Ship Detection based on SAR Imaging Supervised by Noisy Ship Direction

Synthetic aperture radar (SAR) is an active imaging microwave sensor. Since SAR imaging is robust against weather, light, and other external conditions, Ship Detection based on SAR Imaging has high research and application values. However, the variety in size, direction, background and arrangement poses great challenges to SAR image ship detection. Due to the lack of effective constrains, plenty of boxes predicted by the existing methods are quite different from groundtruth, resulting in IOU between predicted boxes and groundtruth being smaller than the threshold and generates false alarms. Therefore, in this letter we propose a method of noisy ship direction classification, which uses the ship pixel distribution in SAR images to classify ship objects into different directions, so as to reduce the shape difference between boxes and groundtruth. Experimental results show that by combining with YOLOv5 framework, our method can significantly improve the ship's detection performance, and reach state-of-the-art.