SAR images are strongly disturbed by multiplicative noise. Recent progress in image denoising could be exploited to improve state of the art SAR image denoising methods.
The aim of this project is to study NL-Bayes [1] and to adapt it for SAR data. A first approach will be to do some variance stabilization by taking the logarithm of the SAR image, but other adaptations could be necessary (for the patch-grouping step, or the denoising part, see [2] for example).
The principle of NL-Bayes is to associate to each patch a prior distribution chosen to be a Gaussian multi-variate model. Therefore a covariance matrix reflecting this local approximation is available at the end of the algorithm. Other approaches, like EPLL [3] use a dictionary of Gaussian models to represent the patches and select the best one for each patch. A comparison between the learned Gaussian of NL-Bayes and the Gaussian of the dictionary will be done, specially for “rare patch” to evaluate the interest of the dictionary ([3]) against the local learning ([1]).
Steps of the project:
– NL-Bayes on logarithm of SAR images
– Study of possible adaptation of NL-Bayes
– Comparison of local Gaussian against EPLL Gaussians (given)
Programming
Some code is available for these different approaches but it could be more efficient to developp the code (even slower) to be able to adapt and compare the different methods.
Matlab or C/C++
Supervision
F. Tupin, A. Almansa, A. Houdard
References
[1] http://www.ipol.im/pub/art/2013/16
A. Buades, M. Lebrun, and J.M. Morel. “A Non-local Bayesian Image Denoising Algorithm.” SIAM Journal on Imaging Sciences, 2014
[2] https://hal.archives-ouvertes.fr/hal-01006733
S. Tabti, C. deledalle, L. Denis, F. Tupin, “Modeling the distribution of patches with shift-invariance: application to SAR image restoration”, ICIP 2014
[3] “From Learning Models of Natural Image Patches to Whole Image Restoration” Daniel Zoran and Yair Weiss, ICCV 2011 (oral presentation), PDF Supp code