In the field of object recognition, the SIFT feature is known to be a very successful local invariant descriptor and has wide application in different domains. However it also has some limitations, for example, in the case of facial illumination variation or under large tilt angle, the identification rate of the SIFT algorithm drops quickly. In order to reduce the probability of mismatching pairs, and improve the matching efficiency of SIFT algorithm, this paper proposes a novel feature matching algorithm. The basic idea is taking the successful-matched SIFT feature points as the training samples to establish a space mapping model based on BP neural network. Then, with the help of this model, the estimated coordinate of the corresponding SIFT feature point in the candidate image is predicted. Finally search the possible matching points around the coordinate. The experiment results show that using the prediction model, the number of mismatching points can be reduced effectively and the number of correct matching pairs increases at the same time