Bibliografie

(eng): Blind source separation algorithms are based on various separation criteria. Differences in convolution kernels of the sources are common assumptions in audio and image processing. Since it is still an ill posed problem, any additional information is beneficial. In this contribution, we investigate the use of sparsity criteria for both the source signal and the convolution kernels. A probabilistic model of the problem is introduced and its Variational Bayesian solution derived. The sparsity of the solution is achieved by introduction of unknown variance of the prior on all elements of the convolution kernels and the mixing matrix. Properties of the model are analyzed on simulated data and compared with state of the art methods. Performance of the algorithm is demonstrated on the problem of decomposition of a sequence of medical data. Specifically, the assumption of sparseness is shown to suppress artifacts of unconstrained separation method.