Abstract: In remote sensing images object detection tasks based on deep learning, the ship usually exhibit features arranged in any direction. The common algorithms of object detection adopt horizontal detection that generally cannot meet the application requirements of such scenarios. Therefore, this paper adds a rotation angle prediction branch to the single-stage Anchor-Free object detector CenterNet, it can output a rotating bounding box for the detection of marine ship objects. At the same time, in view of the problem that maritime ship remote sensing data sets only have horizontal bounding box labels, which cannot be directly applied to rotating boxes object detection, and manual labeling of rotating boxes labels is expensive, an active and transfer learning method of rotating boxes label generation is proposed. Firstly, a horizontal box-rotating box constraint screening algorithm is proposed. The rotating prediction box is supervised and constrained by the horizontal ground truth bounding box. The image with higher detection accuracy is selected and added to the training set. Then this process is iterated to filter out more images. Finally, the automatic labeling of the rotating box of the data set is completed by matching the label categories. In this paper, about 65.59% of the pictures in the remote sensing image data set BDCI of marine ships are finally marked with a rotating boxes, and some unmarked pictures are manually marked as the test set. The pictures marked by the method in this paper are used as the training set for verification. The evaluation index AP50 reaches 90.41%, which is higher than other rotating boxes detectors, indicating the effectiveness of this method.

Key words: remote sensing images, object detector of rotated boxes, transfer learning, Anchor-Free, CenterNet