Improving a denoising strategy to restore linear structures
Encadrants : Florence Tupin, Emanuele Dalsasso
Email : florence.tupin@telecom-paris.fr
Disponible : OUI
Spécialité : IMA
Nombre d'étudiants : 1
Description : SAR (Synthetic Aperture Radar) images have the advantage of being
acquired at any time (day or night and whatever the weather conditions).
However, they are very noisy because of the speckle noise due to
coherent imaging. Deep networks have provided impressive speckle
reduction results but still erase the fine dark linear structures
corresponding to the road network.
Travail demandé : The aim of this project is to develop a strategy to help the network to
preserve this thin information. To do so we propose to introduce the
result of a line detector to modify the loss and constraint the network
to better preserve the linear structures. Different strategies could be
developed but a first step will be to train the network in a supervised
setting [1] using a line detection algorithm [2] to weight the loss.
Liens complémentaires : References
[1] Dalsasso, E., Yang, X., Denis, L., Tupin, F., & Yang, W. (2020). SAR
image despeckling by deep neural networks: From a pre-trained model to
an end-to-end training strategy. Remote Sensing, 12(16), 2636.
https://www.mdpi.com/2072-4292/12/16/2636/pdf
[2] Generalized Likelihood Ratio Tests for Linear Structure Detection in
SAR Images
Nicolas Gasnier, Loïc Denis, Florence Tupin
EUSAR 2021: 13th European Conference on Synthetic Aperture Radar, Mar
2021, Leipzig (virtual), Germany
[3] SAR2SAR: a semi-supervised despeckling algorithm for SAR images
Emanuele Dalsasso, Loïc Denis, Florence Tupin
IEEE Journal of Selected Topics in Applied Earth Observations and Remote
Sensing, IEEE, In press, pp.1-1. ⟨10.1109/JSTARS.2021.3071864⟩