A fast image registration algorithm for super-resolution

Abstract

In this paper, a robust and fast image registration algorithm suitable for super-resolution is proposed, it yields a solution that precisely registers images with subpixel accuracy. The proposed registration process is carried out in three stages. In first stage, the image edges are extracted and then the corner points which are on the image edges are extracted, where the improved Harris corner algorithm is used in order to reduce the calculation amount. In second stage, for the coarse registration, the NCC (Normalized cross correlation) similarity measure is used to get an initial set of corresponding point pairs, and then a statistical method is employed in order to remove mismatched points. In detail, we calculate the shift (Δ x ,Δ y ) of each point pairs, count the frequency of each shift, and then select the shift with most frequency as the image shift, which is expressed as(Δ x 0 ,Δ y 0 ). In third stage, for fine image registration, subpixel image registration is achieved by interpolation. The bicubic interpolation is done in the neighborhood of the inliers (the correspondences with shift (Δ x 0 ,Δ y 0 ) are called as inliers) and the NCC matching and statistical method is used once again to find the correct corresponding point pairs, from which the shifts between the reference and unregistered image are estimated. The experimental results illustrate the registration speed and accuracy of the proposed method improved significantly.