Conventional SIFT (Scale Invariant Feature Transform) has some deficiencies when applied to the problem of face recognition. (i) SIFT provides a limited number of keypoints because it only seeks blob-like structures. Since most faces look generally similar to each other, a limited number of keypoints may not be sufficient to discriminate between them. (ii) Each keypoint extracted from the test image has to match against all keypoints extracted from the database images, thus it is very time consuming. (iii) Since different keypoints may have similar local features (descriptors) so that false matches still frequently arise. In this paper we combine SIFT and FAST (Features from Accelerated Segment Test) keypoint detectors to increase discriminating power between face images. The SIFT descriptors of both types of keypoints are collected where the Laplacian of Gaussian operator (LoG) is used to find the characteristic scale of each FAST keypoint. Moreover, we introduce local constraint on location, scale, or orientation to reduce false matches during keypoint matching. Computer simulation on ORL and YALE database shows that the proposed method not only reduces the number of false matches but also reduces the computational costs. Moreover, from our experiments, recognition rate is obviously improved.