深度卷积特征表达的多模态遥感影像模板匹配方法  

A template matching method of multimodal remote sensing images based on deep convolutional feature representation

作　　者：南轲 齐华 叶沅鑫 NAN Ke;QI Hua;YE Yuanxin(Faculty of Geosciences and Environmental Engineering,Southwest Jiaotong University,Chengdu 611756,China)

机构地区：[1]西南交通大学地球科学与环境工程学院

出　　处：《测绘学报》2019年第6期727-736,共10页Acta Geodaetica et Cartographica Sinica

基　　金：四川省科技计划(2017SZ0027)~~

摘　　要：多模态遥感影像间(光学、红外、SAR等)存在显著的非线性辐射差异,传统方法难以有效地提取影像间的共有特征,匹配效果不佳.鉴于此,本文将深度学习方法引入影像匹配中,提出了一种基于Siamese网络提取多模态影像共有特征的匹配方法.首先通过去除Siamese网络中的池化层和抽取特征来优化该网络,保持特征信息的完整性和位置精度,使其可有效地提取多模态影像间的共有特征,然后采用模板匹配策略,实现多模态遥感影像高精度匹配.通过利用多组多模态遥感影像进行试验,结果表明,本文方法的匹配正确率和匹配精度都优于传统的模板匹配方法.Due to significant non-linear radiometric differences between multimodal remote sensing images (e.g., optical, infrared, and SAR), traditional methods cannot efficiently extract common features between such images, and are vulnerable for image matching. To address that, the deep learning technique is introduced into the present study to design a matching method based on Siamese network, which aims to extract common features between multimodal images. The network is first optimized by removing the pooling layer and extracting the feature layer from Siamese network to maintain the integrity and positional accuracy of the feature information, making it possible the effective extraction of common features between multimodal images. Then, the template matching strategy is adopted to achieve high-precision matching of multimodal images. The proposed method is evaluated by using multiple multimodal remote sensing images. The results show that the proposed method outperforms traditional template-matching methods in both the matching correct ratio and matching accuracy.

关 键 词：多模态影像 影像匹配 深度学习 Siamese网络 

分 类 号：P231[天文地球—摄影测量与遥感]