Triangle Distance IoU Loss, Attention-Weighted Feature Pyramid Network, and Rotated-SARShip Dataset for Arbitrary-Oriented SAR Ship Detection

Abstract

In synthetic aperture radar (SAR) images, ship targets are characterized by varying scales, large aspect ratios, dense arrangements, and arbitrary orientations. Current horizontal and rotation detectors fail to accurately recognize and locate ships due to the limitations of loss function, network structure, and training data. To overcome the challenge, we propose a unified framework combining triangle distance IoU loss (TDIoU loss), an attention-weighted feature pyramid network (AW-FPN), and a Rotated-SARShip dataset (RSSD) for arbitrary-oriented SAR ship detection. First, we propose a TDIoU loss as an effective solution to the loss-metric inconsistency and boundary discontinuity in rotated bounding box regression. Unlike recently released approximate rotational IoU losses, we derive a differentiable rotational IoU algorithm to enable back-propagation of the IoU loss layer, and we design a novel penalty term based on triangle distance to generate a more precise bounding box while accelerating convergence. Secondly, considering the shortage of feature fusion networks in connection pathways and fusion methods, AW-FPN combines multiple skip-scale connections and attention-weighted feature fusion (AWF) mechanism, enabling high-quality semantic interactions and soft feature selections between features of different resolutions and scales. Finally, to address the limitations of existing SAR ship datasets, such as insufficient samples, small image sizes, and improper annotations, we construct a challenging RSSD to facilitate research on rotated ship detection in complex SAR scenes. As a plug-and-play scheme, our TDIoU loss and AW-FPN can be easily embedded into existing rotation detectors with stable performance improvements. Experiments show that our approach achieves 89.18% and 95.16% AP on two SAR image datasets, RSSD and SSDD, respectively, and 90.71% AP on the aerial image dataset, HRSC2016, significantly outperforming the state-of-the-art methods.