(conférencier invité par Farida Cheriet dans le cadre des séminaires spéciaux du dépt. GIGL)

Titre: "Deformable Registration: Setting the State of the Art Using Primal Dual Strategies"

Conférencier: Prof. Nikos Paragios, École Centrale de Paris

Date: lundi 03 mars 2008
Heure: 14h00
Local: M-1010, Pavillon Lassonde, École Polytechnique

In this talk we present a novel, fast, efficient and gradient free approach to dense image registration. In such a context the registration problem is formulated using a discrete Markov Random Field objective function. First, towards dimensionality reduction on the variables we assume that the dense deformation field can be expressed using a small number of control points (registration grid) and an interpolation strategy. Then, the registration cost is expressed using a discrete sum over image costs (using an arbitrary dissimilarity measure) projected on the control points, and a smoothness term that penalize local deviations on the deformation field according to a neighborhood system on grid. Towards a fully discrete approach the search space is quantized resulting in a fully discrete model. In order to account for large deformations and produce a finer and finer resolution a multi-scale incremental approach is considered where the optimal solution is iteratively updated. This is done through successive morphings of the source towards the target image where uncertainties are used to determine the local precision. Efficient linear programming using the primal dual principles is considered to recover the lowest potential of the cost function. Results using synthetic and real data with known deformations demonstrate the extreme potentials of such an approach.

Joint work with B. Glocker, N. Komodakis, N. Navab & G. Tziritas