Extraction of pertinent subsets from time-frequency representations for detection and recognition purposes

A time–frequency representation can highlight non-stationarities in a signal. We propose to extract subsets from the time– frequency representation (TFR) for classi3cation or recognition purposes. We developed two approaches. The 3rst one is developed for TFRs obtained from the short time Fourier transform or the gliding minimum variance method. The extraction of compact subsets is viewed as a segmentation of the TFR, which is performed by morphological 3ltering and watershed segmentation. The second approach is developed when the TFR has been obtained using parametric estimators. We consider a hybrid estimator, the ARCAP method, and use a Kalman 3lter trajectory tracker to extract spectral lines. The proposed methods are illustrated by examples on natural signals: dolphin whistle acoustical signals, cavitation signals and seismic signals produced by snow avalanches. ? 2002 Elsevier Science B.V. All rights reserved.