Using Signal/Residual Information of Eigenfaces for PCA Face Space Dimensionality Characteristics

Principal component analysis has been used since 1990 in many recognition algorithms to get a face feature representation and to exploit the dimensionality reduction characteristic of the principal component analysis (PCA). The way to determine the optimal dimension of the reduced space is still not available. Another critical point when working with PCA is the influence of the training set, denoted here as PCA construction set. In this paper we are working on the behaviour of the signal/residual information of the PCA-eigenspectrum in order to determine an optimal threshold that could be used for the dimensionality reduction. We also study the influence of different sets used to construct the PCA representation. Our experiments are done on the FRGCV21 database, using the BEE PCA baseline software. We also use images from the BANCA database for the construction of the PCA representations