Ship identification based on optical remote sensing images is crucial in both civil and military contexts, such as fisheries management, port surveillance, and marine security. However, the horizontal bounding boxes generated by general object detection methods often involve redundant information, for instance, background and multiple ships, making it difficult to accurately locate ship targets. To resolve this issue, we proposed a novel oriented ship fine-grained identification network, called SFINet. We applied an Oriented RPN to the ship fine-grained recognition task for the first time, Unlike rotating anchors with different pre-set angles, Oriented RPN can generate a set of oriented proposals for regression and classification directly from horizontal anchors. Additionally, due to the diversity of ship types and the difference in the number of instances in each category, we improved the softmax cross-entropy loss to alleviate the problem of imbalance between instances in different ship categories, without increasing extra costs. Extensive experiments show that our SFINet achieved state-of-the-art detection accuracy in two publicly available ship fine-grained datasets, HRSC2016 (64.04% mAP) and FGSD2021 (86.90% mAP).
SFINet: An Oriented Fine-Grained Ship Identification Network Based on Remote Sensing Image
Lect. Notes Electrical Eng.
International Conference on Image, Vision and Intelligent Systems ; 2023 ; Baoding, China August 16, 2023 - August 18, 2023
Proceedings of International Conference on Image, Vision and Intelligent Systems 2023 (ICIVIS 2023) ; Kapitel : 21 ; 206-215
25.02.2024
10 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch