Show simple item record

dc.contributor.authorLiu, Xianhua
dc.contributor.authorRandall, R.
dc.contributor.editorHans-A. Bachor and Massimiliano Colla
dc.identifier.citationLiu, Xianhua and Randall, R.B. 2005. A New Independent Component Analysis Algorithm: Joint Approximate Diagonalization of Simplified Cumulant Matrices, in Bachor H.A. and Massimiliano, C. (ed), Proceedings of The 16th Biennial Congress of the Australian Institute of Physics - "Physics for the Nation", Jan 31-Feb 4 2005, pp. 29-32. Canberra, Australia: Australian Institute of Physics.

This paper proposes a new algorithm to improve robustness, reliability and efficiency for blind signal separation with a different diagonal cumulant maximization criterion. It calculates a fraction of the fourth order cumulant set and avoids the eigenmatrix decomposition to considerably reduce the separation cost for large-scale problems. Experimental separation shows that the new algorithm is robust, reliable and efficient for both large and small-scale separation problems, thus has combined merits of the well-known JADE and Fast ICA algorithms. Mixed music and speech signal separation is presented in this paper.

dc.publisherAustralian Institute of Physics
dc.subjectIndependent component analysis
dc.subjectBlind source separation
dc.subjectfourth order cumulant matrix
dc.titleA new independent component analysis algorithm: Joint approximate diagonalization of simplified cumulant matrices
dc.typeConference Paper
dcterms.source.titleProceedings of the 16th National Congress of Australian Institute of Physics
dcterms.source.seriesProceedings of the 16th National Congress of Australian Institute of Physics
dcterms.source.conferenceThe 16th National Congress of Australian Institute of Physics 2005
dcterms.source.conference-start-dateJan 31 2005
dcterms.source.conferencelocationCanberra, Australia
dcterms.source.placeCanberra, Australia
curtin.accessStatusFulltext not available

Files in this item


This item appears in the following Collection(s)

Show simple item record