A Hierarchical Approach for Face Recognition

In this paper we present a new approach for face recognition based on multi-level Principal Component Analysis (PCA). A hierarchical representation based on multiple eigenspace nesting is proposed for representing faces. The faces are organized into a hierarchical structure, where each node is represented by an eigenspace related to a set of faces from distinct individuals. The leaf-level contains nodes related to single individuals, whose eigenspaces are calculated by several images to consider changes in illumination, expressions, translations and rotations. Recognition is performed by navigating the tree, following the most suitable path to represent the searched image, until a leaf is reached: distance-from-space has been used to establish the best path. A threshold on the distance to each node is adopted to reject unauthorized subjects. The first experiments carried out on the ORL database of faces, which is one of the most common benchmarks in this area, show that the new method achieves performance comparable with other eigenface-based recognition approaches reported in the literature, allowing for an efficient reduction of the number of comparisons needed for the recognition task. Moreover, the proposed hierarchical representation, unlike most of the existing methods, can be efficiently updated as new samples become available.