The Spectral Feature Matching Algorithm Based On Distance Measurement

Image matching technology is the premise and foundation of many research fields.Its research results are widely applied in image registration,image classification and retrieval,object detection,stereo vision and 3D reconstruction.Over the years,many scholars have studied image matching algorithms from various aspects.As an effective mathematical description tool,graph can clearly express the structure of the image and can also preserve the relationship between regions.The structure of the graph is based on the relative relation of the neighborhood information,and the spectral features constructed based on graph are therefore robust to linear brightness change and monotonous brightness change.Therefore,the spectral features have been widely applied in image matching in recent years.This paper mainly studies the spectral feature matching algorithm based on distance measurement.The main research content and research results are as follows:(1)The image matching algorithm based on Euclidean distance spectral feature is proposed in this paper.In order to improve the accuracy and robustness of the image matching algorithm based on spectral features,the algorithm first uses the neighborhood information of the feature points to extract the spectral features which can describe of feature point properties.Then,structure attribute relation graph,in which the nodes is the feature points with described by the spectral features and the edges is the Euclidean distance between the spectral features of the nodes,transform image matching problem into graph matching problem.Finally,the maximum pool matching strategy was introduced to obtain the results of graph matching.The experimental results show that the proposed algorithm improves the accuracy of the matching algorithm based on spectral features and has a high robustness to the outliers.(2)The image matching algorithm based on Mahalanobis distance spectral feature is proposed in this paper.For Euclidean distance measurement,it can not fairly reflect the potential relationship between the dimension components of the data sample,which leads to the poor matching accuracy and stability.The proposed algorithm uses mahalanobis distance instead of the Euclidean distance to measure.Firstly,using Mahalanobis distance to construct a local undirected weighted graph on sub point sets.And the singular value decomposition of the adjacency matrix of the graph is performed,the Mahalanobis distance spectral feature is constructed by the spectral value vector.Then,the matching matrix is constructed based on the Mahalanobis distance spectral feature and the greedy algorithm was introduced to obtain the matching results.Finally,in order to further improve the accuracy of the proposed algorithm,we use the SVM method to eliminate the error matching points.The experimental results show that the algorithm improves the accuracy of spectral feature matching,and the robustness of the matching algorithm is higher when there is a outliers in the image.(3)The image matching algorithm based on Hyperbolic distance spectral feature is proposed in this paper.Although the matching algorithm based on mahalanobis distance improves the matching accuracy,but the limitations of Mahalanobis metric linear transformation can not describe the potential nonlinear relationship of the sample,and limit the scope of its practical application.In order to further broaden the application of matching algorithm,the hyperbolic geometry with better adaptability to the sample data was introduced.A hyperbolic metric matrix is defined according to the statistical properties of the data,and the hyperbolic distance measurement is given.For each feature point to be matched,the sub point sets were selected according to the hyperbolic distance of other feature points to the feature point,and the undirected weighted graph was constructed.Then the hyperbolic distance metric was used to compute the adjacency matrix of the graph,and the hyperbolic distance spectral feature of the point was obtained by spectral decomposition.Finally,the matching mathematical model is established according to the similarity between the hyperbolic distance spectral features and the position relation between the feature points.The greedy algorithm was introduced to obtain the matching results.The experimental results show that the algorithm improves the accuracy and robustness of the spectral feature matching algorithm and is suitable for more image changes.