A Deep Neural Network for Counting Vessels in Sonar Signals

Monitoring the oceanographic activity of ships in restricted areas is an important task that can be done using sonar signals. To this end, a human expert may regularly analyze passive sonar signals to count the number of vessels in the region. To automate this process, we propose a deep neural network for counting the number of vessels using sonar signals. Our model is different from common approaches for acoustic signal processing in the sense that it has a rectangular receptive field and utilizes temporal feature integration to perform this task. Moreover, we create a dataset including 117K samples where each sample resembles a scenario with at most 3 vessels. Our results show that the proposed network outperforms traditional methods substantially and classifies 96% of test samples correctly. Also, we extensively analyze the behavior of our network through various experiments. Our codes and the database are available at https://gitlab.com/haghdam/deep_vessel_counting. © Springer Nature Switzerland AG 2020.