LAVIR-Locally adaptive variational image registration

Variational methods for image registration involve minimizing a nonconvex functional with respect to the unknown displacement between two given images. In this paper, we present a new non-parametric image registration method posed as an optimization procedure, which combines a matching criterium and a smoothing term in an appropriate manner. Through the use of a weight function the model produces a smooth mapping between the two images, S and T, pointwise guided by the gradient variation of the target image S, the balance between the smoothing on the displacement vector field and the matching criterium allows for the occurrence of large deformations. We also present a framework for the fast linearized alternating direction method of multipliers (ADMM) for the numerical solution of the proposed model. The basic idea of the proposed algorithm is to incorporate an acceleration scheme into linearized ADMM. Experiments with both synthetic and real images in different domains illustrate that the proposed approach is efficient and effective.