17 janvier 2018

Catégorie : Doctorant

The SC team of IRIT laboratory is looking for candidates for a Ph.D position on Deep learning for embedded image compression in Earth Observation.

The objective of this Ph.D, co-funded by the CNES, the french spatial agency, is to assess the potential benefit and the feasibility of deep learning techniques for embedded image compression on-board Earth observation satellites.

The thesis will be co-supervised by Marie Chabert (Full Professor), Thomas Oberlin (Assistant Professor) and Charly Poulliat (Full Professor), Signal and Communication team, IRIT Institut de Recherche en Informatique de Toulouse.

The deadline for applications is 31/03/2018.

The desired profile is Master (MSc or equivalent) or Engineer degree in Image and Signal Processing / Computer Science / Applied Mathematics, with excellent academic record and research experience, in-depth knowledge of image and signal processing and computer science with a specialization in one of the following areas: probability and statistics, machine learning, computer vision, or digital/embedded electronics.

This thesis aims to assess the potential benefit and the feasibility of deep learning techniques for embedded image compression on-board Earth observation satellites. Deep learning is currently the subject of many studies. It uses artificial neural networks whose particularity, compared to conventional neural networks, is to include a large number of hidden layers [HS06].

The most straightforward and widespread use of deep learning is automatic classification [LBBH98], which can be performed directly from the data (images, texts or audio signals). The phase of extraction of relevant features, resulting from a prior choice in classical methods, is replaced here by a training phase (learning by example) on a sufficiently representative database. Thus, deep learning can provide very competitive performance when large datasets and computing power are available for the network training phase. Currently, these two conditions are satisfied in many application frameworks, which explains the strong interest aroused by deep learning in numerous domains.

The use of deep learning for the purpose of data compression, which is less intuitive than classification, is the subject of more recent works that provides very encouraging results [TOHVMBCS16]. Deep learning frequently outperforms state-of-the-art methods when the target compression ratios are high enough. However, the required computing power necessary for the training phase, and the complexity of the final architecture should be considered for a fair performance comparison.

The main objective of this thesis is to evaluate, both in terms of performance and complexity, deep learning-based approaches for embedded compression of Earth observation images. In this context, the availability of very large datasets of Earth observation images provides ideal conditions for the training of neural networks. This is a first positive point for the use of these approaches.

However, current studies focus more on their ability to compete with conventional techniques in terms of performance (optimization of a rate-distortion criterion) than in the associated operational complexity (i.e. implementation of the trained network).

This aspect is crucial in embedded image compression and generally leads to the adoption of sub-optimal solutions from the performance point of view. The simplified JPEG2000-like approach recommended by the Consultative Committee for Space Data Systems (CCSDS 122.0-B) is a good example of that. The first question to be answered is “which task must be assigned to the neural network within the classical compression scheme?”. Existing compression methods based on deep learning generally include auto-encoders as a substitute to both the transform & quantization stages involved in classical lossy compression schemes. An auto-encoder is a neural network that has a hidden layer of reduced size which gives access to a compact representation of the data. The results provided by auto-encoders for lossy compression are promising but a structural limitation lies in the fixed compression ratio provided (directly related to the number of nodes in the hidden layer). The recurrent auto-encoders which, at each iteration, encode the residue between the previous reconstruction and the original image, allow progressive coding and multiple compression ratios, as recommended by most compression standards, including the CCSDS. Thus, the first goal of this work will be to determine whether the conventional transform coding can advantageously be replaced by deep neural networks (possibly recurrent auto-encoders) in an embedded context. Different architectures will be exhaustively analyzed for the intended application. Additionally, it is important to notice that other techniques known as deep-compression [HMD16] make it possible to compress the network itself by trimming, quantizing and Huffman coding the network coefficients. The second goal of the thesis would be to integrate other features into the aforementioned scheme. In particular, an in-depth analysis of the network could allow the extraction of relevant information useful for compression itself or, more extensively, for other Earth observation applications: entropy coding, classification, denoising, image restoration, etc.

[LBBH98] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. 86(11), 1998.

[HS06] G. E. Hinton and R. R. Salakhutdinov. Reducing the dimensionality of data with neural networks. Science, 313(5786):504–507, 2006.

[TOHVMBCS16] G. Toderici, S. M. O’Malley, S. J. Hwang, D. Vincent, D. Minnen, S. Baluja, M. Covell, and R. Sukthankar. Variable rate image compression with recurrent neural networks. ICLR 2016, 2016.

[TVJHMSC16] G. Toderici, D. Vincent, N. Johnston, S. Jin Hwang, D. Minnen, J. Shor, M. Covell, Full Resolution Image Compression with Recurrent Neural Networks, arXiv:1608.05148.

[HMD16] S Han, H Mao, WJ Dally, Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding, International Conference on Learning Representations (ICLR'16 best paper award).

[DCCM10] X Delaunay, M Chabert, V Charvillat, G Morin, Satellite image compression by post-transforms in the wavelet domain, Signal Processing 90 (2), 599-610, 2010.

[DCCMR08] X Delaunay, M Chabert, V Charvillat, G Morin, R Ruiloba, Satellite image compression by directional decorrelation of wavelet coefficients, IEEE ICASSP 2008.

[DCMC07] X Delaunay, M Chabert, G Morin, V Charvillat, Bit-plane analysis and contexts combining of JPEG2000 contexts for on-board satellite image compression, IEEE ICASSP 2007.

All applications, including a motivation letter and a C.V., must be sent electronically to Marie Chabert, marie.chabert@enseeiht.fr

(c) GdR 720 ISIS - CNRS - 2011-2018.