Application of Deep Object Detection in Underwater Acoustic Pulse Interception

Abstract

Underwater acoustic pulse interception is an important task for underwater signal processing system, including the detection and identification of unknown acoustic pulses. An acoustic pulse interception method based on deep learning is proposed. The interception system consists of a pulse detection network and a DOA estimation network. The pulse detection neural network is used to achieve multi-pulse detection and bounding box inference on the spectrogram. The phase component of the short-time Fourier transform coefficients in the time-frequency bounding box is extracted. Then the DOA estimation network learns the phase feature to figure out the direction of arrival of each detected pulse by regression. Finally, the number of sources and their DOA estimates could be obtained through such operations as outlier removal and data fusion. Simulation results show that this method is able to achieve reliable pulse detection, source number estimation and high precision DOA estimation in underwater acoustic environment.