Recognizing facial expressions using subspace linear discriminant analysis

Abstract

In this paper, performance analysis of subspace Linear Discriminant Analysis (subspace LDA) method is performed for the solution of facial expression recognition. Subspace LDA is used as feature extractor with the combination of the preprocessing techniques of histogram equalization and mean-and-variance normalization in order to nullify the effect of illumination changes on facial images. The recognition performance of the Principal Component Analysis (PCA) and subpattern-based PCA are compared with the performance of subspace LDA approach to demonstrate the performance differences and similarities between these three types of approaches. In order to compare these methods, three facial expression databases such as John Kanade, FGnet and JAFFE have been used in our work. Person-dependent experiments are performed on these databases separately to represent the facial expression recognition performances of the aforementioned approaches.