DGSP-YOLO: A Novel High-Precision Synthetic Aperture Radar (SAR) Ship Detection Model

With the rapid advancement of deep learning, its application in synthetic aperture radar (SAR) ship target detection has become increasingly prevalent. However, the detection of ships in complex environments and across various scales remains a formidable challenge. This paper introduces DGSP-YOLO, a novel high-performance detection model designed to overcome these hurdles. The model integrates the SPDConv and C2fMHSA modules into the YOLOv8n baseline, significantly enhancing the feature extraction capabilities for small-scale targets. Additionally, the original convolutional blocks have been optimized with GhostConv, ensuring efficient performance and reduced parameter count. To further refine the detection process, the DySample module has been incorporated to mitigate noise interference, leading to the generation of more refined feature maps. The model also employs EIoU to bolster its capacity to process images of varying quality. Extensive experiments on the HRSID, LS-SSDD-v1.0, and SSDD datasets have been conducted to test the model's effectiveness rigorously. The results demonstrate that DGSP-YOLO outperforms other prevalent models, achieving mAP50 and mAP50:95 scores of 94% and 72.2% on the HRSID dataset, and 69% and 25.3% on the LS-SSDD-v1.0 dataset, respectively. On the SSDD dataset, the model achieved an impressive mAP50 and mAP50:95 of 99% and 75.1%, respectively. These outcomes underscore DGSP-YOLO's superior accuracy and overall performance, marking a significant advancement in SAR ship target detection.