Small Sample Size Face Recognition using Random Quad-Tree based Ensemble Algorithm

Abstract

Certain applications such as person re-identification in camera network, surveillance photo verification, forensic identification etc. suffer from a small sample size (SSS) problem severely. Conventional face recognition methods face a great challenge on SSS as the trained feature space is overfitted to the small training set. Interest in combination of multiple base classifiers to solve the SSS problem has sparked renewed research efforts in ensemble methods. In this paper, we propose a novel Random Quad-Tree based ensemble algorithm (R-QT) to address the SSS problem. In contrast to other methods confining the ideas on limited data, R-QT enlarges the training data to obtain more diverse base classifiers. Moreover, R-QT encodes not only discriminant features but also the geometric information across the face region, which further improves the recognition accuracy. Results on five standard face databases demonstrate the effectiveness of the proposed method.