Synthetic Aperture Radar Ship Detection Network Based on Multi-Scale Feature Enhancement

The paper introduces a ship identification network with SAR imagery, termed Multi-scale Feature enhancement YOLOX (MF-YOLOX), specifically designed to tackle the challenge of poor accuracy in recognizing multi-scale targets under complex circumstances posed by existing algorithms. Within the feature extraction network, the EMA module is incorporated to efficiently encode global spatial information and establish both short-term and long-term dependencies. Furthermore, the CIoU loss function is employed to substitute the original IoU in the prediction network, aiming to mitigate the positional bias issue of the prediction frame and enhance the model’s precision in the field of target localization. Experimental outcomes based on the HRSID dataset demonstrate that our approach achieves an $\mathbf{F} 1$ score of 89%, an average accuracy mean Average Precision (mAP) of $\mathbf{9 1 \%}$, and a detection speed of 51 Frames Per Second (FPS). Our approach offers the highest detection accuracy and operating efficiency when compared to some advanced methods, providing technical support for precise ship detection in complex maritime conditions.