Blind Source Separation: A Performance Review Approach

Blind source separation (BSS) has been used extensively for the last three decades to extract different sources such as speech, biomedical, and image sources. The quality of extraction is governed by the extraction metrics of the BSS algorithms. The aim of this work is to review the extraction metrics for 27 different algorithms used in some previous works. Results show that the best algorithms in speech extraction is obtained when using independent low-rank matrix analysis (ILRMA) and canonical correlation (CC) based BSS. For image source separation, ZEODS algorithm has shown the highest extraction evaluation score. For fetal electrocardiogram (FEEG) source separation, the null space idempotent transformation matrix (NSITM) and the generalized functional link artificial neural network (GFLANN) show the highest evaluation scores. For electroencephalogram(EEG) signals de-noising and source extraction, the short and long term prediction blind source extraction (SLTP-BSE), B-spline mutual information independent component analysis (BMICA), and FASTICA, show the highest evaluation scores. The impact of evaluating the extraction algorithms result in optimum source separation.