Image transport regression using mixture of experts and discrete Markov Random Fields

Abstract

The registration of multi-modal images is the process of finding a transformation which maps one image to the other according to a given similarity metric. In this paper, we introduce a novel approach for metric learning, aiming to address highly non functional correspondences through the integration of statistical regression and multi-label classification. We developed a position-invariant method that models the variations of intensities through the use of linear combinations of kernels that are able to handle intensity shifts. Such transport functions are considered as the singleton potentials of a Markov Random Field (MRF) where pair-wise connections encode smoothness as well as prior knowledge through a local neighborhood system. We use recent advances in the field of discrete optimization towards recovering the lowest potential of the designed cost function. Promising results on real data demonstrate the potentials of our approach.