A few-shot learning method for underwater acoustic target recognition based on generative data augmentation

The recognition of underwater acoustic targets (UATR) is of great significance for the protection of marine diversity and national defense security. The development of deep learning provides new opportunities for UATR, but faces challenges brought by the scarcity of reference samples and complex environmental interference. To tackle this problem, we propose a generative discriminative collaborative framework, a variational auto-encoder boosted learning framework based on latent space completion. Rooted in the core contradiction arising from the incompleteness of intra-class manifolds and the instability of discriminative boundaries, this framework incorporates the premise of latent space continuity. Leveraging a structure-preserving generative reconstruction mechanism, it implicitly supplements the original dataset, which in turn enables the reconstruction of intra-class distributions that are more continuous, integral, and discriminative at the feature level. In this paper, we construct a three-stage pipeline system consisting of auto-clean cut unified preprocessing, latent reconstruction variational auto-encoder multi-scale latent space reconstruction, and an acoustic identification model. Furthermore, by establishing a staged modeling workflow, data purification, latent space completion, and discriminative optimization converge on their individual objectives independently while maintaining overall synergy, thus forging a robust recognition paradigm tailored to few-shot learning scenarios.