PEDNet: A Proposal Enhancement Dynamic Network for Fine-Grained Ship Detection in Optical Remote Sensing Images

Abstract

Fine-grained ship detection (FGSD) refers to identify and localize different ships in optical remote sensing images, which is utilized in both civilian and military areas widely. Nevertheless, the region proposal as an essential component of ship detection and recognition, is restricted to design a multitask learning of FGSD, from proposal generation, representation, and utilization. In this paper, we propose a proposal enhancement dynamic network (PEDNet), to deal with the subtasks of FGSD better. First, the lateral scale-equalizing feature pyramid network (LSE-FPN) is designed to deliver equalized scale features for the subsequent generation of more precise proposals. Second, the multi-channel fusion network (MCFN) is proposed to extract independent and discriminative feature representations in proposals. In addition, we present an oriented dynamic prediction network (ODPN), which employs the dynamic training strategy to provide guidance for model training, to leverage high-quality proposals effectively for ship identification and localization. Finally, PEDNet is trained and validated on ShipRSImageNet, FGSD2021, DOSR, and MAR20 datasets. Quantitative and qualitative experimental results reveal that, compared with cutting-edge detectors, AP50 of the PEDNet received by 74.58, 92.50, 61.28, and 80.84, AP75 received by 68.34, 86.26, 46.19, and 69.18, AP $_{\mathrm {50:95}}$ received by 56.25, 68.78, 40.17, and 55.54 on four datasets, respectively. It effectively proves that our method achieves advanced FGSD effectiveness and favorable generalization performance.

Keywords

• Computer vision
• object detection
• fine-grained ship detection (FGSD)
• deep learning
• remote sensing images