[rilk-poste] Poste Restauration d'images de macroscopie confocale

Restauration d'images de macroscopie confocale These DeadLine: 30/09/2009 caroline.chaux@univ-paris-est.fr La thèse se déroulera au Laboratoire d'Informatique Gaspard Monge http://igm.univ-mlv.fr/LabInfo qui est un laboratoire de recherche commun à l'université Paris-Est et le CNRS. Le LIGM est facilement accessible par les transports en commun (RER A, 15 min de Nation) In this work, we propose to investigate deconvolution algorithms in order to restore confocal macroscopy images, and to deal with noise, focus and motion artifacts in a consistent manner. Two preliminary points we have to deal with are the nature of the noise and the point spread function (PSF) models; a study of the characteristics of the noise (i.e. shape of its probability distribution and values of the related hyper-parameters) will be performed, using statistical estimation methods. This is a key point that will guide the choice of restoration method to be designed. A first working direction that will be investigated is a variational approach relying on a convex criterion to be minimized. This criterion consists of different terms which can be decomposed into two categories: data fidelity term and regularization terms. The classical problem can be formulated as the minimization of the sum of two convex functions (fidelity term g(x) and a priori term f(x)) over a real separable Hilbert space H. Two classes of optimization frameworks allowing us to address the resulting minimization problem will be investigated: 1) continuous convex optimization using advanced variational tools and proximal iterative algorithms; 2) discrete convex optimization using combinatorial optimization techniques related to graph theory, such as the celebrated graph cut methods, or related methods such as Continuous Maximum Flows. To the best of our knowledge, no comparisons between the two optimization strategies have been performed on practical image restoration problems and it would therefore be interesting to evaluate both of them in terms of convergence speed, complexity and precision in the context of macroscopy imaging. A combination of the two strategies could also be considered. A second investigation direction is the extension to non convex optimization problems which may correspond to more accurate models of degradation processes. Concerning further research directions, it is important to note that the mentioned variational approaches involve hyper-parameters to be set (scaling factors, exponents,...). Hyper-parameter estimation is a main problem that needs to be addressed. Some methods already exist such as Expectation-Maximization, Monte Carlo Markov Chains, cross-validation but their computational costs are often high. Discrete hyperparameter optimization methods will be investigated. Since the considered confocal macroscopy techniques provide large amounts of 3D+t data, they will be demanding in terms of computational complexity. However, biologists expect image processing delays of at most a few minutes. This means that attention must be paid to the implementation of the proposed methods. Implementations in C/C++ of the proposed algorithms will be essential, taking advantage of parallelism on multi-core or massively parallel architectures (by using libraries such as OpenMP, GPGPU) in order to accelerate the deconvolution process. Obviously these are complex problems that the prospective student is not expected to solve by themselves, but by working in a team. http://gdr-isis.org/rilk/gdr/Kiosque/poste.php?jobid=3440