On the performance of wavelet families in face recognition using a multilayer perceptron neural network classifier

In this paper, we investigate the effects of different wavelet families as well as the effects of number of neurons on a the performance of a neural network based face recognition system. The face images are transformed using multi-level wavelets from which features are extracted. The resulting feature vectors are project over an orthogonal space using a simple PCA (Principal Component Analysis) projection. The uncorrelated transformed feature vectors are then used with an Multilayer Perceptron (MLP) based classifier. Different scenarios in terms of wavelet families and network structures are investigated. Extensive experimental results were performed using the ORL database. We show that certain families together with certain MLP structures give the best results in terms of recognition accuracy.