Random Fields for Image Registration

Опубликовано 17 августа 2016, 21:16

In this talk, I will present an approach for image registration based on discrete Markov Random Field optimization. While discrete optimization often provides strong solutions in purely discrete settings, the task of registration usually involves the estimation of continuous transformation parameters. We tackle this problem by introducing suitable approximation schemes and iterative refinement strategies in a principled manner. This allows us to apply discrete optimization both for linear and non-linear image registration. The performance of our implementations is demonstrated on several clinical applications including atlas generation, multi-modal brain registration, automatic segmentation via atlas-matching, and whole-body MRI stitching, as well as for the non-clinical problem of optical flow estimation.