Dense image registration through MRFs and efficient linear programming.

摘要: In this paper, we introduce a novel and efficient approach to dense image registration, which does not require derivative of the employed cost function. such context, registration problem is formulated using discrete Markov random field objective First, towards dimensionality reduction on variables assume that deformation can be expressed small number control points (registration grid) an interpolation strategy. Then, sum over costs (using arbitrary similarity measure) projected points, smoothness term penalizes local deviations according neighborhood system grid. Towards approach, search space quantized resulting in fully model. order account for large deformations produce results high resolution level, multi-scale incremental considered where optimal solution iteratively updated. This done through successive morphings source target image. Efficient linear programming primal dual principles recover lowest potential Very promising synthetic data with known real demonstrate potentials our approach.