Abstract:

 Linear discriminant analysis(LDA) often encounters a small-sample-size(SSS) problem when used in face recognition because there are not enough  trainingsamples and the within scatter matrix is singular. The traditional LDAbased null space algorithm can effectively avoid the SSS problem by first removing the null space of totalscatter matrix. A new feature extraction method called weighted null space was proposed, and the selection of weight parameters was discussed. Experimental results on face databases demonstrated its validity.

Key words: face recognition; null space; linear discriminant analysis, LDA