Knowledge-inspired and sample-generation-based spectral augmentation for few-shot underwater acoustic target recognition

To address the challenges of data scarcity and non-stationary noise interference in underwater acoustic target recognition in complex marine environments, existing few-shot learning methods often lack effective mechanisms to exploit structured spectrotemporal characteristics under low-signal-to-noise ratio conditions. This paper proposes a sample-generation-based few-shot underwater acoustic target recognition framework with a knowledge-inspired spectral decomposition augmentation strategy and a cross-feature guided recalibration module. The spectral decomposition augmentation module operates in the spatial-frequency domain of spectrograms to construct complementary multi-view samples by preserving coarse spectrotemporal structures while suppressing noise-sensitive fine texture variations. A Siamese network with time-frequency attention is employed to learn robust spectrotemporal representations, while the cross-feature guided recalibration module uses original-view features to perform dynamic gated recalibration on augmented views, thereby reducing sensitivity to noise-induced variations. Multi-view joint learning based on cosine similarity encourages the learning of shared discriminative representations across views. In addition, a task-adaptive fine-tuning mechanism with pseudo-sample generation is introduced to improve adaptation to novel classes. Experimental results on the ShipsEar dataset and the cross-domain DanShip dataset demonstrate that the proposed method achieves superior recognition accuracy and robustness under low-signal-to-noise ratio and cross-domain conditions compared with existing state-of-the-art few-shot learning methods.