Few-Shot Underwater Acoustic Target Recognition Using Domain Adaptation and Knowledge Distillation

The complex dynamics of the marine environment pose substantial challenges for underwater acoustic target recognition (UATR) systems, especially when there are limited training samples. However, existing image-based few-shot learning methods might not be applicable, mainly because they fail to capture the temporal and spectral features from acoustic targets and lack the competent domain adaptation ability due to the inefficient usage of base samples. In this article, we develop a novel Domain Adaptation-based Attentional Time-Frequency few-shot recognition method (DAATF) for underwater acoustic targets. The DAATF explicitly utilizes a self-attention-based feature extractor to capture the time-frequency structural dependencies and constructs an autoencoder-based domain adapter to improve the cross-domain knowledge transfer through reusing the base dataset. In addition, a knowledge distillation module is designed to enable the model to reserve the general feature extraction ability of the pretrained network to avoid overfitting. Extensive experiments are conducted to assess prediction accuracy, noise robustness, and cross-domain adaptation. The obtained results validate that the DAATF can achieve outstanding performance, demonstrating its great potential for practical UATR applications. Furthermore, we provide free and open access to the DanShip data set.