Low complexity method for blind source extraction for stationary mixed kurtosis sign signals

Abstract

This dissertation concerns the problem of how to make the extracting algorithm run in real time and how to reduce the computational complexity for the blind source separation. Our approach is to partition the observed signals into several pieces and to extract the partitioned observations with our proposed approximation activation function performing only the "shift-and-add" operation on the VLSI level. No division and exponential multiplication are needed. Moreover, an optimal initial demixing weight for speeding-up the separating time will be presented. The proposed algorithm is tested on the benchmarks available at http://speech.kaist.ac.kr/~jangbal/ch1bss. The experimental results signify that our solution provides a comparable efficiency as those of other approaches but lower in space and time complexity.