基于卷积神经网络和校正网络相结合的遥感图像配准方法研究

RESEARCH ON THE REGISTRATION METHOD OF REMOTE SENSING IMAGE BASED ON THE COMBINATION OF CONVOLUTION NEURAL NETWORK AND CORRECTION NETWORK

使用仿射变换网络对遥感图像进行空间变换,批量生成训练图像,将特征提取和匹配放在卷积神经网络的端到端架构中,直接预测仿射变换参数;通过采用校正网络对卷积神经网络的结果进行改进,实现遥感图像更加精确的配准。通过与SIFT算法、SURF算法和其他深度学习方法相比,该方法对遥感图像配准的速度和精度均有显著提升。

This paper used the affine transformation network to spatially transform remote sensing images to generate training images in batches. Feature extraction and matching were placed in the end-to-end architecture of the convolutional neural network, and the affine transformation parameters were predicted directly. The results of the convolutional neural network were improved with a correction network to achieve more accurate registration of remote sensing images. The registration results show that compared with SIFT algorithm, SURF algorithm and other deep learning methods, this algorithm can significantly improve the speed and accuracy of remote sensing image registration.