Séminaire du 04/02/2016, 14h00

jeudi 04 février 2016, 14h00

Binary partition trees for multimodal segmentation and mathematical morphology

Conférencier : Miguel Angel Veganzones (Gipsa-lab)

Résumé : Binary partition trees (BPT) have proven to be an efficient approach to encode a hierarchical representation of images. Image segmentations can be obtained from a BPT representation by pruning the hierarchical tree. Recently, has been proved that optimal pruning can be effectively achieved by minimizing an energy function over the space of partitions encoded by the BPT representation, using dynamic programming. In this talk, I will overview our latest research using the BPT energy minimization framework for (i) multimodal segmentation and (ii) mathematical morphology.

When it comes to multimodal segmentation, the fusion of multiple hierarchies remains an open question. Recently, the concept of braids of partitions has been proposed as a theoretical tool and possible solution to this issue, but it has never been investigated in practical scenarios. In the first part of the talk, I will expose our recent proposal for the analysis of multimodal images, based on this notion of braids of partitions. In particular, we develop a method to perform the hierarchical segmentation of such multimodal images, relying on an BPT energetic minimization framework.

Mathematical morphology is grounded on the notion of complete lattices which requires of an ordering of the data samples. For images, the ordering is generally defined based in the space of intensity values of the pixels. For instance, the conventional ordering of the Real field is employed for gray-scale images. In the second part of the talk, I will expose the concept of permutation orderings for images, where the ordering is defined using not only the space of intensity values of the pixels, but their spatial location on the images as well. An example of the proposed approach using BPTs is provided.