Underwater Acoustic Target Recognition Via Dualmodal Mixture-of-Experts Network: a Hierarchical Auditory-Visual Fusion Approach

The development of underwater acoustic target recognition (UATR) technology holds both theoretical and practical significance in advancing detection capabilities for underwater early-warning systems and protecting maritime strategic interests. Current approaches predominantly use timefrequency spectral features integrated with image classification networks, but their recognition accuracy remains constrained by three interrelated challenges: the limited availability of underwater acoustic data, compounded by complex environmental noise interference and diminished target-radiated noise levels, which collectively necessitate substantial reliance on expert knowledge throughout the recognition pipeline. To address the challenge of low recognition accuracy in ship-radiated noise scenarios with limited data samples, the paper propose a DualModal Mixture of Experts model (D-M MoE) that integrates auditory and visual modalities. This framework combines pretrained audio and visual foundation models to perform composite analysis of waveform signals and time-frequency spectrograms, leveraging the powerful feature representation capabilities of selfsupervised learning. Experimental validation on the ShipsEar dataset demonstrates that our method significantly improves performance metrics for underwater acoustic target recognition tasks.