Computational Oriented Proposal for Fine-Grained Ship Detection in Complex Remote Sensing Images

Remote sensing images pose great challenges for fine-grained ship (SH) detection due to their bird's-eye view, complex scenes, and low resolution. One challenge of this problem is the difficulty of precisely localizing various SHs at the individual level. For SH detection, horizontal bounding boxes (HBBs) and oriented bounding boxes (OBBs) are the two common methods. However, when the remote sensing image contains SHs with arbitrary orientations and dense coverage, the former method is prone to poor localization and high detection errors, while the latter is inefficient precision in characterizing and adjusting the object orientation as well as high computational complexity. Furthermore, another challenge is the fine-grained nature of the SH categories and the difficulty of collecting training images that reflect realistic ocean scenarios. To address these challenges, this article proposes the computational oriented proposal (COP), where SHs are extracted from images, and their inherent features are used to calculate orientations and predict sizes to directly generate precise OBBs instead of relying on predefined anchors or other data-driven methods. To enhance SHs while suppressing background interference caused by complex backgrounds, a multiscale object enhancement module (MSOEM) is developed through multiscale attention. To further improve fine-grained SH classification accuracy, we propose an improved multitasking loss function with a contrastive embedding classification loss term, which can increase intraclass similarity and decrease interclass similarity. Moreover, a dataset called fine-grained remote sensing SH detection (FRSSD) is created. Experiments on the HRSC2016, FRSSD, and dataset for object detection in aerial images (DOTA) show that the proposed method outperforms other state-of-the-art object detection methods.