Facial expression recognition based orthogonal supervised spectral discriminant analysis

In recent years, feature extraction methods make an achievement in pattern recognition and computer vision. It extracts not only useful feature for classification, but also reduces the dimension of pattern samples. In this paper, we propose orthogonal supervised spectral discriminant analysis (OSSDA) which motivated by marginal fisher analysis (MFA) and spectral clustering. It put different weights for each sample by the density of itself. OSSDA redefines the between-class scatter matrix and within-class scatter matrix so as to enhance the compactness of inter-class and maximizes the distance between marginal points. Experiment on JAFFE database and Cohn-Kanade database show our method can get better performance than LDA, MFA.