Accueil du site > Equipes > Signal et Image (SI)

**Signal & Image **is a research team of the Laboratoire d’Automatique, Génie Informatique et Signal de Lille (LAGIS – UMRS CNRS 8219) which is associated to the National Center of Scientific Research (CNRS), Ecole Centrale Lille and the Université de Lille1. Five topics are developped in this team.

**Polarimetric Imaging (V. Devlaminck, P. Terrier, J.-M. Charbois)**

Polarimetric imaging, using the polarimetric properties of light to build a multi-dimensional imaging can be traced back to the physical or geometric objects illuminated. Algorithms in image processing or statistical estimation, may be ineffective on this type of imaging because they can provide solutions physically ineligible under Euclidean framework. In recent years, approaches such as Information Geometry have emerged. They are to define a framework directly connected to the data to be processed by constructing appropriate distances and building algorithms to process these images from these distances. These techniques are of interest to solve the difficulties encountered by the Euclidean approach in the case of polarimetric data. We are interested in developing such new data processing tools in the framework of polarimetric data. Determine which solution to choose depending on the context is also a question that we address.

**Information Forensics and Security (P.Bas, W. Sawaya, P. Chainais)**

This project focuses on how to protect contents using data-hiding techniques (Watermarking / Steganography) or authentication. Part of our research activities are devoted to security analysis of (1) watermarking, (2) steganographic (called steganalysis) or (3) authentication schemes. To this end several machine learning techniques such as Independent Component Analysis, Clustering techniques or Classification are used in order to try to infer the secret key used by the analysed scheme. The security analysis enables also to improve the design of new protection schemes that will be immune to the different security attacks. Theoretical security measures are also derived in order to benchmark the practical security of protection schemes.

**Information Processing for Multi Sources Systems (E. Duflos, P. Vanheeghe, F. Septier, C. Garnier, J. Klein, O. Colot, J-M Vannobel, P. Chainais, Y. Delignon)**

We manage both uncertainty and imprecision in the frame of probability theory, belief function theory and fuzzy logic.

**- Statistical Signal Processing :** we develop new Bayesian methods for analysis and estimation purposes. The resulting algorithms are themselves based on Sequential Monte Carlo and MCMC methods for which evolutions are also proposed. The originality of the work carried out are (1) the estimation in high dimension with application to tracking, (2) the development of Bayesian non parametric (BNP) methods for signal processing and (3) the development of multisensor multi target tracking methods based on finite random sets (like the Probability Hypothesis Density (PHD) filter). We have been the first to propose BNP methods for estimation purpose in non gaussian dynamical systems.

**- Uncertainty models and imperfect data processing :** collected data are tainted with several kinds of imperfections : uncertainty, imprecision, incompleteness, unreliability or conflict. In order to be efficiently processed, such data need to be represented by uncertainty models that best appraise what is actually known about them. To that end, we develop models in the framework of the belief function theory and fuzzy logic.

**Semi-supervised classification of color textures by Color Filter Array (CFA) image analysis (L. Macaire, O. Losson, B. Mathon, L. Duvieubourg)**

Single-sensor color cameras acquire CFA images in which pixels are characterized by only one of the three color components red, green, blue. Demosaicing schemes, that estimate the two missing color components in order to build the color images, tend to alter the local texture infor- mation that is useful to discriminate the textures. That leads us to adapt well-known texture features, such as chromatic co-occurrence matrices or local binary patterns, to the specific properties of CFA images. The most discriminant features can be selected in a semi-supervised learning context by defining constraints. These works stem from initial studies about color image segmentation by pixel fuzzy classification and color space selection. Their experimental validation requires representative databases of outside color textures.

**Brain Computer Interfaces (M.-H. Bekaert, F. Cabestaing and C. Lecocq) **

BCI aims at developing brain-computer interfaces for palliative communication, i.e. to enable deeply handicapped people to communicate. BCI is a highly interdisciplinary R&D field and the group has established strong relations with ergo-therapists (CHRU, Lille), ergonomists (Paris 8 University), neurophysiologists (CHRU, Lille), neurosurgeons (Hôpital Laënnec, Nantes). The group activity is not only focused on developing new signal processing techniques for BCI, but also on using efficiently devices and algorithms for helping physicians or therapists to use and evaluate this new technology. Visite the BCI website on http://bci-team.univ-lille1.fr .