Abstract:

In order to better realize image tracking, a feature describing scaleinvariant feature transform (SIFT) algorithm based on the kernel locality preserving projections (LPP) is proposed to search a more robust and convenient feature descriptor. In the algorithm, SIFT properties are well inherited, and the numbers of sub image plies are refined to increase the amount of stable feature points according to the difference of energy characteristics in every scale space. Moreover, a nonlinear dimension falling method based on the kernel LPP is formed to redraw feature descriptors by the kernel technique. With this function, the improved LPP can extract high dimensional features. The result shows that the novel algorithm can obtain more and better feature points, and its calculating speed quite faster than the principal component analysisSIFT in various scenarios like zoom, rotation, blur and illumination variation of images.