面向多模态遥感影像旋转差异的智能匹配方法

Intelligent matching algorithm for rotation difference between multimodal remote sensing images

遥感影像配准是图像处理技术领域的重要研究内容,影像间的大旋转差异严重影响匹配精度的提高。针对这个问题,提出基于双分支神经网络的旋转差异校正方法。利用相位一致性信息特征值与方向角提取影像的旋转特征向量(RVPC);构建并训练用于预测影像间旋转差异角的双分支神经网络R-FCN;将RVPC输入网络并预测旋转差异角,最后基于预测角对影像进行仿射校正。R-FCN网络在SEN1-2数据集上,5°(±2.5°)精度训练准确率达98.17%。

Remote sensing image registration is a research of significance in the field of image processing.Large rotation difference between multimodal images seriously affects the improvement of the final quality of image matching.To solve this problem,a rotation difference correction method based on double branch neural network is proposed.Firstly,with the calculation of eigenvalue and orientation of phase congruency information,the rotation feature vector of the image is extracted,named RVPC.After that,a double branch neural network R-FCN is constructed and fully trained to predict the rotation difference angle between images.Furthermore,two RVPC vectors are input into the network and we get an output prediction vector and then calculate the prediction angle.Finally,the image is affine corrected based on the angle.On public data set SEN1-2,the training accuracy of network R-FCN reaches 98.17%.