Remote Sensing Image Registration Method Based On Convolutional Neural Network

The main function of image registration is to achieve the process of geometrical overlaying two or more images of the area at different times, from different viewpoints or by different sensors. Image registration is the foundation problem of Geographic Science, Medical Imaging, Computer Vision et al. In respect of satellite remote sensing image registration areas, Image registration information for the realization of the application of the object recognition, image fusion, reconstruction, and many other scenes, has an indispensable role.Nowadays a large number of remote sensing image data need to be registered, so the traditional manual image registering procedure, can’t taken real-time require when a large number of image need to be registered. So it’s very important to study on automated techniques, and today it has become a very popular research focus. The current image registration algorithm is divided into two categories: based on region and based on image features. This paper uses a matching algorithm based on local features. I use the Convolution neural network to get feature description.And complete image registration between optical remote sensing images. In this paper, the work done with the following specific points:1. Summarizes the current developments in image registration technique and introduce the Image registration process framework, and make a prospect on image registration techniques.2. Introduces the theory of image registration and convolution neural network, and explain the derivation of the convolution neural network.3. Use Maximally Stable Extremal Regions method to extract image sample and construct the appropriate network architecture, at last train this network4. Use convolution neural network to optimize feature description. So the Image’s new feature expression can be matching well at Feature matching stage. At last use optical remote sensing images doing simulation experiment base on this method.