This work deals with blind source separation in the context of convolutive mixtures. The problem is considered in the case of non linear, non i.i.d. sources. Criteria are studied, which are called contrast functions, and which maximization leads to the separation of a mixture of sources. Among separation methods, global approaches extract all the sources simultaneously, whereas iterative methods extract the sources one by one. An alternative and an improvement of the so-called "deflation" iterative method are proposed. The potential of iterative methods is illustrated through applications to both communication CPM (continuous phase modulation) signals and images. Finally, starting from a semi-blind point of view, quadratic contrast functions are derived. They offer very good performances, even in a blind context, and their optimisation speed is improved.
