Due its possibilities in security systems and robotics, face recognition is one of the most researched areas within the biometric field. In a common scenario from real life face recognition problem, the dimension in the sample space is larger than the number of training samples per class. This is known as the “small sample size problem”. Discriminative Common Vectors (DCV) technique has been used to face this problem successfully. In this paper, we introduce a new approach based on DCV theory to increase its performance in face verification tasks. This modification uses a specific set of projecting vectors selected by an optimization algorithm based on the classifier's performance, and in the fact that no such thing as common vectors exists when this set contains vectors from the range of the within-class scattering matrix (SW ). Based on these two ideas, we may call this approach Discriminative Multi-Projection Vectors (DMPV) as it projects samples in both range and null space of SW. We tested the system with different databases and results show that DMPV outperforms classic DCV method.