Number of components and initialization in Gaussian mixture model for pattern recognition

0410563 - UTIA-B 20010032 RIV AT eng C - Conference Paper (international conference)
Paclík, P. - Novovičová, Jana
Number of components and initialization in Gaussian mixture model for pattern recognition.
Wien: Springer, 2001. ISBN 3-211-83651-9. In: Artificial Neural Nets and Genetic Algorithms. Proceedings. - (Kůrková, J.; Neruda, R.; Kárný, M.; Steele, N.), s. 406-409
[International Conference on Artificial Neural Nets and Genetic Algorithms /5./. Prague (CZ), 22.04.2001-25.04.2001]
R&D Projects: GA MŠMT VS96063; GA AV ČR KSK1075601
Institutional research plan: AV0Z1075907
Keywords : pattern recognition * Gaussian mixture model * kernel density estimate
Subject RIV: BB - Applied Statistics, Operational Research

The method for complete mixture initialization based on a product kernel estimate of probability density function is proposed for mixture estimation using EM-algorithm. The mixture components are assumed to correspond to local maxima of optimaly smoothed kernel density estimate. The gradient method is used for local extrema finding. As the last step, agglomerative hiearchical clustering methods merges closest components together. A comparison to scale-space approaches is given on examples.
Permanent Link: http://hdl.handle.net/11104/0130652