ESD-Ship: An Fast and Accurate SAR Ship Detection Method

To address the issue of low accuracy and lengthy processing time in synthetic aperture radar (SAR) ship detection based on deep learning, this paper proposes an improved ship detection method based on YOLOv7 (You Only Look Once). Building upon the YOLOv7 algorithm, the overall feature learning architecture is simplified by removing unnecessary large target feature processing structures. Subsequently, a downsampling module with enhanced feature extraction capability is constructed by integrating the RepVGGBlock. Finally, the traditional NMS time-consuming and inaccurate positioning issues are optimized by combining Weighted Non Maximum Suppression (NMS) with Cluster-NMS. Compared to the original algorithm, the proposed method achieves a substantial improvement on the HRSID dataset, with the mean average precision (mAP) increasing from 91.8% to 94.9% and FPS rising from 38.3 to 46.3, outperforming the control significantly.