Multi-scale ship detection with high-resolution SAR images based on improved YOLOv8

Abstract

To improve the detection accuracy of SAR ship targets under complex background clutter, this paper proposes an improved YOLOv8 detection method for SAR ship detection tasks. Three optimization strategies are adopted, including CA attention module, improved DSConv convolution module, and Shape-IoU loss function. A series of ablation experiments are performed to verify the effectiveness of different embedding positions, independent modules, and multi-module collaborative optimization. Quantitative results show that the proposed method achieves significant improvements in precision, recall and mAP@0.5 compared with the original YOLOv8 baseline. The introduced CA module enhances multi-scale feature representations of ship targets while attenuating responses from irrelevant background regions. The DSConv module extracts fine-grained ship features to reduce missed and false detections. The Shape-IoU loss function introduces shape-aware regularization to optimize bounding-box shape matching for ship targets. It improves multiple metrics when used alone. However, when jointly optimized with CA and DSConv, the complete model achieves prominent gains in precision, recall and mAP@0.5 at the cost of degraded mAP@0.5:0.95 due to multi-module coupling effects. Comparative experiments with multiple mainstream lightweight models demonstrate that the proposed method achieves superior detection performance under conventional IoU thresholds. Visualization results on four polarization modes further verify the strong anti-clutter capability and stable feature extraction ability of the improved model. Cross-dataset generalization tests show that the model trained on the SAR-SHIP (2021) dataset maintains stable accuracy gains on HRSID and SSDD datasets, proving excellent cross-domain generalization ability. Nevertheless, the multi-module optimization brings a certain performance trade-off. The proposed method sacrifices partial fine localization accuracy and inference speed to obtain higher conventional detection precision. It is suitable for high-precision SAR ship detection scenarios with moderate real-time requirements.