Abstract: This paper proposes a robust image registration algorithm which includes the following two stages: (1) initial matching. SIFT matching method and the normalized cross correlation (NCC) metric modified with adaptive scale and orientation of SIFT features are proposed to find good initial matches, and the geometric consistency check is used to identify false matches; (2) matching propagation. A robust matching propagation using adaptive NCC and local homography constraint starts from the initial correspondences established in the first phase, and the geometrical consistency check is used simultaneously to eliminate the incorrect matches. By using matching propagation, we can get control points used to image registration as many as possible. Initial local homography is estimated using least squares matching algorithm and the initial values of unknown parameters needed for it is provided by adaptive NCC method. Compared to existing point-based image registration methods, the proposed algorithm has better performance in terms of registration accuracy and robustness to geometric deformations within images.