Vous êtes ici : Accueil » Kiosque » Annonce


Mot de passe : 

Mot de passe oublié ?
Détails d'identification oubliés ?


5 mai 2019

PhD Thesis proposal at Bordeaux University : Superresolution of multiscale 3D images from material sciences

Catégorie : Doctorant

Recent and ongoing developments in imaging techniques and computational analysis deeply modify the way materials science and engineering consider their objects of research. The motivation to develop new Superresolution (SR) methods originates in these modifications.
3D imaging has moved from micron-scale resolution to true nano-scale regime allowing the passage from statistical analysis based on averages to analysis of the modulations of the averages, the defects and details at the interfaces as important and defining characteristics for performances. Studying such material is very challenging because multi-modal imaging techniques are often required producing high-resolution (HR) and low-resolution (LR) images having different characteristics. Addressing this problem directly in all its complexity is unrealistic and a step by step approach will be adopted.
Superresolution (SR), the process of obtaining one or more HR images from one or more LR observation(s) has found applications in many real-world problems. Over the past two decades a large number of research papers and books addressing specific practical purposes have been written.
For various inverse problems including image denoising or restoration techniques, stochastic modeling for patch characterization has been recently studied. Using a global Gaussian mixture modeling (GMM), Yu etal. [6] proposed a patch-based Bayesian approach. For the SR problem, Sandeep et al. [3,5] extend the previous modeling, i.e. GMM, for joint HR-LR modeling given the opportunity to compute an optimal estimator based on the conditional expectation. Various works [1, 2, 4] have shown the possibility to consider pdf models which have more flexibility to adapt the shape of data and less sensibility for over-fitting the number of classes than the GMM. Indeed, including Gaussian and Laplacian distributions as special cases, generalized Gaussian mixture modeling (GGMM) are potentially interesting for modeling the statistical properties of various images or features extracted from these images.
In [2, 4], we have proposed different algorithms for estimating GGM parameters. The first goal of this PhD thesis is to extend the previous works to the GGMM case. We intend to capture the HR statistics from the LR image in a more appropriate way by using this richer modeling. The second goal of this PhD thesis is to use GMM and/or GGMM on other representations of the image such as the output of some neural networks.
  • Interdisciplinary project involving 3 laboratories from Bordeaux campus (ICMCB, IMB, IMS) in connection with 2 international projects (European network MUMMERING, ANR/DFG project SUPREMATIM (collaboration with Kaiserslautern University)).
  • PhD thesis (funded by ANR/DFG project SUPREMATIM). Starting date: September 2019.
  • Joint supervision by J-F Aujol, Y Berthoumieu, and D Bernard.
  • Strong background in image processing and applied mathematics is required.
Send a detailed CV, a letter stating the reasons of your application.
  • Jean-Francois.Aujol@math.u-bordeaux.fr
  • Yannick.Berthoumieu@ims-bordeaux.fr
  • Dominique.Bernard@icmcb.cnrs.fr


[1] M. Said Allili. Wavelet modeling using finite mixtures of generalized Gaussian distribu-tions: Application to texture discrimination and retrieval.IEEE Trans. Image Processing,21(4):1452–1464, 2012.

[2] Z. Boukouvalas, S. Said, L. Bombrun, Y. Berthoumieu, and T. Adal. A new Riemannianaveraged fixed-point algorithm for MGGD parameter estimation.IEEE Signal ProcessingLetters, 22(12):2314–2318, Dec 2015.

[3] CA Deledalle, S Parameswaran, and TQ Nguyen. Image denoising with generalized gaus-sian mixture model patch priors.SIAM Journal on Imaging Sciences, 11(4):2568–2609.

[4] F. Pascal, L. Bombrun, J.-Y. Tourneret, and Y. Berthoumieu. Parameter estimationfor multivariate generalized Gaussian distributions.IEEE Trans. Signal Processing,61(23):5960–5971, 2013.

[5] P. Sandeep and T. Jacob. Single image super-resolution using a joint GMM method.IEEETransactions on Image Processing, 25:4233–4244, 2016.

[6] G. Yu, G. Sapiro, and S. Mallat. Image modeling and enhancement via structured sparsemodel selection. InImage Processing (ICIP), 2010 17th IEEE International Conferenceon, pages 1641–1644. IEEE, 2010.

Dans cette rubrique

(c) GdR 720 ISIS - CNRS - 2011-2019.