Ship Detection in SAR Images Based on Oriented Bounding Box and Supervised Contrastive Learning

Ship detection in SAR images is a challenging task due to the small target size, background noise, near-coastline noise, etc. In this paper, we propose to employ existing SAR dataset as a transitive dataset to improve the performance of detection model pre-trained on natural images. Specifically, we propose to use the oriented bounding box (OBB) and the contrastive learning method to fully utilize the transitive dataset. We employ a supervised contrastive learning method based on MoCo v2 and trained the model on the transitive dataset which is modified by using the oriented bounding box annotations generated from the existing horizontal bounding box annotations. We evaluate our method on different detection models and the experimental results show that using OBB and supervised contrastive learning can improve the performance of detection models, especially on inshore ship detection task.