Image Matching Algorithm Based On SIFT Feature Points

Image matching is such a hot and difficult problem in computer vision and imageinformation fusion. It is mainly used to match two or more images, which is acquired atdifferent times, from different sensors or from different viewpoints. It is a fundamentalproblem in image processing and computer vision, and also a challenging topic inMulti-sensor information fusion. Research on this subject has very important theoreticaland practical significance. At present, the image matching technology has been widelyused in target recognition, moving target tracking,3D scene modeling, etc.This paper mainly studies the image matching algorithm based on SIFT (ScaleInvariant Feature Transformation) feature. The features are invariant to image scaling androtation, and partially invariant to change in illumination and3D camera viewpoint. ButSIFT is sort of complicated and time consuming, because it searches for stable featuresacross all possible scales and matches features in high dimensional space. In order toimprove the matching speed of SIFT feature, this paper describes an approximate nearestneighbor search algorithm based on vector angle. First, compute vector angle betweenhigh dimensional vectors and a stochastic selected reference vector, and sort these angles.Then compute the angle of reference vector and the query vector, and find this angle in thesorted angles by binary search algorithm. Finally, take this angle as the center, searchingthe approximate nearest neighbor of the query vector in the setting range.Many contrast experiments have been done to verify the efficiency and superiority ofour method. The experimental results show that the SIFT feature matching can beaccelerated significantly without undermining the performance of feature matching. Inaddition, we also did an experiment that is image matching with the SIFT feature. Theexperimental results show that image matching algorithm based on SIFT is partially staleand robust to image rotation, change in illumination and3D camera viewpoint.