Pre-clustered principal component analysis for fast training of new face databases

Generally the eigenface approach to obtain facial representation for face recognition requires the large computational load originated from solving eigenvalue problem to obtain principal axes. When the new face database is large, this computational load significantly increases. To resolve this difficulty, we propose a novel feature extraction scheme using pre-clustered face data; before a data set for specific problem (on-line data) is given, a number of general face images (off-line data) are clustered by a modified k-means clustering algorithm and a set of principal axes are obtained for each cluster by applying the principal component analysis to the data belonging to each cluster; when a data set for specific problem is given, an appropriate cluster is assigned to this data set by distance measure and features are extracted by using the principal axes of the selected cluster. With this scheme we can reduce the large portion of computational load required to find the principal axes for each on-line data set, with a slight degradation of classification performance. This performance degradation is compensated by using the class-augmented principal component analysis as an on-line feature extractor and nearest neighborhood classifier as a classifier. In experiments, we use Yale face database and ORL face database to compare results of recognition performance and computational efficiency with those of other previous results.