Sonar Foundational Model for Representation Learning and Automatic Target Recognition Systems in Underwater Maritime Environment

与深度学习在其他应用领域的成功相比，深度学习的进步尚未导致使用被动声纳的水下自动目标识别的实际革命。这至少部分是因为声纳数据的多样性和复杂性与相对有限的数据集相结合，尚未导致深度学习系统在一系列水下环境中可靠地优于传统的信号处理方法。本项目旨在通过开发被动声纳的基础模型，实现被动声纳ATR的突破。首先，它将聚合现有的声纳数据集，以训练更通用和强大的声纳表示。其次，它将研究这种表示如何支持ATR，跟踪等数据高效学习以及域概括技术，以保证在不同的海洋条件和水听器类型的概括。

Advances in deep learning have not yet led to a practical revolution in underwater Automatic Target Recognition using passive sonar, in contrast to deep learning's success in other application domains. This is at least in part because the diversity and complexity of sonar data combined with relatively limited datasets has not yet led to deep learning systems that outperform traditional signal processing approaches reliably across a range of underwater environments. This project aims to achieve a breakthrough in passive sonar ATR by developing a foundational model for passive sonar. First, it will aggregate existing sonar datasets to train a more general and robust sonar representation. Secondly, it will study how this representation can support data-efficient learning for ATR, tracking, etc as well as domain generalisation techniques to guarantee generalisation across different oceanic conditions and hydrophone types.