Disadvantages of the several typical variants on Linear Discriminant Analysis (LDA) while dealing with the well-known small sample size problem in face recognition are revealed. A new discriminant analysis method named two subspaces LDA is proposed to deal with small sample size problem. Firstly, all the samples are projected to the nonzero space of the total scatter matrix. Secondly, discriminant analysis is carried out in the zero subspace and nonzero subspace of the within-class scatter matrix respectively. Thirdly, the two kinds of discriminant features are fused to determine class of the samples. Experimental results show that the proposed algorithm is superior to other linear methods.