Fast Rotated Image Matching based on Two-dimensional Orthogonal Expansion by Marginal Eigenvectors

Abstract

Advanced algorithms are proposed for fast image matching based on two-dimensional orthogonal expansion by marginal eigenvectors for a set of training pictures. The original method has been proved to be efficient by use of the vectors of smaller dimensions rather than the number of pixels. It is based on two eigenvalue problems for the average column and row covariance matrices made of the training pictures, and then the computation time for training was reduced in comparison with the methods based on the non-marginal eigenvectors. However, the computation time for recognition or registration remained to be improved. In this paper, the computation cost for registration can be reduced by use of two algorithms of fast computation: fast computation of expansion coefficients and efficient computation of distances in the coefficient space. A rotation-invariant registration method based on the proposed algorithms is designed. The effectiveness of the proposed algorithms can be shown through fundamental experiments with real images.