Context: SAR images can be acquired all time (day and night and whatever the weather) which is a real advantage for tropical areas. Nevertheless the signal is affected by speckle noise which induces strong fluctuations.
Project: Mulitple SAR denoisers have been proposed in the past years like CNN based [1] or statistical approaches. They usually have different advantages and drawbacks. The objective of this project is to take benefit from different approaches to improve the denoising result. A CNN is fed by the noisy images and 3 denoised results presenting different artefacts. These results are obtained by MuLoG filter which processes the logarithm of the noisy image and can be used with different denoisers (TV minimization, DDID, BM3D), providing different results. A direct learning and transfer learning will be compared. The denoising code and CNN architecture will be provided with the training dataset.
Supervision: Florence Tupin, Emanuele Dalsasso
[1] https://arxiv.org/abs/1704.00275, SAR image despeckling through convolutional neural networks
G. Chierchia, D. Cozzolino, G. Poggi, L. Verdoliva
[2] Charles Deledalle, Loïc Denis, Sonia Tabti, Florence Tupin.
MuLoG, or How to apply Gaussian denoisers to multi-channel SAR speckle reduction?
IEEE Transactions on Image Processing 2017
https://www.charles-deledalle.fr/pages/mulog
[3] https://arxiv.org/pdf/1711.06712.pdf
