Application Research of Convolutional Neural Network in Remote Sensing Image Registration

In order to solve the problems of large workload of feature extraction and inaccurate feature matching when traditional image registration algorithms are used in remote sensing image processing, we combine affine transformation network, bilinear interpolation and Convolutional Neural Network to to meet the needs of remote sensing image registration accuracy in practical engineering. Firstly, the reference image is enlarged by designed affine transformation network; then, the convolution neural network extracts different features of reference image and training image, learns mapping relationship between images by labels and predict the transformation relationship between reference image and sensed image directly. The experimental results show that compared with the widely accepted SIFT and SURF, our proposal can significantly improve the accuracy of remote sensing image registration.