| 摘要: |
| 针对卫星合成孔径雷达（Synthetical Aperture Radar,SAR）图像中目标尺度小、特征弱、背景干扰多等不利因素造成的目标检测算法普遍精度不高、漏检误检严重的问题，提出了一种基于YOLOv8n改进的新算法。该算法通过移动翻转瓶颈卷积（Mobile Inverted Bottleneck Convolution,MBConv）来提升骨干网络特征提取效率，使用可变形注意力机制（Deformable Attention,DAttention）来扩大骨干网络的特征感受视野，在特征融合网络中引入并行补丁感知注意力（Parallelized Patch-aware Attention,PPA）以提升特征融合精度。该算法改进了原网络在小尺度目标检测能力上的不足。在公开数据集SAR Ship Dataset上对舰船目标进行检测识别，与原网络相比，精确率提升了1.8%，召回率提升了4.5%，平均精度mAP@50提升了1.8%。同时，在其他数据集上进行模型泛化实验，结果显示改进算法泛化能力较强，具有一定的工业应用价值。 |
| 关键词: SAR图像 小尺度目标检测 可变形注意力 移动翻转瓶颈卷积 并行补丁感知注意力 |
| DOI：10.20079/j.issn.1001-893x.240803001 |
|
| 基金项目: |
|
| YOLOv8n-DESP:an Improved Algorithm for Small-scaleShip Object Detection in SAR Image with Complex Background |
| LIU Xue |
| (National Key Laboratory of Complex Aviation System Simulation,Chengdu 610036,China) |
| Abstract: |
| The precision and recall rate of object detection in synthetical aperture radar（SAR） image with complex background are both low and hard to raise.For above problems,an improved YOLOv8n algorithm is proposed which has better performance by using mobile inverted bottleneck convolution(MBConv),deformable attention(DAttention) and parallelized patch-aware attention(PPA) modules.The MBConv can increase the efficiency of feature extraction,and the DAttention module broadens the field of perception in the backbone.Besides,by introducing the PPA module in the neck,the precision of feature fusion is raised up.By testing in the SAR Ship Dataset,the results show that the precision,mAP@50 and recall rate are raised by 1.8%,1.8%,and 4.5%,respectively.Meanwhile,experiments in other four datasets are carried out and the results prove the performance of YOLOv8n-DESP are better than that of YOLOv8n.In brief,YOLOv8n-DESP is robust and generalized and has industrial application value. |
| Key words: SAR image small-scale object detection deformable attention mobile inverted bottleneck convolution parallelized patch-aware attention |
《电讯技术》编辑部 