A PDF-MATCHED SHORT-TERM LINEAR PREDICTABILITY APPROACH TO BLIND SOURCE SEPARATION

This paper presents a PDF-matched measure of short-term linear predictabil- ity of signal as a merit function for blind sources separation (BSS) of linearly mixed signals/images. BSS based on the proposed merit function (PSLP-BSS) achieves both high separation performance and low computational complexity. Despite the conventional predictability measurement, the proposed one does not need any long-term predictor, and it achieves 50.5 times less computational complexity by utilizing only short-term linear (STL) predictors. Furthermore, PSLP-BSS not only recovers signals with maximized predictability, but also increases non-gaussianity that concludes more independent recov- ered signals. This is because the coefficients of STL predictors are chosen in an objective probabilistic algorithm based on an assumed high kurtosis PDF for source signals. Source signals are recovered bynding an un-mixing matrix that maximizes the proposed mea- sure of short-term predictability for each extracted signal. The un-mixing matrix can be obtained as the solution to a generalized eigenvalue problem, and signals can be extracted simultaneously using the fast eigenvalue routine. The dominance of PSLP-BSS to con- ventional one has been demonstrated by many tests performed over articial mixtures of audio signals (music and speech) and articial mixtures of gray-scale images. Keywords: Blind source separation, Gaussianity, Gradient ascent, Eigenvalue routine, Kurtosis, Predictability, Short term linear (STL) predictors