Abstract: Independent component analysis (ICA) is a generalization of principle component analysis (PCA), which encounters some problems in face recognition.These problems are analyzed and three approaches proposed to improve the performance of ICA-based face recognition.First, a kernel based ICA method with KPCA is presented, which maps samples onto the feature space and obtains nonlinear independent components.Second, Fisher discriminant information is defined to select optimal discriminant independent components.Finally, a method is proposed to obtain the optimal feature representation using optimal discriminant independent components for unknown face samples.Experiments using ORL database indicate that the improved method outperforms FLDA, with recognition stability better than FLDA in case the number of features is small.

Key words: independent component analysis, face recognition, kernel principle component analysis, optimal discriminant analysis