Starting date: October or November 2020
Duration: 12 months (with a possible extension of the duration)
Application deadline date: June 30th 2020
Decision announcement date: July 10th 2020
In the ADEME IMOTEP project context (Innovation MOTEur Propre - see here at the bottom of the web page for a presentation of the partners of this project), this PostDoc subject is proposed by the Image team of the Hubert Curien laboratory in collaboration with the “Laboratory of Tribology and Dynamic of Systems” (LTDS) of the “Ecole Centrale” in Lyon and IREIS (Institut de Recherche En Ingénierie des Surfaces), the research and development department of HEF group.
The work will also be carried out in collaboration with a PhD student whose PhD thesis started on December 2018. The subject proposes to address the denoising or source separation tasks applied to cyclostationary signals, using the framework of deep learning methodology. Cyclostationary signals are random signals whose statistical characteristics have periodicities [GN06]. Signals of this type are encountered in many fields such as telecommunications, radar signals, acoustic signals, mechanical signals related to rotations in engines, etc. This kind of signals can also be encountered in the field of imaging, where the cyclostationarities appear in the two spatial dimensions. This problematics can then be extended to every dimensionality of data (temporal, spatial, spatio-temporal, etc.). Hence, one of the goal of this post doc is to find new advanced general methodologies to process such data.
Whatever the dimensionality of the problem, the processing of these signals is often related to the denoising and the separation of a mixture of several signals (or sources). One can notice that the denoising task can also be seen as the separation of the useful signal from the noise [BAM97, AMX01]. It is also possible to consider improving the considered methods by detecting (or recognizing) the type of signals involved in the mixture, which then makes it possible to choose the most appropriate models. Our current developments explore also the synthesis of appropriate (or physically based) signals to improve deep learning architectures.
Recently, new methods of denoising and source separation, exploiting the current advances in machine learning based on deep learning (LSTM-RNN [KY17], deep NN [KK18], Autoencoders [GP17], etc.) have emerged for speech or audio signal processing. These methods can be inspired by image denoising approaches [TF19].
Our objective is to extend and generalize these works to cyclostationary signals with supports of different dimensionalities, and to study how to improve the training of architectures with synthetic datasets. For multidimensional data, the work will focus on the denoising of images and videos that contain some periodic geometrical shapes. Applications of the methods will be done on: (1) measurement signals (tribological measurements) acquired on test benches for the engines of the future and (2) scanning electron microscopy image sequences data.
We are looking for a motivated student with a PhD degree in signal / image processing, in computer vision or in applied mathematics with an experience in signal / image processing. A good background in software development (algorithmic, Matlab/Octave/Scilab or Python, ...) is expected. Knowledge in machine learning would also be appreciated.
Net salary: around 2200 euros. Teaching activities are eventually possible.
Your application should include the following documents:
[AMX01]K. Abed-Meraim, Y. Xiang, J.H. Manton, Y. Hua, "Blind source-separation using second-order cyclostationary statistics" IEEE Transactions on Signal Processing, 49(4), 2001.
[BAM97] A. Belachrouni, K. Abed-Meraim, J.-F. Cardoso, E. Moulines, "A Blind Source Separation Technique Using Second-Order Statistics" IEEE Transactions on Signal Processing, 45(2), 1997.
[GN06] W. A. Gardner, Antonio Napolitano, Luigi Paura, "Cyclostationarity: Half a century of research", Volume 86, Issue 4, April 2006, Pages 639-697.
[GP17] E. M. Grais, M. D. Plumbley, "Single Channel Audio Source Separation using Convolutional Denoising Autoencoders" In: 5th IEEE Global Conference on Signal and Information Processing (GlobalSIP2017), 14 - 16 November 2017, Montreal, Canada.
[KK18] M. Kulin, T. Kazaz, I. Moerman, E. De Poorter, "End-to-end learning from spectrum data: A deep learning approach for wireless signal identification in spectrum monitoring applications", IEEE Access, 2018. https://arxiv.org/abs/1712.03987
[KY17] M. Kolbæk, D. Yu, Z.-H. Tan, J. Jensen, "Joint Separation and Denoising of Noisy Multi-talker Speech using Recurrent Neural Networks and Permutation Invariant Training" In: IEEE Int. Workshop on Machine Learning for Signal Processing, 2017, Tokyo, Japan. https://arxiv.org/abs/1708.09588
[TF19] C. Tian, L. Fei, W. Zheng, Y. Xu, W. Zuo and C. Lin, “Deep Learning on Image Denoising: An overview”, arXiv 2019, https://arxiv.org/abs/1912.13171.
(c) GdR 720 ISIS - CNRS - 2011-2020.