YOLO-SSOD: Enhancing SAR Ship Detection with Bi-Level Routing Attention in YOLOv10

Maritime surveillance has gained more attention in recent years due to deep-learning vision methods becoming more and more accurate and discriminative. Among different ways to capture maritime images, Synthetic Aperture Radars (SARs) for ships have become an important method due to their advantages in various weather and lighting conditions. Therefore, this work presents a hybrid model approach using Bi-Level Routing Attention (BRA) among backbone and neck layers to improve the YOLOv10n original model when facing SAR images in a maritime context. The BRA insertion tries to capture relevant features by focusing on small objects in the images. The experiment results reveal that the proposed models achieved 97.70% mean Average Precision (mAP) against 97.36% of the original YOLOv10 retrained for SAR context on the SAR Ship Dataset. Also, the proposed model outperformed all analyzed models in the last five years using different approaches on the same dataset. The most recent model achieved 97.72% while our model 97.77%, using a different combination of training and test image amounts.