Radar image denoising with CNN priors

Context

In lectures 3 and 7 we explain how to model radar image noise and how to adapt non-local patch based denoising approaches (that assume additive gaussian noise) to radar images, where the noise has different characteristics.

More recently, some convolutional neural network based approaches [1] claim to achieve denoising performance beyond many of the leading patch based non-local denoising methods for additive Gaussian noise.

On the other hand splitting techniques can be used to turn any denoiser (including a CNN-based one) into a regularizer for many different inverse problems [2,3,4].

Objective

The objective of this project is to understand the splitting (plug & play) approach in either [2,3] or [4] and how to plug-in an appropriate data-fitting term for radar image denoising based on the course materials. A more detailed explanation is available in [6]. Alternatively one could train a neural network end-to-end to restore noisy radar images [5].

The proposed framework is built in such a way that it can use an off-the-shelf denoiser model as regularizer without need for retraining. Retraining the denoiser on radar image data may improve the results but may exceed the scope of this project.

The most important analysis and programming part involves writing the proximal operator for the data fitting term, i.e. \(-\log P(\tilde{u} | u)\) the negative log-conditional-likelihood of the noisy image \(\tilde{u}\) given the clean image \(u\).

Subprojects

This project may be split into three subtasks:

  1. Understand one of the splitting techniques [2,3,4] and how to adapt them to radar noise [6], using an off-the shelf gaussian denoiser like [1].
  2. Retrain the denoiser [1] using noiseless radar images and simulated gaussian noise. Analyze the effects of using the retrained denoiser as a regularizer in the previous task.
  3. Train a denoising network [1,5] on radar images with realistic noise.

Programming

Matlab or Python

Supervision

Andrés Almansa & Florence Tupin

Bibliographic References

[1] Zhang, Kai, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. 2017. “Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising.” IEEE Transactions on Image Processing 26 (7): 3142–55. doi:10.1109/TIP.2017.2662206. [preprint]

[2] Meinhardt, Tim, Michael Moeller, Caner Hazirbas, and Daniel Cremers. 2017. “Learning Proximal Operators: Using Denoising Networks for Regularizing Inverse Imaging Problems,” April. http://arxiv.org/abs/1704.03488.

[3] Fan, Kai, Qi Wei, Wenlin Wang, Amit Chakraborty, and Katherine Heller. 2017. “InverseNet: Solving Inverse Problems with Splitting Networks,” December. http://arxiv.org/abs/1712.00202.

[4] Zhang, Kai, Wangmeng Zuo, Shuhang Gu, and Lei Zhang. 2017. “Learning Deep CNN Denoiser Prior for Image Restoration,” In CVPR 2017. http://arxiv.org/abs/1704.03264.

[5] Chierchia, G., D. Cozzolino, G. Poggi, and L. Verdoliva. 2017. “SAR Image Despeckling through Convolutional Neural Networks.” In (IGARSS 2017) IEEE International Geoscience and Remote Sensing Symposium, 5438–41. IEEE. [doi:10.1109/IGARSS.2017.8128234] [arXiv:1704.00275]

[6] Deledalle, Charles-Alban, Loïc Denis, Sonia Tabti, and Florence Tupin. 2017. “MuLoG, or How to Apply Gaussian Denoisers to Multi-Channel SAR Speckle Reduction?” IEEE Transactions on Image Processing 26 (9): 4389–4403. [doi:10.1109/TIP.2017.2713946] [preprint] [arXiv:1704.05335][hal-01388858]

This entry was posted in Projets2018. Bookmark the permalink.