Study On Integrated Complementary Invariant Feature Multi-source Remote Sensing Image Registration Methods

Remote sensing image registration is not only the key technology using in thefield of imaging mosaic, changing detection, information fusion, target recognitionand tracking, but also one of the critical steps for a variety of remote sensing imageanalysis purposes including weather forecast, map updates, and so on, whose maintask is to achieve the aligned position in space of the image data from one objectwhich obtained in different time phase, at different angles or from different sensors.Currently, feature-based registration method is one of the registration technologyresearch focuses. This thesis surveys and analyses the existing feature-based imageregistration techniques, and emphasis is given to study the integrating multiplefeatures multi-source remote sensing image registration problems.This thesis comprehensively analyzes various factors of affecting imageregistration which includes the existing local invariant feature extraction anddescription algorithm, matching and searching strategy as well as matchingoptimization extraction algorithm, then two integration feature registration algorithmswhich respectively adapt the optical remote sensing image and SAR image areproposed, and a number of actual remote sensing images are used to experiments inorder to verify the correctness of the proposed methods. The research achievementsare as follows:(1)In the process of studying local invariant feature extraction and descriptionalgorithm, the thesis found that popular SIFT algorithm consumes too long time andit’s worse in terms of its affine invariance, in order to resolve this problem, the thesispropose a new method to save time, enhance affine invariance and improve theregistration accuracy, which reduce the dimension of SIFT descriptor firstly to savethe time of constructing descriptor, then combined the affine invariance of affineinvariant moment to improve the affine invariance of descriptor, finally achieved theexcepted result.(2) Due to the limitations of the single feature, the thesis propose a newregistration algorithm based on integrating MSER and SIFT-AIM complementaryinvariant feature through comprehensive analysis and discussion of the affectingprocess between various feature properties and feature registration algorithm. Thenew registration algorithm can be used to eliminate or reduce the registration errorcaused by the limitation of single feature through two steps from coarse registration to fine registration. Firstly, the method take coarse registration using MSER feature, andthe spatial geometric deformation was corrected; then image affine invariance wasimproved and image registration time was saved through further fine registrationusing SIFT-AIM feature, finally, enhanced image registration was achieved.(3)After comprehensive analysis of various feature properties, thesis find it’salways failed when the algorithm that is successful in optical image registration wasused to deal with SAR image registration. To address this problem, thesis proposed aregistration method integrating Canny edge and SIFT complementary invariantfeature. To begin with, the registration algorithm split segment using canny edge, andcoarse registration was done utilizing split segment, so the spatial deformation of SARimage was corrected initially. Lastly, the fine registration was achieved through thestable SIFT feature points extracted from canny edge.The achievement in this thesis can enrich the relative theories and methods of theremote sensing image feature extraction and registration, and establishes a goodfoundation for the follow-up of multi-source remote sensing image fusion.