Annonce

Les commentaires sont clos.

Internship subject : Deep Learning for Synthetic Aperture Radar Tomography

21 December 2022


Catégorie : Stagiaire


I - Context

Synthetic Aperture Radar (SAR) [4] is now a well-established remote sensing imaging modality for earth observation from satellites or planes. Among others, one of the qualities of SAR imaging is that the sensor is active, enabling to take images of the scene even during night or in the presence of cloudy weather conditions. Beyond single SAR image processing, tomographic SAR (TomoSAR) [6] uses a stack of complex images, acquired using slightly different points of views, to try to infer the tridimensional (3D) information of the scene. Precisely, the aim of TomoSAR is to recover a full 3D reflectivity profile, enabling to separate the different scatterers in the scene which are located at different heights and projected in the same range/azimuth resolution cell.

This intership will focus on the TomoSAR problem for urban scenes, which is of great interest for environmental risk management and urban planning, among others. One of the main specificities of urban TomoSAR is that there are usually only a small number of significant scatterers within a resolution cell. Based on this, several methods based on compressed sensing (CS) have been developped during the last decades : for instance, SL1MMER [9], truncated singular value decomposition (TSVD)-based CS and alternating direction method of multipliers (ADMM)-based l1 algorithm, to only name of few. Although such approaches have obtained good results for some scenes, they are largely impeded by a very high computational cost due to their iterative structure. In addition, they often heavily rely on handcrafted priors, which might only imperfectly model real datasets.

More recently, deep learning methods have been shown to be sucessful for a wide range of imaging problems. For TomoSAR, a conditional generative adversarial network is for instance used in [8]. However, likely due to the lack of massive datasets to train the neural networks and the difficulty of inverse problem to be solved, the use of deep learning for TomoSAR still remains largely to be explored. To bypass this issue, a promising research path is to incorporate some physical knowledge about the task at hand, enabling to design neural networks architectures requiring less training samples. To do that, a few works have studied the applicability of Algorithm Unrolling (AU) [2] to the TomoSAR problem [5, 7, 3]. The main insight of AU is to design neural networks architectures mimicking the structure of conventional CS-iterative algorithms : only a few parameters of such iterative algorithms are then learnt, requiring far less training samples than fully black-boxes neural networks. Compared to conventional CS algorithms, AU-based neural networks are generally much faster at test time and they lead to better results as they enable to have more data-driven approaches.

 

II - Internship main research paths

The goal of this internship is to further explore algorithm unrolling to design new neural networks for TomoSAR. Several pathways are to be investigated. First, deep equilibrium models [1] can, for the sake of simplification, be roughly considered as an extension of AU enabling to mimick iterative algorithms having an infinite number of layers. Such models have not so far been used for TomoSAR, which will be the first step of this internship. In addition, we believe that adding some further regularization in TomoSAR AU-based neural networks might improve their results, which will be tackled in a second part of the internship. On top of this two main subjects, a careful reflexion about relevent datasets to use for the neural network training will have to be considered.

 

III - Candidate

The candidate must follow a Master 2 program (or equivalent) and have good knowledge of signal/image processing, as well as (deep) machine learning. Ideally, Python and its learning modules (Pytorch) should be known. In addition, knowledge about remote sensing is a plus, as well as in convex optimization.

During the internship, the candidate will acquire knowledge in image/signal processing, deep learning, convex optimization and inverse problems. The skills learnt can be useful in various domains : remote sensing, medical imaging, astrophysics...

 

IV - Contact

The internship will be held in the IMAGES team (T ́el ́ecom-Paris), under the supervision of Christophe Kervazo and Florence Tupin.

Contact: christophe.kervazo@telecom-paris.fr;florence.tupin@telecom-paris.fr

Possibility to continue as PhD : very likely.

More information on https://sites.google.com/view/christophekervazo/ and https://perso.telecom-paristech.fr/tupin/

 

Références

[1] Shaojie Bai, J Zico Kolter, and Vladlen Koltun. Deep equilibrium models. Advances in Neural Information Processing Systems, 32, 2019.

[2] Karol Gregor and Yann LeCun. Learning fast approximations of sparse coding. In Proceedings of the 27th international conference on international conference on machine learning, pages 399–406, 2010.

[3] Yunqiao Hu, Xiaoling Zhang, Shunjun Wei, Yu Ren, Nan Wang, and Jun Shi. Adatomo-net : a novel deep learning approach for sar tomography imaging and autofocusing. In IGARSS 2022-2022 IEEE International Geoscience and Remote Sensing Symposium, pages 587–590. IEEE, 2022.

[4] Alberto Moreira, Pau Prats-Iraola, Marwan Younis, Gerhard Krieger, Irena Hajnsek, and Konstantinos P Papathanassiou. A tutorial on synthetic aperture radar. IEEE Geoscience and remote sensing magazine, 1(1) :6–43, 2013.

[5] Kun Qian, Yuanyuan Wang, Yilei Shi, and Xiao Xiang Zhu. γ-net : Superresolving sar tomographic inversion via deep learning. IEEE Transactions on Geoscience and Remote Sensing, 60 :1–16, 2022.

[6] Clement Rambour, Alessandra Budillon, Angel Caroline Johnsy, Loic Denis, Florence Tupin, and Gilda Schirinzi. From interferometric to tomographic sar : A review of synthetic aperture radar tomography- processing techniques for scatterer unmixing in urban areas. IEEE Geoscience and Remote Sensing Magazine, 8(2) :6–29, 2020.

[7] Muhan Wang, Zhe Zhang, Yue Wang, Silin Gao, and Xiaolan Qiu. Tomosar-alista : Efficient tomosar imaging via deep unfolded network. arXiv preprint arXiv :2205.02445, 2022.

[8] Shihong Wang, Jiayi Guo, Yueting Zhang, Yuxin Hu, Chibiao Ding, and Yirong Wu. Tomosar 3d reconstruction for buildings using very few tracks of observation : A conditional generative adversarial network approach. Remote Sensing, 13(24) :5055, 2021.

[9] Xiao Xiang Zhu and Richard Bamler. Super-resolution power and robustness of compressive sensing for spectral estimation with application to spaceborne tomographic sar. IEEE Transactions on Geoscience and Remote Sensing, 50(1) :247–258, 2011.