An Improvement Of Matching Criteria And Method Of SIFT Feature And Its Application

Feature matching is a significant area in computer vision, meanwhile, SIFT feature is one of the most popular feature nowadays. The purposes of this paper are as follow:First, modifying the original SIFT feature matching criteria; Second, approaching a new Gabor-SIFT feature matching method; Third, applying the above results to build a 3D object model so as to deal with the problem that is when the perspective changes a lot, SIFT feature does not perform well.First of all, by analyzing the original feature matching method, this paper discovers that some corresponding feature points have been detected in both images while these points are not considered as matched in the original method. So, this paper presents a new feature matching criteria, which increases matching feature points by 20%.Moreover, in SIFT-based visual search, the high dimension of SIFT feature will consume a large amount of time under the original matching method. Since the dimension and distinctiveness of SIFT are in conflict, how can we maintain a high distinctiveness and a fast speed at the same time? Since Gabor feature can represent low level image information and the dimension can be low if combining method is used. Thus, this paper approaches a new Gabor-SIFT feature matching method:First this paper uses low dimensional Gabor features to screen most unmatched key points at the beginning. For the remaining key points, this paper uses high dimensional SIFT features to have a second match. The experiment results show that the new Gabor-SIFT matching method is about one time faster than the original method, meanwhile, the matching results almost remain the same.Finally, since SIFT feature belongs to local image features, its matching performance is poor when the angle of view changes more than 30 degree. Based on the above two results, this paper build a 3D object model by several images under different views. This model can achieve a good matching performance under any view of the object, while only require four times more of the SIFT feature.