ShadowFPN-YOLO: A Real-Time NMS-Free Detector for Remote Sensing Ship Detection

The deployment of advanced satellites has significantly enhanced computer vision applications in remote sensing. Among these, ship detection plays a vital role in ensuring maritime safety and managing ship traffic. However, the overhead perspective in remote sensing images often results in ship bounding boxes containing substantial background or overlapping ships, which can obscure key features necessary for accurate detection. To address these challenges, this paper proposes a real-time detector, ShadowFPN-YOLO, featuring the novel ShadowFPN module. ShadowFPN dynamically applies shadows to multi-scale feature maps by randomly masking contiguous regions, effectively reducing background interference and enhancing the network’s focus on the essential features of target ships. Furthermore, the backbone network integrates generalized efficient layer aggregation network with reparameterized cross stage partial modules to further enhance performance. Additionally, the adoption of a dual-head strategy eliminates the dependency on non-maximum suppression (NMS) for managing overlapping boxes, offering a more robust and accurate solution for object detection in densely packed scenarios. Experimental results demonstrate the strong performance of our method, achieving mAP values of 55.70 at 430 FPS on the DIOR-Ship dataset and 55.85 at 497 FPS on the HRSID dataset, all while maintaining the fewest parameters and the lowest computational cost compared to the latest YOLO models. These results underscore the efficiency and effectiveness of our approach for real-time remote sensing ship detection.