Blind Source Separation using Orientation Histograms in Joint Mixture Distributions

J. Yamashita and Y. Hirai (Japan)


Blind Source Separation, Orientation Histograms, Fourier Transformation


Blind source separation (BSS) is one of the most promising research activities in the neural network lit erature. In the previous work we proposed a BSS algo rithm, which could find source orientations from ori entation histograms in a joint distribution of observed mixtures [1]. Although the algorithm works well for simultaneous mixtures of sources with supergaussian probability densities, many natural sources have sub gaussian densities and propagation delays. In this pa per we extend the algorithm by introducing Fourier transformation for the preprocessing of observed mix tures. It will be shown that the extended algorithm can separate convoluted mixtures of sub- and super gaussian sources with different delays. Since our algo rithm does not suffer from arbitrariness in permutation and amplitude of recovered sources as ICA algorithms do, it could be much simpler and easier to construct a real-time BSS system.

Important Links:

Go Back