Feature Reduction fo Face Recognition by PDBNN

Abstract

Selecting proper features for efficient face recognition is an essential task in neural network design. Most of recognition or classificaiton algotithms using features in uniform manner for each class. We believe this constraint could be relaxed to achieve betted recognition performance. In this paper, we present a PDBNN based feature reduction algorithm that deletes some feature vectors which contribute the least among the whole feature set. The delection is performed on each facial basis. By applying the proposed algorithm, we performed some face recognition experiments on an 151 people facial database. The experimental results show the recognition accuracy improved from th eoriginal 87.91% (2500 features) to 91.22% by using only 500 features.