| 摘要: |
| 合成孔径雷达（Synthetic Aperture Radar, SAR）在海上目标监测与舰船检测中具有重要应用价值。然而，舰船目标在SAR图像中通常尺度较小、对比度较低，SAR成像过程中产生的乘性斑点噪声会显著干扰目标特征表达，导致小目标检测任务面临较大挑战。针对上述问题，本文在YOLO11n单阶段目标检测框架基础上提出一种面向SAR图像舰船小目标检测的改进方法DFES-YOLO。该方法通过引入去噪模块（Denoiser Module, DM）抑制乘性斑点噪声对特征表达的干扰，从而增强小目标特征的可分辨性；同时设计特征提取模块（Feature Extraction Module, FEM）以提升深层语义信息表达能力，并结合高效多尺度注意力机制（Efficient Multiscale Attention, EMA）增强多尺度特征建模能力。此外，构建小目标检测颈部结构（Small Object Neck, SON），以强化浅层细粒度特征与高层语义信息的融合，从而提高小目标检测性能。在SSDD和HRSID数据集上的实验结果表明，所提出的DFES-YOLO在精确度、召回率、mAP50和mAP50:95等指标上均取得明显提升，表明该方法在复杂噪声环境下的SAR舰船小目标检测任务中具有良好的鲁棒性和泛化能力。 |
| 关键词: 合成孔径雷达、舰船检测、小目标检测、YOLO单阶段检测算法、斑点噪声抑制 |
| DOI： |
| 分类号:TN957.51 |
| 基金项目: |
|
| DFES-YOLO: A Denoising-Enhanced Framework for Detecting Small Ships in SAR Images |
|
|
|
|
| Abstract: |
| Synthetic Aperture Radar (SAR) has significant application value in maritime object monitoring and ship detection. However, ship objects in SAR images are typically characterized by small scale and low contrast. Moreover, the multiplicative speckle noise generated during the SAR imaging process significantly interferes with object feature representation, posing substantial challenges for small-object detection tasks. To address these issues, this paper proposes an improved method, termed DFES-YOLO, for small ship detection in SAR images based on the YOLO11n single-stage object detection framework. Specifically, a Denoiser Module (DM) is introduced to suppress the interference of multiplicative speckle noise on feature representation, thereby enhancing the discriminability of small-object features. In addition, a Feature Extraction Module (FEM) is designed to improve deep semantic feature representation, and an Efficient Multiscale Attention (EMA) mechanism is incorporated to strengthen multi-scale feature modeling capability. Furthermore, a Small Object Neck (SON) structure is constructed to enhance the fusion of shallow fine-grained features and high-level semantic information, thereby improving small-object detection performance. Experimental results on the SSDD and HRSID datasets demonstrate that the proposed DFES-YOLO achieves significant improvements in precision, recall, mAP50, and mAP50:95, indicating that the proposed method exhibits strong robustness and generalization capability for small ship detection in SAR images under complex noise environments. |
| Key words: Synthetic Aperture Radar (SAR) Ship Detection Small Object Detection YOLO Speckle Noise Suppression |