引用本文:孙祎芸,樊臻,董山玲,郑荣濠,兰剑.基于双流对称特征融合网络模型的海洋船舶目标识别[J].控制理论与应用,2022,39(11):2009~2018.[点击复制]
SUN Yi-yun,FAN Zhen,DONG Shan-ling,ZHENG Rong-hao,LAN Jian.Marine ship target recognition using two-stream symmetric feature fusion convolutional neural network[J].Control Theory and Technology,2022,39(11):2009~2018.[点击复制]
基于双流对称特征融合网络模型的海洋船舶目标识别
Marine ship target recognition using two-stream symmetric feature fusion convolutional neural network
摘要点击 1013  全文点击 276  投稿时间：2022-01-05  修订日期：2022-09-27
查看全文  查看/发表评论  下载PDF阅读器
DOI编号  10.7641/CTA.2022.20007
  2022,39(11):2009-2018
中文关键词  识别算法  特征级融合  卷积神经网络  空间注意力机制
英文关键词  recognition algorithm  feature level fusion  convolutional neural networks  spatial attention mechanism
基金项目  NSFC-浙江两化融合联合基金(U1809202)资助.
作者单位邮编
孙祎芸 浙江大学 310027
樊臻* 浙江大学 310027
董山玲 浙江大学 
郑荣濠 浙江大学 
兰剑 浙江大学 
中文摘要
      海洋船舶目标识别在民用和军事领域有着重要的战略意义, 本文针对可见光图像和红外图像提出了一种 基于注意力机制的双流对称特征融合网络模型, 以提升复杂感知环境下船舶目标综合识别性能. 该模型利用双流对 称网络并行提取可见光和红外图像特征, 通过构建基于级联平均融合的多级融合层, 有效地利用可见光和红外两种 模态的互补信息获取更加全面的船舶特征描述. 同时将空间注意力机制引入特征融合模块, 增强融合特征图中关 键区域的响应, 进一步提升模型整体识别性能. 在VAIS实际数据集上进行系列实验证明了该模型的有效性, 其识别 精确度能达到87.24%, 综合性能显著优于现有方法.
英文摘要
      The marine ship target recognition has important strategic significance in civil and military fields, this paper proposes a two-stream symmetric feature fusion convolutional neural network model based on attention mechanism for visible and infrared images, in order to improve the comprehensive recognition performance of ship targets in complex perception environment. The model uses the two-stream symmetric network to extract visible and infrared image features in parallel. By constructing a multi-level fusion layer based on cascade average fusion, the complementary information of visible and infrared modes is effectively used to obtain more comprehensive ship feature description. At the same time, the spatial attention mechanism is introduced into the feature fusion module to enhance the response of key regions in the fusion feature map and further improve the overall recognition performance of the model. A series of experiments on the VAIS real data set have proved the effectiveness of the model, its recognition accuracy can reach 87.24%, and its comprehensive performance is significantly superior to the existing methods.