Combining Invariant and Corner-Like Features to Optimize Image Matching

Significance and usefulness of local invariant features and traditional corner-like features have been widely proven in the literature. In this paper, we novelly combine the two types of features to select salient keypoints with the invariant and corner-like properties, which are highly distinctive and improving match performance. We use moment-derived complex image patterns (e.g., corner, T-junction, sectional cut, and chess-cross) to find corner-like features. We further optimize the matching results by finding corner-like patterns in the invariant matched point correspondences; and rebuff point correspondences that have dissimilar pattern responses which are most likely false matches.