A joint artificial and semantic feature space mixup for deep-model-based passive underwater acoustic multi-target recognition

Due to the limitation of sonar beam resolution, the passive radiated noises from different targets may overlap with each other within the same beam, which gives rise to the multi-target passive recognition problem. This problem is typically addressed by optimizing a multi-label loss function within the multi-target data. Furthermore, the complexity of the ocean acoustic environment results in significant intra-class diversity, which is particularly pronounced in a limited set of data, leading to severe distribution shifts. Additionally, nonlinear interactions occur when radiated noise from multiple targets propagates through underwater acoustic channels. To address these challenges, this article proposes a joint artificial and semantic feature space mixup for underwater acoustic multi-target recognition. Specifically, besides just mixing multiple targets in the original signal space, this study also mixes them in the canonical artificial feature space (e.g., Mel and short-time Fourier transform spectrograms) and semantic feature space (i.e., hidden features from deep models). By constructing multi-target data across different spaces, the study attempts to guide the deep network to learn the potential diversity of multiple targets with limited data. The study has also derived theoretical proofs for the rationality of this method in mitigating the impact of distribution shifts and nonlinear interactions. Extensive experiments demonstrate the consistent efficacy of this method when incorporating different deep models and artificial features.