Research On Keys Techinques Of Multimodal Remote Sensing Image Registration

With the rapid development of sensor technology,the amount of multimodal remote sensing images are exploding.The resolutions of these multimodal are also growing a lot.The fusion and analysis of multimodal remote sensing images are becoming a hot research direction but needs much further study work.As the first and fundamental step of fusion and analysis of multimodal images,multimodal remote sensing image registration provides profound foundation for further object detection,object recognition,and change detection.This dissertation focuses on feature-based registration methods,pixel-based registration methods and automatic registration evaluation methods and presents a indepth study on these topics.With feature-based registration methods,this dissertation first presents a clear discussion on feature types and corresponding matching methods suited for multimodal remote sensing image registration.We give out directions on the types of features and the ways to incorporate those features.Upon this,we presents two registration methods with different kinds of local features:(1)for local features which could be compared quantitively across modalities,we design an uniform objective function which based on different characters of local features and structural features,and obtained closed-form solutions;(2)for local features which could only be compared qualatively across modalities,we creatively use class labels which could be across modalities with robust registration methods to register multimodal images.With pixel-based methods,this dissertation first analyzes the theory foundation of those methods,and discusses the working part and deficient part of those methods.Base on mutual information methods,we present two registration methods:(1)in order to ensure the contents to be registered are the same and incorporate image features,we present labeled mutual information.Mutual information is only conducted on image contents with same labels,which effectively avoids non-injective mapping and incorporates spatial structural information;(2)we present region growth and model growth methods – dual growth mutual information,which solves problems of traditional mutual information like the slow convergence rate and likely-falling into local maximums.The proposed method could improve registration success rate and accuracy and decrease computation load at the same time.With automatic image registration evaluation methods,this dissertation starts from parameter estimation,and lays out the theory foundation of uncertainty estimation of feature-based registration methods.By incorporating Bootstrap resampling methods,we could implement the uncertainty estimation.Through experiments the correctness of the proposed method was demonstrated and the method was used to study the impact of different registration elements.Another presentation of this chapter is a rectified registration consistency.After theoretically analyzing the deficiency and impact of original registration consistency,by incorporating a new item,the proposed method could obtain an accurate and automatic evaluation of registration methods without increasing the computation load.The proposed methods and models are experimented with lots of real multimodal remote sensing images and automatic registration evaluations.The feasibilities are soundly proved.