Detection Method for Echolocation Clicks Based on LSTM Networks

Abstract

Odontocete clicks are a kind of short-time echolocation signal with high frequency. Research on the detection method of clicks is helpful for the accurate detection of marine mammal vocalization, so as to better protection of marine mammals. A method based on image processing is proposed to detect odontocete echolocation clicks. The collected data are divided into fixed-length frames and generate spectrograms. The spectrograms are filtered to remove noise and enhance the line-shape clicks. Considering that the echolocation signals are like lines in time-frequency domain, line detection is subsequently used to obtain the precise position of the lines. Finally, a Long Short-Term Memory network was trained to obtain a detector to distinguish clicks. The performance of the proposed method was evaluated using real audio recordings. The experimental results indicate that comparing with the traditional energy detector method, the proposed algorithm shows higher recall and precise under low signal-to-noise ratio (SNR). The proposed method can provide technical support in odontocete survey to accurately determine the species and better for marine bioacoustics study.