This thesis deals with the blind deconvolution of noisy data. We consider the case of seismic data. The inversion of the model need to select higher order statistics according to the distribution of the signals. To solve that, we use the assumptions of whiteness or of nongaussianity. We propose blind déconvolution algorithm in time domain and frequency domain. We measure whiteness by mutual information rate and nongaussianity with the negentropy. Afterwards, we study the sensitivity of the different algorithm with respect to a white Gaussian additive on the data. Theoretically and in practice on real and synthetic data, non-gaussianity appears as the method which provides the better trade off between déconvolution quality and noise amplification.
