多源遥感影像学习型特征双向一致性配准

Multi-source remote sensing image bidirectional consistent registration based on learning feature

针对多源遥感影像之间非线性辐射和几何畸变的差异严重影响配准质量的问题,本文提出一种具有双向一致性变换适用于多源遥感影像的配准方法。首先,利用微调的ResNet101网络模型提取多源遥感影像学习型特征,在特征匹配阶段,为提高同名特征匹配的可靠性,设计了一种双向一致性特征匹配网络模型;然后,基于小型轻量级网络加权回归变换模型参数,实现多源遥感影像稳健可靠的配准。试验利用Google Earth影像、卫星影像、无人机影像、Google Earth-卫星-无人机混合影像4种不同数据源对本文方法进行测试,并与具有代表性的多种方法进行比较,结果表明本文方法在配准精度、效率、稳健性方面具有优势,基本实现了2像素以内的自动配准精度。

A robust registration method with bidirectional consistent transformation is proposed to solve the problem of multi-source remote sensing image relatively poor registration effect due to large nonlinear radiation and geometric distortion.First,the fine-tuned ResNet101 network model was utilized to extract images learning features.To improve the reliability of the corresponding feature matching,we designed a bidirectional consistency feature matching model in the feature matching stage.Then,robust registration was achieved by using a parametric regression transformation model with weight based on a small lightweight network.In our experiments,we tested the proposed algorithm using Google Earth images,satellite images,UAV images and Google Earth-satellite-UAV images,and compared it with the several typical methods.The results show that the proposed method has advantages in the accuracy,efficiency and robustness,and achieves automatic registration accuracy almost within 2 pixels.