Novel Image Registration Method Based on Local Structure Constraints

Abstract

This letter presents an effective approach to reduce the ambiguity of matching results for image registration based on a coarse-to-fine strategy. In the coarse registration stage, we compute initial transformation parameters via the descriptors. In the fine registration stage, we propose a new matching strategy for an iterative closest point framework, in which the matching pairs are determined by a bidirectional matching criterion in terms of feature similarity and spatial consistency. In this letter, the spatial consistency includes not only spatial distance but also local structure constraints on reference and sensed images. Comparative experiments on multispectral and viewpoint-altered images show that the proposed algorithm achieves higher performance in accuracy and robustness.