基于深度学习的遥感图像匹配方法

Remote Sensing Image Matching Method Based on Depth Learning

目前,图像自动配准技术已成为图像配准领域中的研究热点之一。如何提高匹配精度是图像配准的关键步骤,基于此本文提出基于深度学习的匹配方法。首先,描述了卷积神经网络模型的网络结构,通过改善经典模型的网络结构,将其应用到影像匹配任务当中;其次,利用训练好的自适应网络模型来获取控制点的特征表达;最后,将控制点的特征表达通过欧式距离算法进行相似度匹配。实验结果表明,本文方法大幅降低了图像匹配粗匹配的错误率,为后续配准建立了良好基础,且对数据源具有良好的稳健性。

At present,automatic image registration technology has become one of the hotspots in the field of image registration.How to improve the matching precision is the key step of image registration.Based on this,we propose a matching method based on depth learning.First of all,the network structure model of the convolution neural network is described,through the network structure to improve the classic model and its application to image matching task.Secondly,the trained adaptive network model is used to obtain the characteristic expression of the control points.Finally,the control points are characterized by the expression of similarity matching algorithm for Euclidean distance.The experimental results show that the proposed method can greatly reduces the error rate of image matching rough matching,and establishes a good foundation for subsequent registration,and has good robustness to the data source.