Applications of Wavelet to Independent Component Analysis

Independent component analysis (ICA)/blind source separation (BSS) has received many attentions in neural network and signal processing in recent years. In applications the observed signals are often corrupted with high noise (low SNR, low sample size, non-Gaussian noise), the source number is unknown, and the sources are non-stationary, which are not well correspond to the ideal ICA models and as a result the effectiveness of the classic ICA algorithm is definitely decreased. In this paper we review and give some new ideas to tackle all these problems by using the well-known mathematics tool _iscrete wavelet transform (DWT). Using DWT we can remove the corrupted high noise; efficiently detect the number of the sources and move the conventional time-domain natural gradient ICA algorithm to the wavelet-domain natural gradient algorithm because this procedure makes the algorithm brief and has a wider applications in real world signal processing