Inference bay' esienne pour l'estimation de d eformations larges par champs gaussien: application au recalage d'images multi-modales

Image registration aims to estimate the global deformation between a target image I1 and a reference image I2. In this context, we will focus on estimating a random field U on the I₁ domain ω = [0, 1]² based on observations of U on a finite set of curves β ϵ ω. Indeed, we present a new multimodal image registration method based on Gaussian random fields. The proposed method first find the optimal correspondences between curves β_s then estimate the deformation vector field on ω. The optimal solution is computed using Maximum Likelihood and Bayesian inference. Based on results using both real and simulated data, the resulting deformation has the advantage of being exact on the observations as being sufficiently smooth over the whole ω.