A new blind signal separation approach based on multistage QR factorization

In this paper, two-step separation technique based on subband QR factorization is introduces. The first step is the sphering step and the second step is the rotation step. The sphering matrix is estimated from the time representation of the mixture signals but the rotation matrix is estimated in the wavelet domain of the whitened mixture-signal. The conventional sphering method is used while the rotation method is achieved using the QR factorization in the wavelet domain. After decomposition of the pre-whitened signals into subbands using the wavelet packet transform, the QR factorization of the cross correlation matrices of all subbands is used to estimate the rotation matrix and consequently estimating the separation matrix. The proposed method has been tested with several types of signals. The results indicate that this method is promising for the signal separation problem of speech, image, and noisy signals.