An efficient fusion strategy called discriminant feature fusion strategy for supervised learning is proposed to seek the optimal fusion coefficients of feature fusion. Contributions of this paper lie in: 1) creating a constrained optimization problem based on maximum margin criterion for solving the optimal fusion coefficients, which causes that fused data has the largest class discriminant in the fused feature space; 2) keeping an unique solution of optimization problem by transforming the optimization problem to an eigenvalue problem, which causes the fusion strategy to reach a consistent performance. Besides of the detailed theory derivation, many experimental evaluations also are presented in this paper.