摘要
针对合成孔径雷达(synthetic aperture radar,SAR)图像的成像机理导致SAR图像受噪声影响严重且SAR目标类别存在不均衡问题,使得SAR目标特征刻画较困难。针对上述问题,提出一种基于动态聚合网络(dynamic aggregation network,DANet)的SAR目标检测方法。该网络以无锚框全卷积一阶段目标检测(fully convolutional one-stage object detection,FCOS)网络为基础网络框架。DANet在特征金字塔网络中嵌入具有动态坐标注意力的动态聚合模块,以提高噪声影响下SAR目标特征学习能力。为了缓解SAR数据集中不同类别之间数量不平衡的问题,DANet在检测头的回归分支引入类平衡动态交并比(intersection over union,IoU)损失函数,通过动态非单调机制和类平衡因子引导模型参数更新。在MSAR-1.0数据集和SAR-AIRcraft-1.0数据集上的实验结果表明,DANet在上述SAR目标检测数据集上的目标检测准确率分别达到了70.25%和52.36%,相比基准网络FCOS,其平均精度分别提高了2.73%和3.67%。与其他相关算法相比,DANet在SAR图像目标检测精度方面具有明显优势。
Due to the imaging mechanism of synthetic aperture radar(SAR)images,SAR images are severely affected by noise and there is an imbalance in SAR target categories,making it difficult to characterize SAR target features.In view of these problems,a SAR target detection method based on dynamic aggregation network(DANet)is proposed.This network is based on the fully convolutional one-stage object detection(FCOS)network without anchor boxes.DANet embeds a dynamic aggregation module with dynamic coordinate attention in the feature pyramid network to improve the SAR target feature learning ability under noise influence.In order to alleviate the problem of imbalanced numbers between different categories in SAR datasets,DANet introduces a class balanced dynamic intersection over union(IoU)loss function in the regression branch of the detection head,which guides model parameter updates through dynamic non monotonic mechanisms and class balance factors.The experimental results on the MSAR-1.0 dataset and SAR-AIRcraft-1.0 dataset show that DANet achieved target detection accuracies of 70.25%and 52.36%,respectively,on the aforementioned SAR target detection datasets.Compared with the benchmark network FCOS,its average accuracy increased by 2.73%and 3.67%,respectively.Compared with other related algorithms,DANet has significant advantages in SAR image target detection accuracy.
作者
倪康
贾文杰
邹旻瑞
郑志忠
NI Kang;JIA Wenjie;ZOU Minrui;ZHENG Zhizhong(School of Computer Science,Nanjing University of Posts and Telecommunications,Nanjing 210023,China;Jiangsu Province Engineering Research Center of Airborne Detecting and Intelligent Perceptive Technology,Nanjing 210049,China)
出处
《系统工程与电子技术》
北大核心
2025年第8期2527-2539,共13页
Systems Engineering and Electronics
基金
国家自然科学基金(62101280)
江苏省自然科学基金(BK20210588)
中国博士后科学基金(2023M731781)
江苏省航空对地探测与智能感知工程中心开放基金(JSECF2023-01,JSECF2023-05)
雷达成像与微波光子技术教育部重点实验室(南京航空航天大学)基金(NJ20230005)
南京邮电大学引进人才科研启动基金(NY222107)资助课题。
关键词
合成孔径雷达
注意力机制
目标检测
特征金字塔
损失函数
synthetic aperture radar(SAR)
attention mechanism
object detection
feature pyramid
loss function
