ACmixSAR-YOLO11: A SAR Ship Detection Model Based on YOLO11

Hybrid deep learning frameworks integrating convolutional neural networks (CNNs) and Transformer-based attention mechanisms have recently achieved promising results in synthetic aperture radar (SAR) ship detection. However, they still suffer from inherent limitations, including missed detections of small-scale ships under large variations in target scales, multiplicative speckle noise, and complex sea clutter interference. To address these issues, this paper proposes ACmixSAR-YOLO11, a SAR ship detection framework built upon YOLO11. The proposed model incorporates three complementary improvements to enhance feature representation and multi-scale detection capability. First, an ACmixSAR module is designed by integrating convolutional operations with self-attention mechanisms. This design strengthens feature discriminability. Second, a SARNoiseGate gating mechanism is embedded into the ACmixSAR module, which is designed to suppress low signal-to-noise ratio (SNR) feature responses while preserving discriminative ship features. Third, we develop a scale-weighted complete intersection over union (SW-CIoU) loss function, which assigns adaptive regression weights based on target sizes to rebalance the optimization across small, medium, and large ships. All improvements below are reported as relative percentage changes. Experimental results on the SSDD dataset demonstrate that the proposed ACmixSAR-YOLO11 outperforms the original YOLO11 Baseline, achieving a 1.50% increase in precision, a 0.15% growth in recall, a 1.47% enhancement in mAP@0.5, and a 4.26% promotion in mAP@0.5:0.95. Significance testing was applied to mAP@0.5 only (paired one-sided t-test, p = 0.0058 < 0.01); the other three metrics are reported as three-seed means and spreads. As supplementary evidence, experiments on HRSID show improved recall, mAP@0.5 and mAP@0.5:0.95 under conventional patch-level splitting at the cost of slightly reduced precision. However, this patch-level performance gain is partially inflated by scene overlap between training and test patches. Further validation on a strictly leakage-free scene-level split of HRSID shows that the overall mAP@0.5 gain diminishes to within random seed fluctuation, while improvements in mAP@0.5:0.95 and precision remain observable. These results demonstrate that the proposed framework effectively enhances detection accuracy with a moderate parameter increase (9.69 M → 11.75 M) while maintaining real-time inference.