hal-02088220
https://univ-rennes.hal.science/hal-02088220
doi:10.1109/ACSSC.2018.8645115
[UNIV-RENNES1] Université de Rennes 1
[LTSI] Laboratoire Traitement du Signal et de l'Image
[STATS-UR1] Statistiques-HAL-UR1
[UR1-HAL] Publications labos UR1 dans HAL-Rennes 1
[UR1-MATH-STIC] UR1 - publications Maths-STIC
[TEST-UR-CSS] TEST Université de Rennes CSS
[UNIV-RENNES] Université de Rennes
[UR1-MATH-NUM] Pôle Rennes 1 - Mathématiques - Numérique
[UR1-BIO-SA] Pôle Rennes 1 - Biologie-Santé
Independent Component Analysis Based on Non-polynomial Approximation of Negentropy Application to MRS Source Separation
Saleh, M.
Karfoul, Ahmad
Kachenoura, A.
Albera, Laurent
Senhadji, Lotfi
[SDV.IB] Life Sciences [q-bio]/Bioengineering
COMM
Approximation algorithms
Cost functions
Independent component analysis
Iterative methods
Magnetic resonance
Magnetic resonance spectroscopy
Polynomial approximation
Probability density function
Separation
Asymmetric probability densities
FastICA algorithms
Gradient ascent
ICA algorithms
Nonlinear objective functions
Number of iterations
Numerical results
Quasi-optimal
Source separation
In this paper, a new ICA algorithm based on non-polynomial approximation of negentropy that captures both the asymmetry of the sources' PDF and the sub/super-Gaussianity of this latter is proposed. A gradient-ascent iteration with quasi-optimal stepsize is used to maximize the considered cost function. With this quasi-optimal computation in the case of highly non-linear objective function, the main advantages of the proposed algorithm are 1) its robustness to outliers compared to kurtosis-based ICA method especially for situations of small data size, and 2) its ability to capture sources' asymmetric probability density functions which is a property that can't be fulfilled in classic ICA algorithms like FastICA. Numerical results reported in the context of source separation of brain magnetic resonance spectroscopy show the superiority of the proposed algorithm over the FastICA algorithm in terms of both source separation accuracy and the number of iterations required for convergence. © 2018 IEEE.
2018-10-28
en
IEEE Computer Society
52nd Asilomar Conference on Signals, Systems and Computers, ACSSC 2018
Pacific Grove, United States