基于图注意力网络的半监督SAR舰船目标检测  

Semi-supervised SAR Ship Target Detection with Graph Attention Network

作　　者：吕进东 王彤[1] 唐晓斌[2] LÜJindong;WANG Tong;TANG Xiaobin(National Key Laboratory of Radar Signal Processing,Xidian University,Xi’an 710071,China;Academy of Electronics and Information Technology,China Electronic Technology Group Corporation,Beijing 100041,China)

机构地区：[1]西安电子科技大学雷达信号处理国家级重点实验室,西安710071 [2]中国电子科技集团电子科学研究院,北京100041

出　　处：《电子与信息学报》2023年第5期1541-1549,共9页Journal of Electronics & Information Technology

基　　金：国家重点研发计划(2016YFE0200400)。

摘　　要：基于深度学习的合成孔径雷达(SAR)舰船目标检测近年得到了快速发展。然而,传统有监督学习需要大量的标记样本来训练网络。针对此问题,该文提出一种基于图注意力网络(GAT)的半监督SAR舰船目标检测方法。首先,设计了对称卷积神经网络用于海陆分割。随后,完成超像素分割并将超像素块建模为GAT的节点,利用感兴趣区域池化层提取节点的多尺度特征。GAT采用注意力机制自适应地汇聚邻接节点特征实现对无标记节点的分类。最后,将预测为舰船目标的超像素块定位到SAR图像中并获得精细检测结果。在实测高分辨SAR图像数据集上验证了所提方法。结果表明该方法可以在少量标记样本下,以低虚警率实现对舰船目标的可靠检测。Recently,ship target detection in Synthetic Aperture Radar(SAR)imagery based on deep learning has been widely developed.However,a large number of labeled samples are needed in traditionally supervised learning to train the network.Therefore,a semi-supervised SAR ship target detection approach based on Graph ATtention network(GAT)is proposed.Firstly,a symmetric convolutional neural network is designed to realize land-ocean segmentation.Secondly,the super-pixel segmentation is completed and the super-pixels are modeled as nodes of the GAT.The multi-scale features of a node are extracted by region of interest pooling layer.Attentional mechanisms are used in GAT to concatenate adaptively the neighbor node’s features and classify the unlabeled nodes.Finally,the super-pixels predicted as ship targets are located in SAR image and the fine detection results are obtained.The proposed method is verified on the measured high resolution SAR images dataset.The results show that this method can effectively detect ship targets with low false alarm rate by using a small number of labeled samples.

关 键 词：雷达目标检测 图注意力网络 半监督学习 舰船目标检测 

分 类 号：TN957.51[电子电信—信号与信息处理]