An analysis of appearance-based statistical methods and autoassociative neural networks on face recognition

Abstract

An experimental Study on the face recognition problem was performed using Autoassociative Neural Networks (AANN) and appearance-based statistical methods namely, Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), and Independent Component Analysis (ICA). Various experiments were conducted using FERET database and FERET Evaluation Methodology to evaluate the performance of the four methods on frontal images under different conditions such as rotations, illumination changes and scale reductions. The results show that PCA outperforms LDA, ICA and AANN in general and under different illumination conditions, and LDA follows PCA according to the performajice results. On the other hand, PCA and LDA show a great sensitivity to rotations while ICA and A ANN are not sensitive to rotations. The results also show that all the approaches are sensitive to scale reductions. Compatibility- of experimental evaluations with the theoretically expected results is also demonstrated.