Ellipse Polar Encoding for Oriented SAR Ship Detection
Jun Liu, Lin Liu, Jiarong Xiao
Abstract
Ship detection in synthetic aperture radar (SAR) images using deep neural networks often relies on horizontal bounding box (HBB), which fail to capture ship orientation and aspect ratio accurately. To address this limitation, oriented ship detection methods based on oriented bounding box (OBB) are gaining attention. However, most available existing OBB methods suffer from boundary discontinuity problem, leading to convergence problems and unsatisfied orientation detection performance. In this paper, we propose an innovative oriented SAR ship detection method using ellipse polar encoding (EPE). By representing the ship detection box as an ellipse and employing a set of vectors from the center to the boundary as encoded parameters, our method exhibits smooth variations, enhances convergence, and efficiently decodes into OBB. We further develop a lightweight and effective oriented SAR ship detection network based on this methodology. To account for SAR image characteristics, such as speckle and deformation causing deviations from true ellipses, we introduce an intersection over union (IOU) weighted EPE loss. Experimental results on the Rotated Ship Detection Dataset in SAR images (RSDD-SAR) demonstrate the effectiveness of our proposed method in significantly improving detection performance compared to other oriented target detection methods.