作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2019, Vol. 45 ›› Issue (11): 249-255. doi: 10.19678/j.issn.1000-3428.0052166

• 图形图像处理 • 上一篇    下一篇

基于U-net分割与HEIV模型的遥感图像配准方法

陈辰1a, 周拥军1b, 李元祥1a, 庹红娅1a, 周瑜2, 骆建华1a   

  1. 1. 上海交通大学 a. 航空航天学院;b. 船舶海洋与建筑工程学院, 上海 200240;
    2. 解放军信息工程大学 测绘学院, 郑州 450052
  • 收稿日期:2018-07-19 修回日期:2018-09-30 发布日期:2018-10-19
  • 作者简介:陈辰(1995-),女,硕士研究生,主研方向为图像配准、图像分割;周拥军(通信作者),讲师、博士;李元祥,副教授;庹红娅,副研究员;周瑜,助理研究员、硕士;骆建华,教授。
  • 基金资助:
    国家自然科学基金(41274012);工信部民机专项(MJZ-2016-S-44)。

Remote Sensing Image Registration Method Based on U-net Segmentation and HEIV Model

CHEN Chen1a, ZHOU Yongjun1b, LI Yuanxiang1a, TUO Hongya1a, ZHOU Yu2, LUO Jianhua1a   

  1. 1a. School of Aeronautics and Astronautics;1b. School of Naval Architecture, Ocean and Civil Engineering, Shanghai Jiao Tong University, Shanghai 200240, China;
    2. Institute of Surveying and Mapping, PLA Information Engineering University, Zhengzhou 450052, China
  • Received:2018-07-19 Revised:2018-09-30 Published:2018-10-19

摘要: 在利用航拍遥感图像进行土地测量与变化检测时,需要对图像进行配准处理。为实现目标区域的高精度匹配,提出一种遥感图像配准方法。对图像进行U-net分割,以适用于小样本数据集的处理,针对不同区域特征的误差,将变量含异质噪声模型应用于配准参数估计,提高目标区域的配准精度。实验结果表明,与基于Harris角点的配准方法相比,该方法的全局平均配准精度提高41.39%,与基于SIFT特征点的配准方法相比,其感兴趣区域的平均配准精度提高16.67%。

关键词: 图像配准, 图像分割, 变量含异质噪声模型, 结构化总体最小二乘, 目标区域权值

Abstract: Image registration is required for the aerial remote sensing images when it is utilized for land mapping and change detection.In order to achieve high-precision matching of target regions,a remote sensing image registration method is proposed.To suit for the small sample data sets,the images are segmented by U-net.Considering the errors of different regional features,Heteroscedastic Errors-in-Variables(HEIV) model is applied to the registration parameter estimation to improve the registration accuracy of the target region.Experimental results show that compared with the Harris corner-based registration method,the global average registration accuracy of the proposed method is improved by 41.39%,and compared with the Scale-Invariant Feature Transform(SIFT) feature-based registration method,the average registration accuracy of the regions of interest is increased by 16.67%.

Key words: image registration, image segmentation, Heteroscedastic Errors-in-Variables(HEIV) model, Structured Total Least Squares(STLS), weights of target region

中图分类号: