Muti-source Remote Sensing Image Registration Based On Convolutional Neural Network And SIFT

Remote sensing image registration aims to estimate an optimal geometric transformation between two images of the same scene captured at different time,different viewpoints,and with different sensors.Remote sensing image registration is widely used in sensor proofreading,data fusion,earth resource detection,and so on.Meanwhile,it is an indispensable step in solving many problems such as target recognition and change detection,image fusion,and scene reconstruction.In recent years,the study of deep learning continues to heat up.Convolutional neural net-work(CNN)has achieved great success in visual tasks such as image retrieval and image classification.The CNN feature has become an important image features.In this thesis,we try to apply CNN features to remote sensing image registration.We first study the use of different CNN features for image registration and compare it with the traditional SIFT features.On this basis,CNN features and SIFT features are merged to form new features for image registration to further improve the accuracy of image registration.The advantages of CNN features in remote sensing image registration are verified by experiments.The fusion feature obtains more correct matching points in the feature matching stage of image registration,so that the multi-source remote sensing images registration can be achieved better.This thesis mainly completes the following content:1.Summarize the development of the current image registration technology,expound the traditional image registration process,and the future development orientation of image registration technology is prospected.2.The basic theoretical knowledge of image registration technology is introduced,and the theoretical knowledge and principle derivation of convolutional neural network are analyzed.3.We use convolutional neural network to extract image features and apply them to image registration.The performance of CNN features and SIFT features are compared,which reflects the advantages of CNN features in image registration.4.Combine the CNN features with SIFT features and import the fusion feature to the improved SIFT algorithm to implement the image registration.And compare it with some improved SIFT algorithms by experiments,it shows that the fusion feature can improve the accuracy of image registration and achieve better image registration.