He Rui, Han Dezhi, Shen Xiang, Han Bing, Wu Zhongdai, Huang Xiaohu
College of Information Engineering, Shanghai Maritime University, Shanghai, China.
School of Computer Science, The University of Sydney, Sydney, New South Wales, Australia.
PLoS One. 2025 Jul 30;20(7):e0327362. doi: 10.1371/journal.pone.0327362. eCollection 2025.
Synthetic Aperture Radar (SAR), renowned for its all-weather monitoring capability and high-resolution imaging characteristics, plays a pivotal role in ocean resource exploration, environmental surveillance, and maritime security. It has become a fundamental technological support in marine science research and maritime management. However, existing SAR ship detection algorithms encounter two major challenges: limited detection accuracy and high computational cost, primarily due to the wide range of target scales, indistinct contour features, and complex background interference. To address these challenges, this paper proposes AC-YOLO, a novel lightweight SAR ship detection model based on YOLO11. Specifically, we design a lightweight cross-scale feature fusion module that adaptively fuses multi-scale feature information, enhancing small target detection while reducing model complexity. Additionally, we construct a hybrid attention enhancement module, integrating convolutional operations with a self-attention mechanism to improve feature discrimination without compromising computational efficiency. Furthermore, we propose an optimized bounding box regression loss function, the Minimum Point Distance Intersection over the Union (MPDIoU), which establishes multi-dimensional geometric metrics to accurately characterize discrepancies in overlap area, center distance, and scale variation between predicted and ground truth boxes. Experimental results demonstrate that, compared with the baseline YOLO11 model, AC-YOLO reduces parameter count by 30.0% and computational load by 15.6% on the SSDD dataset, with an average precision (AP) improvement of 1.2%; on the HRSID dataset, the AP increases by 1.5%. This model effectively reconciles the trade-off between complexity and detection accuracy, providing a feasible solution for deployment on edge computing platforms. The source code for the AC-YOLO model is available at: https://github.com/He-ship-sar/ACYOLO.
合成孔径雷达（SAR）以其全天候监测能力和高分辨率成像特性而闻名，在海洋资源勘探、环境监测和海上安全方面发挥着关键作用。它已成为海洋科学研究和海洋管理的重要技术支撑。然而，现有的SAR舰船检测算法面临两个主要挑战：检测精度有限和计算成本高，这主要是由于目标尺度范围广、轮廓特征不清晰以及复杂的背景干扰所致。为应对这些挑战，本文提出了AC-YOLO，一种基于YOLO11的新型轻量级SAR舰船检测模型。具体而言，我们设计了一个轻量级跨尺度特征融合模块，自适应地融合多尺度特征信息，增强小目标检测能力，同时降低模型复杂度。此外，我们构建了一个混合注意力增强模块，将卷积操作与自注意力机制相结合，在不影响计算效率的情况下提高特征辨别能力。此外，我们提出了一种优化的边界框回归损失函数，即最小点距离交并比（MPDIoU），它建立了多维几何度量，以准确表征预测框与真实框在重叠面积、中心距离和尺度变化方面的差异。实验结果表明，与基线YOLO11模型相比，AC-YOLO在SSDD数据集上参数数量减少了30.0%，计算量减少了15.6%，平均精度（AP）提高了1.2%；在HRSID数据集上，AP提高了1.5%。该模型有效地平衡了复杂度与检测精度之间的权衡，为在边缘计算平台上部署提供了可行的解决方案。AC-YOLO模型的源代码可在以下网址获取：https://github.com/He-ship-sar/ACYOLO。
