engleski

Blind separation of statistically dependent sources

Blind source separation is a field developed in signal processing and neural network communities over last 15-20 years. It found numerous applications in science and engineering such as acoustics, biomedical signal analysis, communications, image segmentation and deconvolution, spectroscopy, bioinformatics, finance, etc. The basic static linear blind source separation problem is efficiently solved by means of independent component analysis under standard assumptions: sources are statistically independent and non-Gaussian, and column-rank of the unknown basis or mixing matrix equals the unknown number of sources. However, in a number of applications statistical independence assumption does not hold completely. Examples include biomedical data sets such as EEG, fMRI, etc. We shall present methodology aimed to resolve this issue, and demonstrate its efficiency in novel algorithms for single channel blind image and signal deconvolution, blind separation of the images of human faces, as well as for unsupervised decomposition of low-dimensional multispectral images.

blind source separation; statistically dependent sources; dual tree wavelets; independent component analysis; mutual information

nije evidentirano