ARS-Det: an axis-based anchor-free rotation detector for ship in remote sensing images
Lei Zhang, Yuehuang Wang, Wen Chen · 2021
Ship detection in remote sensing images is a challenging and important task. Many methods of deep learning have been widely used in this field in recent years and most of them rely on a lot of predefined rotated anchor boxes, which not only lead to inaccurate angle predictions but also introduce excessive hyper parameters and high computational cost. In this paper, we propose an anchor-free rotation ship detector (ARS-Det) using axis-based representation scheme to address these issues. Our ARS-Det regards the detection of arbitrary-oriented objects as predicting the points at both ends of the longest axis and width of the objects. Attention FPN model is designed to capture multi-scale and key features of arbitrary-oriented target. Then, points assignment and orientation center-ness computation are proposed to screen of positive samples accurately. Experiments on public ship dataset HRSC2016 show that our method achieves state-of-art performance on ship detection in remote sensing images.