Image Tie Point Matching Based On Graph Theory

Image interpretation is one of the most important research fields in remote sensing and photogrammetry,and image matching is the prerequisite for image interpretation.Recently,image matching is deeply researched and some of the matching algorithms(e.g.,Scale-invariant feature transform and speeded up robust features)are successfully applied in commercial softwares.More and more applications of big data are in the ascendant,as a typical big data,remote sensing images are also included in the tide.Image matching,which is a basic technology for image processing,is challenged by multi-sorce remote sensing images because of radiometric distortions between image pairs.Besides,the observation targets of remote sensing satellite and aerial camera shift from densely populated urban areas to depopulated zone(e.g.,desert,forest,and Gobi and et.al).The image textures of these areas are characterized with repetitive patterns and lack of contrast,which is known as poor textures.Poor textural image matching is another challenging task except for multi-source image matching.As the postprocessing of matching,matching bulenders detection is also challenged by complicated conditions.Althogh some detection algorithms,such as random sample consensus and reweighted least squareare,are cabable of this task for detecting most of the gross errors,whereas there are new traits of poor textural and multi-sensor image matching results(e.g.,the number and rate of gross errors are much more than ever before),traditional used detecting methods sometimes failed in the detection.At last,although numberous matching algorithms have good performance in matching,there are seldom research for evaluating the matching result in the view of reconstruction of geometric structure.To solve these challenges,this paper proposes a graph based image matching algorithm aiming to solve the problems in matching poor textural and multi-sensor image,a graph based matching blunder detection method is also proposed for gross error detection.These two methods are evalutated by the proposed criterias by comparing with traditional matching algorithms.The main contributions and innovations of this paper are summarized as follows:(1)The mainstream matching methods are sysmaticly introducted in this paper.These algorthms are summarized into four levels:point feature matching(bold and corner feature),line feature matching,region feature matching and relational matching.For every matching level,representative matching methods are introduced and gave a detailed merits and demerits of these methods.(2)A systematic evulation criterions are proposed in this paper to quantify the matching result.The ultimate purpose of image matching is to retrieve geometric information of images,thus positional accuracy is the lifeblood of image matching quality.This paper considers every steps of matching and their effects on positional accuracy,and two main parts of matching quality are included,that is,the quality of point feature,and the quality of feature descriptor.The qualities of point feature consist of four parameters,efficiency,positional distribution,repitive rate,and locational accuracy;the qualities of feature descriptors are matching recall,matching precision,positional accuracy and positional accuracy.Some of the mainstream matching methods are evaluated by use these criterions,which give directions of pratical use in design of matching method.(3)A graph based image matching algorithm is also proposed in this paper.Firstly,corner feature detector algorithm is applied in extracting evenly distributed image featue points;then the approximate neighbor searching algorithm is applied for giving constraints on the number of feature points in image pairs,all the features are treated as graph nodes and an affinity tensor is built between the two graphs with its elements represent the similariy of feature triplet.At last,the tensor power iteration algorithm is used to obtain the initial correspondences of node pairs,and the final matching of feature points are acquired by discretization of matrix.The experimental results show that i the avarage matching recall in poor textural image matching is 40%;and in multi-sensor images,the matching recall is 50%,and other criterions are also improved compared to traditional feature matching methods.(4)The matching bulender detection is also included in the framework of graph theory.The geometric invariant,which is the critical factor in gross error detection,can be easily exploited in affinity tensor because the geometric and radiometric information is all included in the tensor.Inspired by data snooping algorithm,this paper proposes a graph based gross error detection method,this method is an iterative process and affinity tensor is used in the process.The matching points are treated as two attributed graphs;every graph node represents the support from other nodes.In every iterative step,the graph node with the smallest support is eliminated,the iteration stops until all the induced graph keeps stable.When applied in real data experiments,compared to traditional used random sample consensus algorithm,the graph based algorithm has higher accuracy and lower loss.And the synthetical data experiments show that the detection accuracy is 100%,and the false rate is only 3.5%.Based on the mentioned above research,this paper proposes a graph based algorithm,and compares the proprosed algorithm to state-of-the-art algorithms such as SIFT?SURF?ORB?BRISK?LIOP?MROGH?NG-SIFT?OR-SIFT and GOM-SIFT.