Chi-square law mixture for an unsupervised time-frequency pattern extraction - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2003

Chi-square law mixture for an unsupervised time-frequency pattern extraction

Résumé

Our purpose is to develop an unsupervised method of extraction and characterization of time-frequency (TF) patterns. The new idea is the notion of probability model in the time-frequency representation. Each TF point is considered as a random variable having a density distribution, which depends of its belonging class and of the TF estimator. We consider square modulus based estimates such as spectrogram and scalogram, which induces chi-square mixture densities. Density parameters are estimated by Maximum Likelihood. The classification is managed by a region-growing algorithm in a feature space built from moments evaluated in local cells. We here propose a new feature based on entropy we must consider in the continuous case. We derive the analytical expressions of the generalized Rényi entropy for a non-central chi-square distribution. We therefore can represent what we refer to as an entropy map where entropy is calculated from local histogram evaluated in gliding time-frequency cells. We present both feature space segmentation and entropy map on real signals: seismic signals (avalanches and seisms), acoustical signals (cavitation and concert halls), bioacoustical signals. Application of the proposed algorithm is also possible on other representation than time-frequency representation.
Fichier principal
Vignette du fichier
manuscript_N_Martin.pdf (466.27 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)
Loading...

Dates et versions

hal-01737186 , version 1 (19-03-2018)

Identifiants

  • HAL Id : hal-01737186 , version 1

Citer

Nadine Martin, Cyril Hory, Mathieu Huchard. Chi-square law mixture for an unsupervised time-frequency pattern extraction. Tenth International Congress on Sound and Vibration, Jul 2003, Stockholm, Sweden. ⟨hal-01737186⟩

Collections

UGA CNRS LIS
52 Consultations
37 Téléchargements

Partager

Gmail Facebook X LinkedIn More