一种改进的基于深度学习的遥感影像拼接方法

发布时间：2018-02-05 20:48

本文关键词： 卷积神经网络(CNN) 图像配准十字点集遥感影像拼接　出处：《计算机工程与应用》2017年20期 　论文类型：期刊论文

【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。
[Abstract]:In view of the two main processes of remote sensing image mosaic: image registration and point transformation, this paper studies the feature points matching in remote sensing image mosaic. A method of image registration using hierarchical convolution features is proposed, which uses convolutional Neural Networks. CNNs adaptively extract hierarchical convolution features of feature points and correlate Filter through correlation filters. CFC) carries on the correlation analysis to the different depth convolution characteristic layer by layer, then synthetically calculates the characteristic point position, then simplifies the traditional point transformation method. A cross point set transform method is proposed. The transform parameters are calculated according to the registered feature points, and the remote sensing images are stitched. The experimental results show that. Compared with the traditional SIFT(Scale Invariant Feature transform method, this method has higher precision and better robustness.
【作者单位】：北京林业大学信息学院;中国科学院遥感与数字地球研究所数据技术部;
【基金】：海南省重大科技计划项目(No.ZDKJ2016021) 国家重点研发计划(No.2016YFB0501504) DBAR中科院“一带一路”专项项目北京市自然科学基金(No.6164038)
【分类号】：TP751

【相似文献】