Conditional variance of differences: A robust similarity measure for matching and registration

Abstract

This paper presents a new similarity measure, the sum of conditional variance of differences (SCVD), designed to be insensitive to highly non-linear intensity transformations such as the ones occurring in multi-modal image registration and tracking. It improves on another recently introduced statistical measure, the sum of conditional variances (SCV), which has been reported to outperform comparable information theoretic similarity measures such as mutual information (MI) and cross-cumulative residual entropy (CCRE). We also propose two additional extensions that further increase the robustness of SCV(D) by relaxing the quantisation process and making it symmetric. We demonstrate the benefits of SCVD and improvements on image matching and registration through experiments. © 2012 Springer-Verlag Berlin Heidelberg.