Post-doc position: Regularization of SAR images using energy minimization and non-local means approaches

SAR images are difficult to process due to both speckle noise and their very high dynamic, especially in urban areas. The recent launch of a new generation of SAR sensors (TerraSAR-X, CosmoSkyMed) has increased the need for regularization methods making easier the use of such data (amplitude data, but also interferometric data). The aim of this project is to develop new approaches based on energy minimization and / or non-local means. The following research axes should be investigated.

The first axis deals with the statistical modelling of the problem. What is a good model for SAR images, both for the SAR data term (i.e., likelihood) and the prior on the solution? Strong scatterers are very important features in urban areas. They are not well modeled by standard approaches such as total variation minimization.

The second axis is dedicated to the optimization problem. In the case of SAR data, the likelihood models are non convex. Methods devised either in the continuous domain [1] or in the discrete domain using graph-cut optimization [2] have been recently developed. A comparison between both frameworks should be led to analyze their advantages and drawbacks.

The third objective is to derive new optimization algorithms dedicated to the case of SAR data. Indeed, a compromise between memory/computational cost and optimum quality can be obtained using fast and approximate optimization algorithms [3]. Based on the new statistical models investigated in the project, new optimization methods will be studied, either in the continuous or discrete domain.

Finally, non-local approaches deriving from the NL-means principle have shown promising results in the context of SAR image restoration. One may hope to get the most of non-local and energy-minimization approaches by designing a combined scheme.

This project is supported by DGA and is developed in collaboration between 3 laboratories: Telecom ParisTech – LTCI (Paris), Ecole Normale Supérieure – CMLA (Cachan) and CPE Lyon – LHC (Saint-Etienne). The developed algorithms will be evaluated by the society IPSIS to measure their interest for defence applications. The position is opened for 18 months (starting in September or October 2009) and localized in Paris at Telecom ParisTech, in the Image Processing group http://www.tsi.enst.fr/~bloch/tii/index.html.

[1] G. Aubert, J-F. Aujol, A Variational Approach to remove Multiplicative Noise, SIAM Journal on Applied Mathematics, volume 68, number 4, pages 925-946, January 2008

[2] H Ishikawa, Exact optimization for Markov random fields with convex priors, IEEE trans. on Pattern Analysis and Machine Intelligence, volume 25, number 10, 2003.

[3] L. Denis and F. Tupin and J. Darbon and M. Sigelle, SAR Image Regularization with Fast Approximate Discrete Minimization, IEEE Transactions on Image Processing, in press, 2009 (doi: 10.1109/TIP.2009.2019302)

Contact: florence.tupin@telecom-paristech.fr, loic.denis@cpe.fr

The applicant should have a PhD in image processing. An experience in energy minimization approaches would be highly appreciated. Knowledge of SAR image domain is a plus.