Dense Ship Detection in SAR Images via Comprehensive Confidence Score-Based Label Assignment

Convolutional neural network (CNN) based ship detection in synthetic aperture radar (SAR) images has achieved impressive attention in recent years. In practice, bounding boxes (B-Boxes) of dense inshore ships in training data have large overlapping areas, which lead to indistinct labels and increase the difficulty to train CNN-based detectors with high performance. To address this issue, a comprehensive confidence score-based detector (CCSDet) is proposed in this article. In CCSDet, we refine the ground-truth B-Boxes (labels) in the overlapping regions of dense ships based on the comprehensive confidence score, which considers both the centrality scores of the labels and the prediction quality of the model in the training stage. The refinement of labels reduces the ambiguity of overlapping ship regions in the training process. Besides, a pixel sample selection strategy is introduced to encourage our detection model to focus on both the total dense target areas and the center of dense ships when the localization loss is calculated. Extensive experiments conducted on the public SAR ship datasets show that our method outperforms the existing methods in the case of densely docked ship detection in SAR images.