A New Ship Detection Method with Limited Labeled Data in SAR Imagery

The ship detection from synthetic aperture radar remains a challenging problem for maritime surveillance applications. In the fully-supervised object detection methods, a lot of labeled bounding boxes (training data) are required, but the exact instance-level annotations are difficult to obtain. The classical approaches that are based on statistical characteristics and polarization characteristics do not require any annotation, but the statistical distribution model cannot match the data perfectly, which often leads to high false alarm rates. In order to solve this problem, a dynamic data augmentation method (DDAM) is proposed for dealing with missing labels and lacking annotations for ship detection. In DDAM, we just annotate a small number of ship targets in training data before training, then the proposed learning framework can dynamically update training data with pseudo-annotated data in iterations. The detection performances are evaluated by mean Average Precision (mAP). Extensive experiments on the Sentinel-1A SAR data show the effectiveness of this new method.