FDLDA: An Fast Direct LDA algorithm For Face Recognition

Feature extraction is one of the hot topics in face recognition. However, many face extraction methods will suffer from the "small sample size" problem, such as Linear Discriminant Analysis (LDA). Direct Linear Discriminant Analysis (DLDA) is an effective method to address this problem. But conventional DLDA algorithm is often computationally expensive and not scalable. In this paper, DLDA is analyzed from a new viewpoint via SVD and an fast and robust method named FDLDA algorithm is proposed. The proposed algorithm achieves high efficiency by introducing the SVD on a small-size matrix, while keeping competitive classification accuracy. Experimental results on ORL face database demonstrate the effectiveness of the proposed method.