基于低秩双线性池化注意力网络的舰船目标识别  

Ship target recognition based on low rank bilinear pooling attention network

作　　者：关欣 国佳恩 衣晓 GUAN Xin;GUO Jiaen;YI Xiao(Naval Aviation University,Yantai 264001,China;Unit 91422 of the PLA,Yantai 265200,China)

机构地区：[1]海军航空大学,山东烟台264001 [2]中国人民解放军91422部队,山东烟台265200

出　　处：《系统工程与电子技术》2023年第5期1305-1314,共10页Systems Engineering and Electronics

基　　金：国防科技卓越青年科学基金(2017-JCJQ-ZQ-003);泰山学者工程专项经费(ts 201712072)资助课题。

摘　　要：针对多模态舰船图像融合识别质量不高等问题,构建了一种端到端的低秩双线性池化注意力网络。首先对各模态原始特征向量基于跨模类别中心进行注意力加权重构,使不同模态特征更好地关注公共类别信息;然后采用双线性池化捕获不同模态图像的交互信息,并引入权重低秩分解降低网络参数规模;最后依靠特征级联实现模态信息的交互与互补,并设计联合损失提升网络跨模态融合识别效果。实验结果表明,相比现有融合方法,所提方法可有效提升多模态遥感舰船图像的融合识别效果,在公开的遥感舰船数据集上取得了较高的识别准确率。In order to solve the problem of low quality of multimodal ship image fusion recognition,an end-to-end low rank bilinear pooling attention network is constructed.Firstly,the original feature vector of each modal are reconstructed based on the cross-modal category center,so that different modal features can pay more attention to the common category information.Then,bilinear pooling is used to capture the interactive information of different modal images,and the weight low rank decomposition is introduced to reduce the scale of network parameters.Finally,the interaction and complementarity of modal information are realized by feature cascaded,and the joint loss is designed to improve the effect of network cross-modal fusion recognition.The experimental results show that compared with the existing fusion methods,the proposed method can effectively improve the fusion recognition effect of multimodal remote sensing ship images,and achieve high recognition accuracy on the public remote sensing ship datasets.

关 键 词：舰船识别 双线性池化 跨模类别中心 注意力加权 跨模联合损失 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]