Abstract: This paper presented a method of two-dimensional locality preserving projection based on Ll-norm(2DLPP-Ll). I}he proposed approach has two advantages compared with the conventional I_2-norm based two-dimensional locality preserving projection(2DLPP). Firstly, it is more robust against outliers because Ll-norm is insensitive to noises. Moreover, it does not require the eigenvalue decomposition. Experiments on two face databases and one hand-written digit dataset illustrate that compared with 2DI_PP,the proposed method exhibits better performance when there arc outliers in training sets.

Key words: Feature extraction, I_l-norm, Locality preserving proj ection, Facc recognition