Elsevier

Neurocomputing

Volume 50, January 2003, Pages 485-488
Neurocomputing

Letters
Face recognition based on linear classifiers combination

https://doi.org/10.1016/S0925-2312(02)00674-4Get rights and content

Abstract

In this paper, an approach that uses a combination of linear classifiers is applied to face recognition. We propose a novel criterion for the combination, the maximum complementariness criterion, which is used to construct the fitness function for a genetic algorithm (GA). A GA is then used to generate the rational weights for the classifiers. Experiments show that our approach can successfully improve the performance of the classification for face recognition using the commonly used Fisherface features.

Introduction

In face recognition, many techniques have been proposed. Among them, the most notable technique involves extracting a Fisherface feature to perform the classification [1], which is a successful technique to a certain degree. To further improve its performance, in this paper, we propose an approach that combines linear classifiers. Recently, the idea of combining several techniques has received considerable attention and it has been used in different applications. Kittler et al. [3] present an elementary theoretical framework, which can be used to generate some simple combination rules when adopting different assumptions and different styles of expression, and a study on different strategies for combining the classifiers was presented by Kuncheva [4]. Nevertheless, research is still needed to solve the problem of how to create an appropriate criterion for combining the classifiers. A novel criterion for the combined classifiers, called the maximum complementariness criterion (MCC), is proposed in our paper to try to solve this problem.

Section snippets

Description of the approach used to combine the classifiers

First, we need to construct classifiers that are uncorrelated as much as possible. Obviously, this is because there is less redundant information among them and this provides a more reliable result when we use the combined classifiers. The orthogonal wavelet transform is used as a preprocessing tool for the images because of its ability to eliminate correlation and its unique capability to decompose an image. We select a conventional Daubechies’ orthogonal wavelet base to apply a discrete

Experimental results and conclusions

In our experiments, we used one of the NUST603 facial-image databases [2] from the Nanjing University of Science and Technology, China. It contains images of 18 people, where each person has 12 images that are of size 64×64 pixels. Fig. 1 shows a typical example of the images for one person. Note that there are significant changes in both the facial expression and the pose. We randomly choose 3 samples from each class to construct the classifiers, and 3 other samples were used to search for the

References (4)

There are more references available in the full text version of this article.

Cited by (12)

View all citing articles on Scopus
View full text