YOLO-SAIL: Attention-Enhanced YOLOv5 With Optimized Bi-FPN for Ship Target Detection in SAR Images

Synthetic Aperture Radar (SAR) is useful for monitoring sea surfaces and detecting targets on ships. However, interpreting SAR images can be challenging due to the high density of ships, an imbalanced foreground-to-background ratio, and the small size of targets. It has recently become increasingly popular to apply deep learning algorithms to the identification of ships in SAR images. Nevertheless, the presence of intricate backgrounds and multi-scale vessels makes it difficult for deep networks to detect distinctive
targets, in part due to the presence of intricate backgrounds and multi-scale vessels. As a solution to these
challenges, this research article introduces a new ship identification model known as YOLO-SAIL. The improved Bi-directional Feature Pyramid Network (Bi-FPN) has been replaced by a conventional Path Aggregation Network (PAN) to extract more powerful semantic features and sharpen the distinction between multi-scale targets. Secondly, a Normalized Attention Mechanism (NAM) is introduced to optimize attention toward dense ship targets by merging contextual cues with global interdependencies. Finally, a new layer of feature fusion is introduced to locate smaller ships. We evaluated YOLO-SAIL against two public benchmark
datasets, the SAR Ship Detection Dataset (SSDD) and the High Resolution SAR Images Dataset (HRSID). The proposed model, YOLO-SAIL, achieves an F1-score of 98.22% in SSDD datasets and 94.68% in HRSID datasets, outperforming most existing models.