计算机工程与应用 ›› 2017, Vol. 53 ›› Issue (20): 180-186.DOI: 10.3778/j.issn.1002-8331.1702-0105

• 图形图像处理 • 上一篇    下一篇

一种改进的基于深度学习的遥感影像拼接方法

雒培磊1,李国庆2,曾  怡1   

  1. 1.北京林业大学 信息学院,北京 100083
    2.中国科学院 遥感与数字地球研究所 数据技术部,北京 100094
  • 出版日期:2017-10-15 发布日期:2017-10-31

Modified approach to remote sensing image mosaic based on deep learning

LUO Peilei1, LI Guoqing2, ZENG Yi1   

  1. 1.School of Information Science and Technology, Beijing Forestry University, Beijing 100083, China
    2.Satellite Data Technology Division, Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences, Beijing 100094, China
  • Online:2017-10-15 Published:2017-10-31

摘要: 针对遥感影像拼接的两个主要过程:图像配准和点变换,分别进行了深入研究。对遥感影像拼接中的特征点匹配问题,提出了一种利用分层卷积特征进行图像配准的方法。该方法利用卷积神经网络(Convolutional Neural Networks,CNN)自适应地提取特征点的分层卷积特征,通过相关滤波器(Correlation Filter,CF)对不同深度的卷积特征逐层进行相关性分析,进而综合计算特征点的位置。然后对传统的点变换方法进行简化,提出十字点集变换方法。根据配准的特征点计算变换参数,实现遥感影像的拼接。实验结果表明,该方法与传统的基于SIFT(Scale Invariant Feature Transform)的拼接方法相比,精度较高且具有较好的鲁棒性。

关键词: 卷积神经网络(CNN), 图像配准, 十字点集, 遥感影像拼接

Abstract: Image registration and coordinate transformation are two important processes of remote sensing image mosaic in different backgrounds, which are studied respectively in this paper. Focused on the image registration, a method based on hierarchical convolutional features is proposed. This method adaptively obtains image features from CNN(Convolutional Neural Networks) and then sends the features derived from different images in different depth to CF(Correlation Filter) to compute the similarity between them. Therefore the locations of the feature points are computed with considering the different depth hierarchical convolutional features. In order to simplify the coordinate transform method, the cross points method is proposed. According to the feature points location from the image registration, the transform parameters can be computed. Then the remote sensing images can be mosaicked by converting all pixels from one image to another. The experimental results show the effectiveness and robustness of the proposed method by comparing to traditional mosaic method based on SIFT(Scale Invariant Feature Transform).

Key words: Convolutional Neural Networks(CNN), image registration, cross points, remote sensing image mosaic