A Hierarchical Underwater Acoustic Target Recognition Method Based on Transformer and Transfer Learning

Underwater acoustic target recognition (UATR) is one of the essential research directions in the underwater acoustic signal processing field. The machine learning-based recognition methods have solved some performance bottleneck problems of traditional recognition methods. One of the challenges faced now is the lack of trainable data. This paper uses the transfer learning (TL) strategy for UATR with limited training samples. We transfer the pre-trained Swin Transformer model on the ImageNet dataset to underwater acoustic signals. This paper presents a Hierarchical Underwater Acoustic Transformer (HUATrans) model for ship-radiated noise recognition. The Transformer architecture has better feature extraction capability than Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). The proposed model uses the Swin Transformer based on the shifted window self-attention mechanism instead of the traditional Transformer architecture and improves the patch embedding module. It reduces the training time and saves a lot of computational resources. At the same time, this paper transfers the Transformer architecture pre-trained in the image domain to the UATR domain, which alleviates the burden on model training due to a lack of training data. Experimental results show that the HUATrans model with the TL strategy achieves excellent recognition performance on ship-radiated noise datasets DeepShip and ShipsEar.