Ship Classification Using Swin Transformer for Surveillance on Shore

Abstract

Ship image classification technology is one of the core technologies for intelligent maritime surveillance system. It is fundamental that ships and their types are accurately identified for analysing and understanding in maritime scenes. Recently, the transformer-based model successfully applied in the field of natural language processing, and they have surpassed convolutional neural networks in image classification tasks, with Swin Transformer as the leader. Swin Transformer builds a hierarchical pyramid structure and a shifted window scheme on the basis of multi-head self-attention mechanism. These qualities reduce the complexity of models, and makes it as a general backbone for computer vision. In this study, we use the well-known ship image dataset called Seaships to investigate the effectiveness of Swin Transformer. We find that its hierarchical pyramid structure, multi-head self-attention mechanism and shifted window scheme play a key role in ship image classification. The results show that Swin Transformer achieves an accuracy of 93.5% in ship image classification, and outperforms typical convolutional networks and Vision Transformer.