An Anchor-Free Feature-Enhanced Network for Arbitrary Oriented Ship Detection in High Resolution SAR Images

Ship detection in synthetic aperture radar (SAR) images plays a crucial role in maritime transportation. However, SAR ship detection faces challenges, such as diverse ship scales, complex background environments, severe noise interference, and the difficulty of distinguishing densely docked ships, all of which hinder the precision of ship target detection. To address these issues, an anchor-free feature-enhanced network for arbitrary-oriented ship detection (AFFE-AOSD) is proposed, which significantly enhances the detection performance for multiscale ship targets and densely docked ship formations, while also improving angle estimation accuracy. First, we designed an efficient feature integration module to enrich gradient flows, tailored specifically for the ship target detection task. This improves the detection of small targets. In addition, to enable more precise angle estimation, a two-stage angle prediction module with two stage circular smooth label is introduced. Furthermore, an angle-aware dynamic sample assignment method and an associated loss function that consider angles have been devised to improve the network’s accuracy in rotation predictions. Experiments on SSDD+ and RSDD datasets demonstrate significant improvements: AFFE-AOSD outperforms the baseline on SSDD+, improving precision by 3%, F1 score by 5.1%, and AP50 by 4.7%, and achieves the best angle prediction accuracy on RSDD with a mean error of 4.47° and a median error of 2°.Moreover, the proposed method consistently achieves real-time inference speeds exceeding 55 frame per second across different scenes, confirming its effectiveness in balancing accuracy and efficiency for practical maritime applications.