A Lightweight Wavelet Convolution-Hybrid Attention Network for Underwater Acoustic Target Recognition

Underwater Acoustic Target Recognition (UATR) is an important technology for underwater security and information acquisition. Due to various factors such as noise interference, the accuracy of ship radiated noise recognition is seriously limited, considering the cost, efficiency and real-time deployment of underwater recognition system, we propose a new lightweight wavelet convolution-hybrid attention network (WT-HAN). Aiming at the contradictory problems of the number of model parameters, computational complexity and recognition performance in UATR, we introduce wavelet convolution and design a new lightweight dual-domain attention module. By utilizing the characteristics of wavelet convolution with low number of parameters and large receptive field and the ability to enhance low-frequency feature response, the number of model parameters is greatly reduced while maintaining high accuracy. Furthermore, the combination of lightweight multiscale spatial attention and efficient channel attention (ECA) realizes the high accuracy enhancement with rather low parameter cost. The experimental results show that the proposed WT-HAN model can achieve much better recognition accuracies compared with existing mainstream models. Benefiting from its extremely lightweight architecture, the proposed model can be efficiently deployed on resource-constrained edge devices and underwater embedded platforms, providing strong technical support for real-time underwater perception tasks.