Methodological Research Of Remotely Sensed Imagery Registration Based On Feature Points And Mutual Information

Image registration,the basis and premise of implementing image fusion,temporal series analysis and change detection,an important part of image processing,has been widely used in remote sensing,health service,agriculture,military affairs and other fields.With the advances of sensors,various types of remote sensing data resources become increasingly diverse and the application is more and more wide.However,disparity exists among remote sensing images captured at different times or even at the same times in the same area because of influences from sensor properties,illumination,time and angle of imaging.Registration for remote sensing images is more difficult and also more urgently required compared with that of other types of images.This thesis integrated the method based on feature point and mutual information to explore a new approach of remote sensing image registration,which might deal with the limitation of present avaliable methods and algorithms in accuracy,efficiency and automation,which had a clear significance.The main work of the thesis includes the following aspects:(1)Summarizing the content of image registration and its research status systematically.This thesis described the pedigree of the existing two kinds of mainstream remote sensing image registration methods based on gray scale and feature respectively and the processes,characteristics and limitations of subclasses in each one.(2)Analyzing and comparing properties,advantages and disadvantages of four common feature point extraction operator:Moravec,Harris,SUSAN and SIFT in efficiency,scale invariance,rotation invariance,timeliness et.al.by taking TM image data as experimental data.(3)Proposing a method of removing mismatching points automatically based on affine transformation.This method successively iterates until the set conditions are met by means of eliminating matching paris with errors greater than the set threshold from the results of affine transformation based on the matching points.Experiments show that the method can automatically remove mismatching pairs according to the preset requirements,and improve the registration accuracy.(4)Presenting a remote sensing image registration strategy that combines the feature points and mutual information.The method based on the strategy takes theSIFT-based rough registration result as the initial solution of mutual information searching.Therefore,the new method combines the advantage of high speed from feature-based registration method and that of high accuracy from gray-based registration method.An empirical registration study using three groups of different remote sensing images as the experimental data shows that the results based on the image registration strategy proposed in this thesis is superior to existing similar research achievements.