Efficient MRF deformation model for non-rigid image matching

摘要: We propose a novel MRF-based model for deformable image matching (also known as registration). The deformation is described by field of discrete variables, representing displacements (blocks of) pixels. Discontinuities in the are prohibited imposing hard pairwise constraints model. Exact maximum posteriori inference intractable and we apply linear programming relaxation technique. show that, when reformulated form two coupled fields x- y-displacements, problem leads to simpler which sequential tree-reweighted message passing (TRW-S) algorithm [Wainwright-03, Kolmogorov-05]. This enables registration with large at single scale. employ fast updates special type interaction was proposed [Felzenszwalb Huttenlocher-04] max-product belief propagation (BP) introduce few independent speedups. In contrast BP, TRW-S allows us compute per-instance approximation ratios thus evaluate quality optimization. performance our technique demonstrated on both synthetic real-world experiments.