In this study, we proposed an improved LPP method named Scatter-Difference Discriminant Locality Preserving Projection (SDDLPP). It considers discriminant information by maximizing the scatter-difference, which makes it have better classification capability. SDDLPP also avoids the singularity problem for the highdimensional data matrix and can be directly applied to the small sample size problem while preserving more important information. Comparative recognition performance results on public face and palmprint databases also demonstrate the effectiveness of the proposed SDDLPP approach.