An exploration of multimodal similarity metrics for parametric image registration based on particle filtering

This paper presents an analysis of different multimodal similarity metrics for parametric image registration based on particle filtering. Our analysis includes four similarity metrics found in the literature and we propose a new metric based on the discretization of the kernel predictability, function recently introduced by Gómez-García et al. (2008), that we call histogram kernel predictability (HKP). Hence the metrics studied in this work are mutual information, normalized mutual information, kernel predictibility with gaussian and truncated parabola functions, and HKP. The evaluations include tests varying the number of particles in the filter, the type of pixel sampling, the number of bins used to calculate the histograms, the noise in the images, and the computation time. Furthermore, we also conducted a geometric analysis to inspect convexity properties of the metrics under discussion. The overall evaluation suggests that the normalized mutual information is the best similarity metric for parametric image registration.