A Lightweight CNN for Noise Source Detection in Bearing-Time Sonar Images

In current broadband passive sonar systems, a human operator observes a series of sonar measurements that represents different energy levels in the bearing and time dimensions. Detecting weak noise sources amidst ambient sea noise can be challenging. This paper proposes a lightweight neural network based on U-Net that detects the presence of low SNR noise sources in a multipath setting. The architecture we designed predict only the direct path. Then, a linear detector can detect efficiently the presence of the acoustic source from the enhanced direct path.