Face Recognition Using A Radial Basis Function Classifier

Face recognition is probably the most natural way to perform a biometric authentication between human beings. However, the available technology for automatic systems still presents some drawbacks and is far away from human performance. In this paper we use the same DCT feature extraction approach presented in previous ICCST'03 and ICCST'05. However, we improve the experimental results using a Radial Basis Function (RBF) neural network in combination with the coding of the recognized class. We explain why the RBF, do not have the limitations of other classifiers such as the MLP. We also propose a method for dealing with the high number of classes associated to the task of face recognition which takes into account the limitations of the RBF as classifiers, and discuss the weakness of these methods when the number of training samples is limited. We have performed an exhaustive study about the neural network architecture and parameters, which has let us to establish relevant conclusions about the optimal configuration.