UATR-DIFF-Transformer: Enhanced Differential Optimization for Underwater Acoustic Target Recognition

Accurate recognition of underwater acoustic targets (UATR) using radiated noise is critical for numerous applications in underwater acoustics, including surveillance, navigation, and marine research. In recent years, deep learning has garnered significant attention in underwater passive target recognition (UATR). However, current methods exhibit limitations in effectively leveraging critical spectral-temporal information from complex underwater signals. To address this, we propose UATR-DIFF-Transformer, an architecture integrating a differential attention mechanism with polynomial loss optimization. The differential attention mechanism captures relative variations in spatial and frequency domains, enabling the model to prioritize discriminative features and achieve enhanced noise-robust characteristics. Simultaneously, the Poly Loss function enhances the model's ability to achieve superior intra-class compactness and inter-class separability. Experimental results on the publicly available Shipsear and Deepship datasets demonstrate that our proposed UATR-DIFF-Transformer framework achieves performance gains in classification accuracy. Compared to traditional Transformer models, the proposed architecture exhibits superior robustness to signal-to-noise ratio (SNR) variations, with its performance degradation under fluctuating SNR conditions being markedly mitigated. Additionally, by optimizing computational efficiency, our model achieves higher computational efficiency, with reduced FLOPs and parameter counts. These results high-light the potential of UATR-DIFF-Transformer in advancing the development of advanced UATR systems and provide a foundation for future research in underwater acoustic signal processing.