Multiscale MRF optimization for robust registration of 2D biological data

摘要

Discrete formulations of image registration offer the promise of dense deformations via optimizations robust to large motions or poor initialization. However, many available efficient algorithms are not well suited to medical or biological data. We propose a novel multiscale Markov Random Field formulation for image registration, which reduces the number of labels needed at each scale while preserving the ability to represent dense, fine-grained feature matches. The multiscale nature of the algorithm also allows arbitrary sub-voxel accuracy, and we further propose a simple extension which grants a measure of rotational invariance to an arbitrary feature matching term.