Object detection is a favourite research technology in the field of maritime computer vision and current research mostly adopts region models based on convolutional neural networks (CNNs). Swin-transformer brings a new and effective approach, with its hierarchical structure and attention mechanism-based block, well compensating for the weakness of CNNs that cannot take into account global features. In this paper, we propose a method for applying the Swin-Transformer to maritime object detection, obtaining significantly improved results on the well-known Seaships dataset. Experiments demonstrate that Swin-Transformer, after pre-trained and fine-tuning, has a detection mean average precision of 96.61%, outperforming other CNNs. Furthermore, Swin-Transformer has demonstrated its capability to detect features of the irregular ship shapes in complex backgrounds, and the detection of general cargo ships and passenger ships is excellent.
Object Detection in Maritime Scenarios Based on Swin-Transformer
Lect. Notes Electrical Eng.
International Conference on Computing, Control and Industrial Engineering ; 2021 ; Hangzhou, China October 16, 2021 - October 17, 2021
6th International Technical Conference on Advances in Computing, Control and Industrial Engineering (CCIE 2021) ; Chapter : 77 ; 786-798
2022-07-06
13 pages
Article/Chapter (Book)
Electronic Resource
English