The traditional heuristic feature selection methods usually neglect the correlations between features, and thus lead to suboptimal feature subset. Therefore, a method of manifold discriminant feature selection(MDFS) is proposed. The method captures the manifold structure of the dataset by incorporating both neighbor and label information, and then the objective function can be formulated by minimizing the difference between intra and inter scatters. Besides, the structured sparse regularization term is further added to reduce the redundant information. Finally, a new iterative algorithm is presented for optimization. The experimental results on three popular datasets, i.e., ORL, COIL20, and Isolet1 dataset, show that, compared with existing related methods, the proposed method achieves better clustering performances in terms of accuracy and normalized mutual information. Thus the effectiveness of the proposed method can be verified.