A Lightweight Fusion Model Based Jamba and Cross Fusion Mixture of Experts for Underwater Acoustic Target Recognition

To alleviate the problems of low recognition accuracy and weak robustness in intelligent underwater acoustic target recognition systems, a lightweight fusion model based the time and frequency attention module, MobileJiT and cross fusion mixture of experts is proposed. based on Jamba module sequence sensitivity and MobileViT network lightweight, we constructed MobileJiT module as the backbone network to extract global and dependency features. We subsequently adopt a time and frequency attention module as the branch network to extract local features. Finally, we adopt a cross fusion mixture of experts module to fuse the features of the backbone network and branch network, and input the fused rich features into mixture of experts for robust recognition. Experimental results indicate that the proposed lightweight model achieves a balance between recognition accuracy and robustness, achieving recognition accuracy of 97.63% and 97.5% on two public datasets, surpassing state-of-the-art models by 0.63% to 2.94%.