Split-and-merge EM for vine image segmentation

With the goal of recovering the 2D tree structure present on grape vine binary images, in this paper we propose to use Mixture of Gaussians for canes segmentation. The main idea behind our approach is to use information criteria from model selection theory to guide directly the split-and-merge steps for learning Mixture of Gaussians via Expectation Maximization. A novel information criteria we found experimentally is able to adapt to canes at different image scales. We show results of cane segmentation using our criteria in comparison to standard ones as Akaike and Bayesian information criteria. Finally we provide directions on how this work could be extended in the future.