Contents I. Introduction
In recent years, image registration has become an extremely important technique. It plays an important role in the fields of computer vision, pattern recognition, environmental monitoring, medical image analysis and remote sensing. Precise registration of remote sensing images is widely used in important fields such as society and science. The characteristics of remote sensing images are large field of view, complex terrain, low resolution, and low accuracy of registration methods, etc., which provide more areas of interest and bring more complex background information, making remote sensing images. It is extremely difficult to carry out the registration of. At this stage, how to find accurate registration techniques and efficient methods is a hot issue in the research field. At present, remote sensing image registration technology mainly includes traditional methods of artificial design and methods based on deep learning. The traditional method is that researchers manually design targeted solutions according to the problems existing in specific remote sensing images. However, this method has poor generalization, high labor cost, and low running time. With the rise of deep learning methods[1], more and more scholars have begun to try to introduce neural network methods to study remote sensing image registration. In 2017, Rocco[2] proposed an end-to-end neural network structure for image registration. The pre-trained residual network is used to extract feature information, the correlation between features is established by finding the cross-correlation relationship between features, and the transformation parameters are obtained by regression, and the final registration result is obtained after two-step transformation. This method utilizes the characteristics of neural network adaptive learning, reduces a lot of labor cost overhead, and obtains an efficient registration method. Later, Yang proposed a registration method that combines deep learning with traditional methods. First, the pre-trained network is used to extract image features, and then traditional methods are used to find the relationship between features, and finally the registration results are obtained. However, this method has great limitations, and the running time is long, and the timeliness is not high. In response to the above problems, Kim [3] proposed an improved method based on end-to-end network. Considering the authenticity of remote sensing images, Rocco’s use of Thin Plate Spline(TPS) transformation will distort the remote sensing images and lose the information meaning of the original images. Kim uses Rotation transformation instead of TPS transformation to finely classify the rotation angle, obtain transformation parameters according to the angle range, and obtain the registration result. In order to prevent remote sensing images from being affected by some nonlinear factors such as illumination, time, white fog, etc., Pearson correlation is used to improve the cross-correlation relationship and improve the matching accuracy. Considering the improvement of the learning ability of the neural network, in 2020, Park[5] proposed a dual-stream bidirectional network for remote sensing image registration. By adding an input branch, the target image is subjected to random color jitter transformation and input to the neural network for learning. The source image and the target image, the source image and the target image after random color dithering are two-stream bidirectional registration channels, and the loss function is improved at the same time, and the optimal registration result is obtained. Inspired by the above method, the present invention explores a registration method with higher accuracy and faster operation, aiming at the accuracy of registration and the timeliness of network operation.
