OpenSARWake: A Large-Scale SAR Dataset for Ship Wake Recognition With a Feature Refinement Oriented Detector
Chengji Xu, Xiaoqing Wang
Abstract
Synthetic aperture radar (SAR) is used to persistently monitor marine areas in all weather conditions for excellent ship and wake identification. Current deep learning-based ship wake detection methods rely on supervised learning. However, no publicly available large-scale SAR dataset is available to support this learning method. Owing to the diversity of ship wake characteristics in SAR images and the complexities of sea states, contemporary computer vision algorithms are generally ineffective for SAR image recognition tasks. To overcome these issues, this article presents the new, well-annotated <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">OpenSARWake</i> dataset dedicated to oriented ship wake detection. This collection provides 3,973 images containing two polarization modes and 4,096 instances. Their image features are used to train a novel two-stage SWNet feature refinement detector that adopts a sophisticated HR-FPN* backbone for SAR ship wake detection. The detector recognizes nearly all wake characterizations. The trained SWNet achieves state-of-the-art detection performance with 49.0% mAP, outperforming most benchmarking algorithms. The <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">OpenSARWake</i> dataset is available at https://github.com/libzzluo/OpenSARWake.