Publications abstracts

Journal papers

  • Marc Castella, Jean-Christophe Pesquet and Athina P. Petropulu, A Family of Frequency- and Time-Domain Contrasts for Blind Separation of Convolutive Mixtures of Temporally Dependent Signals. Accepted for publication in IEEE Trans. on Signal Processing
    • [pdf]
    This paper addresses the problem of blind separation of convolutive mixtures via contrast maximization. New frequency domain contrast functions are constructed based on higher-order spectra of the observations. They allow to separate mixtures of sources which are spatially independent, and temporally possibly non linear processes. Using Parseval's formula, the former criteria yield a general class of time-domain contrasts, which extends to the convolutive case results that have been previously obtained in the context either of instantaneous mixtures or of i.i.d. sources. A Monte-Carlo simulation study is carried out for comparison between the different contrasts, thus providing a guideline about the choice of an appropriate contrast.

    Conferences

  • Marc Castella, Jean-Christophe Pesquet and Athina P. Petropulu, New contrasts for blind separation of non iid sources in the convolutive case. EUSIPCO 2002, Vol.2, pp.107-110, Toulouse, France [pdf]

    • This paper deals with the separation of convolutive mixtures of independent source signals. Starting from a frequency point of view, we consider a joint diagonalization criterion. Integration of the latter over the frequency domain leads to contrasts which are valid for both iid and non iid sources. A generalization of these contrasts is proposed: it aims at obtaining contrasts with improved statistical performances. A link with existing time-domain contrasts is established. A simple gradient based method for the optimization of the contrasts is proposed and evaluated through simulation tests.
  • Marc Castella and Jean-Christophe Pesquet, Source separation of a class of non linear time series. NSIP 2003, Grado, Italy [pdf]

    • This paper deals with blind source separation of non linear signals. We therefore consider non i.i.d. source signals which are mixed by an unknown convolutive filter. Previous approaches succeeded in designing contrasts which are valid for non i.i.d. sources, but which cannot be handled easily in practice. Starting from this point, we propose some improvements and simplifications of the contrasts for the general class of non i.i.d. multiplicative-like signals: the obtained criteria are also expected to exhibit better statistical performances. Classical optimization algorithms are used for the local maximization of the contrasts, and genetic algorithms are tested in order to reach the global maximum of our criteria. Simulations prove the validity of our contrasts and illustrate the difficulty to choose the most appropriate one.
  • Marc Castella, Antoine Chevreuil et Jean-Christophe Pesquet, Séparation aveugle d'un mélange convolutif de sources non linéaires par une approche hiérarchique. GRETSI 2003, Paris, France
    • [pdf]
    Cette communication concerne la séparation aveugle de sources dans le cas de mélanges convolutifs de sources non linéaires (et donc non i.i.d.). Ce cadre général, bien que peu souvent envisagé dans la littérature, est pourtant central d'un point de vue applicatif.
    Une méthode itérative a été proposée \cite{simon/loubaton-sp01}; cependant la déflation entraîne une accumulation d'erreurs parfois gênante. Nous proposons une approche hiérarchique inspirée de \cite{inouye-i3esp99} pour laquelle ce phénomène d'accumulation n'est pas sensible.
    Nous démontrons le bien-fondé de la méthode et sa convergence vers un filtre séparant dans le cas général de sources non linéaires et non i.i.d. Enfin des simulations valident ces considérations théoriques pour des modèles classiques de processus non linéaires (ARCH).
  • Marc Castella, Eric Moreau and Jean-Christophe Pesquet, A quadratic MISO contrast function for blind equalization. ICASSP 2004, Montréal, Canada
    • [pdf]
    This paper is concerned with blind separation of convolutive mixtures of mutually independent signals. We consider the MISO extraction of one source signal based on the maximization of a contrast function (CF): a new, so-called ``reference'' CF is proposed, which is based on cross-statistics between the estimated output and a reference signal. The proposed CF is valid both for i.i.d. and non i.i.d. sources. It presents the advantage over other CFs to be a quadratic function, which makes its optimization much easier to realize. Finally, simulations demonstrate the validity of this CF and show that it leads to improved separation performances.

    Dernière MAJ: