Local discriminant time-frequency atoms for signal classification

Three methods to select discriminant time-frequency atoms from the Gabor ti me-frequency dictionary are proposed. The first method performs a discrimin ant pursuit in its selection, the second method leads to atoms with the mos t discriminant power, and the third method combines the first two methods. The time-frequency atoms selected by the methods extract discriminant featu res among different classes of signals. Experimental results on the classif ication of simulated data sets (triangular waveforms) and real data sets (s peech signals) using the extracted features are presented. (C) 1999 Elsevie r Science B.V. All rights reserved.