SIFT features for face recognition

Scale invariant feature transform (SIFT) has shown to be very powerful for general object detection/recognition. And recently, it has been applied in face recognition. However, the original SIFT algorithm may not be optimal for analyzing face images. In this paper, we analyze the performance of SIFT and study its deficiencies when applied to face recognition. We propose two new approaches: Keypoints-Preserving-SIFT (KPSIFT) which keeps all the initial keypoints as features and Partial-Descriptor-SIFT (PDSIFT) where keypoints detected at large scale and near face boundaries are described by a partial descriptor. Furthermore, we compare the performances of holistic approaches: Fisherface (FLDA), the null space approach (NLDA) and Eigenfeature Regularization and Extraction (ERE) with feature based approaches: SIFT, KPSIFT and PDSIFT. Experimental results on ORL and AR databases show that our proposed approaches KPSIFT and PDSIFT can achieve better performance than the original SIFT. Moreover, the performance of PDSIFT is significantly better than FLDA and NLDA. And PDSIFT can achieve the same or better performance than the most successful holistic approach ERE.