系统工程与电子技术 ›› 2023, Vol. 45 ›› Issue (5): 1297-1304.doi: 10.12305/j.issn.1001-506X.2023.05.05

• 电子技术 • 上一篇    

基于卷积特征的场景地标检索方法

董思强1,*, 邓年茂1, 刘琰1, 张玉宝2   

  1. 1. 北京控制与电子技术研究所, 北京 100045
    2. 中国航天科工集团第四研究院, 北京 102308
  • 收稿日期:2022-04-06 出版日期:2023-04-21 发布日期:2023-04-28
  • 通讯作者: 董思强
  • 作者简介:董思强(1981—), 男, 高级工程师, 博士研究生, 主要研究方向为视觉导航、制导与控制、深度学习
    邓年茂(1963—), 男, 研究员, 博士, 主要研究方向为视觉导航、制导与控制、光电技术
    刘琰(1977—), 男, 研究员, 硕士, 主要研究方向为制导与控制、视觉导航、深度学习、测试技术
    张玉宝(1978—), 男, 高级工程师, 硕士, 主要研究方向为制导与控制、深度学习、测试技术

Scene landmark retrieval method based on convolution feature

Siqiang DONG1,*, Nianmao DENG1, Yan LIU1, Yubao ZHANG2   

  1. 1. Beijing Institute of Control and Electronic Technology, Beijing 100045, China
    2. The 4th Academy of China Aerospace Science and Industry Corporation, Beijing 102308, China
  • Received:2022-04-06 Online:2023-04-21 Published:2023-04-28
  • Contact: Siqiang DONG

摘要:

针对场景中地理目标的检索任务主要解决在视角变化、光照变化甚至遮挡等情况下对地理目标的检索匹配问题, 也称为实例目标检索,使用高性能的卷积网络构建用于实例目标检索任务的三输入孪生网络架构, 采用三元组损失函数进行训练, 并使用区域建议网络准确定义目标区域, 生成准确并具有鲁棒性且固定长度的图像特征向量。检索时根据地理场景的特点采用图像全局特征进行粗检索, 采用局部特征进行精检索, 并配合查询扩展的方法实现了精确的实例目标检索结果。实验表明, 所提方法与其他具有代表性的检索方法相比, 在公开数据集测试中取得了有竞争力的结果。

关键词: 卷积网络, 地标匹配, 孪生网络架构, 三元组损失函数, 实例目标

Abstract:

Retrieval task of scene landmark matching mainly solves the retrieval and matching problem of geographical targets under the conditions of view angle change, illumination change, and even occlusion. It is also called instance target retrieval. In this paper, we design a three streams' siamese network architecture with high performance convolutional network, the triple loss function, and the regional proposal network is used to accurately define the target area, so generate a accurate and robust image feature representation with fixed length. According to the characteristics of the scene landmark matching, the global features of the image are used for coarse retrieval, and the local features are used for fine retrieval. And combined with the query expansion method, accurate instance target retrieval is achieved. Experiments show that compared with other representative retrieval methods, this method can achieve competitive results in public data set testing.

Key words: convolutional network, landmark matching, siamese network architecture, triplet loss function, instance target

中图分类号: