We propose a Markov Random Field formulation for the linear image registration problem. Transformation parameters are represented by nodes in a fully connected graph where the edges model pairwise dependencies. Parameter estimation is then solved through iterative discrete labeling and discrete optimization while a label space refinement strategy is employed to achieve sub-millimeter accuracy. Our framework can encode any similarity measure, allows for automatic reduction of the degrees of freedom by simple changes on the MRF topology, and is robust to initialization. Promising results on real data and random studies demonstrate the potential of our approach.     «

We propose a Markov Random Field formulation for the linear image registration problem. Transformation parameters are represented by nodes in a fully connected graph where the edges model pairwise dependencies. Parameter estimation is then solved through iterative discrete labeling and discrete optimization while a label space refinement strategy is employed to achieve sub-millimeter accuracy. Our framework can encode any similarity measure, allows for automatic reduction of the degrees of freed...     »