Abstract: Based on the equivalence between canonical correlation analysis (CCA) and Fisher linear discriminant analysis (FLDA), nonlinear discriminant features of face images are extracted with kernel CCA.These features are equivalent to those extracted with KFDA.Experimental results demonstrate that KCCA is similar to GDA and significantly better than FLDA.

Key words: canonical correlation analysis, Fisher discriminant analysis, kernel method, small sample size problem, face image discrimination