Dual LDA for Face Recognition

The complete theory for Fisher and dual discriminant analysis is presented as the background of the novel algorithms. LDA is found as composition of projection onto the singular subspace for within-class normalised data with the projection onto the singular subspace for betweenclass normalised data. The dual LDA consists of those projections applied in reverse order. The experiments show that using suitable composition of dual LDA transformations gives as least as good results as recent state-of-the-art solutions.