ShipSense YOLO: a YOLOv8-based algorithm for remote sensing ship image detection

As a critical task in marine monitoring, optical remote sensing ship image detection is widely applied in maritime safety, environmental protection, and shipping management. Based on the YOLOv8 baseline model, this paper first incorporates a Deformable Attention module to enhance the focus on salient features. Second, an auxiliary head (Auxiliary head) is combined with the original detection head to improve the recognition of small targets and training stability. Finally, the NWD loss function is introduced to optimize the inter-sample differences during model training. Experimental results demonstrate that the improved ShipSense-YOLO model exhibits significant improvements in ship detection tasks compared to the baseline YOLOv8. Under the same computational resources, the improved model achieves higher accuracy on the test set, specifically: Precision (P) increases from 87.7% to 89.1%, Recall (R) increases from 78.8% to 79.1%, mAP50 increases from 84.9% to 86.2%, and mAP50-95 increases from 47.2% to 48.4%. Despite the introduction of new modules, the model's inference speed remains at a low latency, indicating that the improved model has good real-time processing capabilities. In summary, the proposed improvement method effectively enhances the accuracy and robustness of ship detection, providing new insights and methods for remote sensing ship image detection tasks.