Image registration using BP-SIFT

Scale Invariant Feature Transform (SIFT) is a powerful technique for image registration. Although SIFT descriptors accurately extract invariant image characteristics around keypoints, the commonly used matching approaches of registration loosely represent the geometric information among descriptors. In this paper, we propose an image registration algorithm named BP-SIFT, where we formulate keypoint matching of SIFT descriptors as a global optimization problem and provide a suboptimum solution using belief propagation (BP). Experimental results show significant improvement over conventional SIFT-based matching with reasonable computation complexity.

► We apply geometric information as constraints in image registration by BP with SIFT. ► BP corrects mismatched keypoints according to descriptors and geometric information. ► BP-SIFT finds the consistent geometric correspondence between two sets of features. ► Results show significant improvement of our method over conventional SIFT. ► Mismatched keypoints, which are just removed by RANSAC, correctly matched by BP-SIFT.