Even though noise present in satellite optical images is a mixture of Poisson and Gaussian noise, the great majority of denoisers work only in the case of Gaussian noise.
A common way to circumvent this inadequacy consists in
- Approximating the mixed noise by a Gaussian noise with luminance-dependent variance; and
- Applying a variance stabilizing transform (Anscombe) that transforms this luminance-dependent variance in approximately constant variance.
Even though the first approximation is sufficiently accurate, the second one may produce artifacts in dark areas as observed for instance in [1].
An alternative way to apply a Gaussian denoiser in a non-Gaussian setting is to plug in the Gaussian denoiser into a plug and play ADMM method like the one proposed in [2].
The purpose of this project is to compare the performance of the plug & play ADMM approach in [2] versus the traditional Anscombe VST approach. For this you shall use one of the two top-performing Gaussian denoisers: NLBayes [3] or FFDnet [4].
Time permitting you can also test the performance of the P&P-ADMM approach in the presence of compression noise in addition to mixed noise as described in [5].
Supervision
Andrés Almansa
References
[1] Preciozzi, J., Gonzalez, M., Almansa, A., & Musé, P. (2017). Joint denoising and decompression: A patch-based Bayesian approach. In (ICIP) IEEE International Conference on Image Processing (Vol. 2017–Sept, pp. 1252–1256). IEEE. [doi:10.1109/ICIP.2017.8296482] [preprint]
[2] Chan, S. H., Wang, X., & Elgendy, O. A. (2017). Plug-and-Play ADMM for Image Restoration: Fixed-Point Convergence and Applications. IEEE Transactions on Computational Imaging, 3(1), 84–98. https://doi.org/10.1109/TCI.2016.2629286
[3] Lebrun, M., Buades, A., & Morel, J.-M. (2013). Implementation of the “Non-Local Bayes” (NL-Bayes) Image Denoising Algorithm. Image Processing On Line, 2013, 1–42. https://doi.org/10.5201/ipol.2013.16
[4] Zhang, K., Zuo, W., & Zhang, L. (2018). FFDNet: Toward a Fast and Flexible Solution for CNN based Image Denoising. IEEE Transactions on Image Processing. https://doi.org/10.1109/TIP.2018.2839891
[5] González, M., Preciozzi, J., Musé, P., & Almansa, A. (2018). Joint denoising and decompression using CNN regularization. In CVPR Workshop and Challenge on Learned Image Compression (pp. 2598–2601). Salt Lake City, Utah, United States. Retrieved from https://hal.archives-ouvertes.fr/hal-01825573
