SIFT (scale invariant feature transform) proposed by Lowe has been widely and successfully applied in object detection and recognition. SIFT features are invariant to image scale and rotationand are shown to provide robust matching across a substantial range of affine distortion, change in 3D viewpoint, addition of noiseand illumination variation. But the dimension of SIFT feature vector is highand the matching of feature points needs long time. This study presents an improved face recognition algorithm based on SIFT, through reducing the dimension of feature vector by modified local descriptorand novel feature matching scheme. In fact, considering the physical meaning of different regions on human face, such as eyes, nose or mouth etc., feature matching could be performed in corresponding areas, not in global scale. Compared with well-established face recognition algorithms, namely Eigenfaces and Fisherfaces, experimental results demonstrate that improved SIFT descriptors applied in face recognition present higher success rate of matching in some degree and the matching speed is persuadably faster.
展开▼