Metric learning and anomaly detection for passive sonar

Passive sonar is used to detect, localise, and classify ships. There is a great interest in automatic classification, as a complement to manual classification performed by a sonar operator. In recent years, there have been rapid developments in deep learning, which can be useful for these applications. These methods often require large amounts of training data. In practice, data for interesting targets can be scarce or non-existent. Therefore, it is of interest to investigate few-shot methods, which can learn to recognise new sound sources from small amounts of reference data, as well as methods which can determine if a sound deviates from the normal picture. In both cases, large amounts of training data are required, where the data do not necessarily need to include interesting targets. In this work, a large dataset has been created, including over 250 hours of single-hydrophone recordings for 794 unique ships. The dataset has been used for evaluation of methods for metric learning and anomaly detection. Metric learning is a method for performing few-shot classification and is also useful as a pre-processing step for anomaly detection. Several network architectures and anomaly detection algorithms have been tested. The results show that several methods are promising.