Introducing diversity to Normalized Cross Correlation for dense image registration

Normalized Cross Correlation (NCC) has been extensively used for image registration, but applying NCC alone does not result in sufficient accuracy for many scenarios. In this paper, we propose a simple yet accurate dense image registration method by introducing diversity to “candidates” of NCC matches. We then select the best match using Belief Propagation (BP) to incorporate non-local geometric information into the calculation. We compared our proposed method with a control method when diversity is not incorporated and a state-of-the-art image registration method, SCoBeP.