Feature extraction is one of the hot topics in the field of pattern recognition. In this paper, a new feature extraction method called Fisher discriminant minimal criterion is proposed to improve the performance of feature extraction. Conventional Fisher discriminant criterion is inversed and null space of between-class scatter matrix is defined in this algorithm. Therefore, limitation of final eigenvectors' dimensions determined by class number is overcome and more effective classification information can be achieved. Experimental results on face databases demonstrate the effectiveness of the proposed algorithm.