Remote-sensing Registration Study Based On Affine-invariant Features

In recent years,the acquisition of spatial data such as remote sensing,sensor networks,and drones has made revolutionary progress.Additionally,the memory price drops significantly.These changes give rise to the objective need to obtain knowledge from spatial data.Collecting data of the same geographical scene with different sensors at different times by different viewpoints can effectively expand the scope of observations and improve the accuracy of the object recognition.Spatial data registration is a key technology for practical applications such as spatial data splicing,deformation detection of target area,and multi-source target data fusion.Remote sensing image data registration is an important branch of spatial data registration,which is the process of unifying image data of the same scene measured by multiple sets of different coordinate systems into the same coordinate system.However,due to the diversity of landscape scenes,combined with the diversity of image acquisition sensors,acquisition methods and acquisition perspectives,there is no universal registration method that can solve all remote sensing image registration requirements.In addition,the registration process still faces the problems such as large affine variation between images,high similarity of the internal scene texture,and small overlap between image pairs to be registered,which seriously affect the accuracy and efficiency of registration.In this paper,the study is conducted to explore solutions to the main problems in the registration of large-scale affine transformation remote sensing images,based on the research into affine transformation physical models between images and the basic principles and implementation methods of classical image feature matching.The main work of the thesis is as follows:(1)On the basis of defining the target of affine transformation image registration,the affine transformation category between different viewpoints and perspective images is studied.In this framework,the affine transformation relations and geometric properties,the feasibility of the extraction of affine invariant feature primitives and the method of feature matching based on invariant feature primitives are deeply analyzed,(2)A set of feature-based matching quantitative analysis and evaluation models are proposed for objective measurement of the effect of registration.Besides,a calculation method of evaluating indicator is designed.The mismatch rate and under-match rate of the indicator are based on the number of matching features.The matching score is a quantitative evaluation of the registration effect from the perspective of the matching quality of feature elements.(3)With the foundation of the study of classical feature point's extraction theory and methods,a neighborhood partition algorithm based on extended centroid affine estimation is proposed to determine the neighborhood points' range of feature description of affine transformation image to solve the problem of poor consistency in description of feature points with the same name in large affine transformation images.This algorithm is combined with SIFT feature description.Compared with SIFT and SURF methods,it is verified that the method can not only improve the distinguishability between feature points,but also have certain advantages over anti-affine transformation in image registration applications.(4)Based on the study of classical feature extraction theory and methods,this paper analyzes the overlap coverage of feature regions deeply and proposes a model for measuring overlap coverage of feature regions.The experiment analysis demonstrates that the higher the degree of overlapping coverage between feature regions,the worse the distinguishability of feature regions,which leads to a decrease in the effective utilization rate of the primitives and the accuracy of feature matching.Through analyzing the experiment data,it is proved that the accuracy of the feature region extraction and the same-name feature matching will decrease with the increase of the affine transform.(5)For the problems existing in(4),an affine invariant feature region overlap optimization algorithm is proposed.Combined with the classic MSER feature region extraction method,the proposed method could effectively reduce the MSER feature area overlap coverage.Experiments have verified the effectiveness of the optimization method.Moreover,through the means of reasonably reducing the overlapping coverage feature area,the distinguishability of the feature area and the ability to resist affine transformation can be effectively improved.(6)Based on the study of multi-scale auto-convolution map feature description and relative improvement methods,a local multi-scale auto-convolution normalized histogram(LMSA_H)constructor is proposed to extract affine invariant feature descriptors of feature regions.This paper proposes a LMSA_H feature description algorithm based on the region of overlap reduction optimization combined with the MSER features.The experiment verifies that this algorithm can further improve the distinguishability and anti-affine transform ability of the feature region.The algorithms and models of feature points and regions extraction and description proposed in this paper demonstrates important application value in image registration and target recognition fields.In the last chapter of this paper,the algorithms proposed in this paper are summarized thoroughly that the feature point extraction algorithm based on extended neighborhood centroid estimation for extended centroids,the feature region detection algorithm based on overlap reduction optimization,and the feature description algorithm based on LMSA_H.Additionally,the advantages and disadvantages and applicable occasions of each method are analyzed to provide reference suggestions for actual remote sensing image registration applications.