Abstract: Aiming at the problem that traditional algorithms were not suitable for image mosaic with large changes in appearance, we  proposed a remote sensing image mosaic method based on convolutional neural network, which enabled the model to realize the registration and mosaic of remote sensing images through deep learning. The algorithm was compared with the traditional algorithms through two experiments. Firstly, the Euclidean distance was used as the evaluation indicator, and statistics of two methods were performed on different remote sensing image data sets to evaluate their image registration abilities. Secondly, effects of remote sensing image mosaic realized by two algorithms in the real remote sensing image mosaic application scene were compared. The experimental results show that the convolutional neural network model has a better registration ability for images with large external deformations. Therefore, for the remote sensing images with large changes in appearance, the proposed algorithm can be used to replace the traditional algorithm to realize image mosaic and obtain a more accurate panoramic image.

Key words: satellite remote sensing, convolutional neural network, image mosaic, scale-invariant feature transform (SIFT)