods.