DENSE SHIP DETECTION GUIDED BY CENTRALITY PRIOR INFORMATION IN SAR IMAGES

The detection of densely distributed ship targets is one of the hot issues in the context of convolutional neural network (CNN)-based synthetic aperture radar (SAR) image processing. In this case, the bounding boxes of the ships may overlap with each other. Traditional detectors do not specifically consider the processing of overlapping areas, resulting in low detection performance. To address this problem, we proposed a new SAR ship detector, where classification confidence score-based method is developed to consider the centrality prior information among the overlap areas. Then, in the shallow layers of the network, the auxiliary heads are used to guide the network to learn the features related to centers of ships. Experimental results on the open datasets with dense ships show that our method achieves the better detection performance without the obvious increase of computation burden compared with the current state-of-the-art detectors.