Statistical segmentation of spectrograms: a complementary approach to track detection algorithms - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2013

Statistical segmentation of spectrograms: a complementary approach to track detection algorithms

Florian Dadouchi
  • Fonction : Auteur
  • PersonId : 947124
Cornel Ioana
Jerome I. Mars

Résumé

An efficient statistical method for the segmentation of spectrograms of narrow band bioacoustics sounds is proposed. The method is not an alternative but a complement to standard time-frequency trackers (particle filters, Viterbi-based tracking, Kalman filtering, graph search, time-frequency-phase tracker, etc.) and operates as a pre-processing step: 1. Identification of regions of interest (ROI) in long recordings; 2. Generation of efficient binary spectrograms; 3. Design of time-varying filters to enhance the signal to noise ratio. The approach is designed to deal with real world constraints: large databases, non-stationary/colored noise and multi-component non linear frequency modulations. The method performs a double binary hypothesis test, one on time-frequency bins, the second on regions of the time-frequency plane. An overview of the author's methodology is the following: First step: The background noise in the spectrogram is modeled as a chi-squared distribution with two degrees of freedom and its parameter is estimated for each time-frequency bin using a minimal statistics approach (small values are realizations of noise only). An adaptive threshold is then computed using a constant false a larm probability (pfa) for thresholding (Neyman-Pearson (NP)). Second step: The second binary hypothesis test applies on regions of the binary spectrogram. The histogram of the number of detections in regions of the binary spectrogram is computed. The probability distribution of the number of false detections is modeled as a Binomial distribution which parameter is estimated. Regions are decided as "hosting signal" or "not hosting signal" using a NP approach. To enhance the robustness of the overall methodology, a multi-thresholding approach is applied. The optimal pfa at step 1 is unknown, N binary spectrograms are then generated. Step 2 solves the binary hypothesis test for each of the N binary spectrograms using a unique false alarm prob ability and selects the one maximizing the number of detections.
Fichier non déposé

Dates et versions

hal-00877149 , version 1 (26-10-2013)

Identifiants

  • HAL Id : hal-00877149 , version 1

Citer

Florian Dadouchi, Cedric Gervaise, Cornel Ioana, Julien Huillery, Jerome I. Mars. Statistical segmentation of spectrograms: a complementary approach to track detection algorithms. DCLDE 2013 - 6th International Workshop on Detection, Classification, Localization, and Density Estimation of Marine Mammals using Passive Acoustics, Jun 2013, St Andrews, United Kingdom. p19. ⟨hal-00877149⟩
331 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More