Research On Matching Methods Of Infrared And Visible Images Based On The Similarity Of Global Structures

Infrared and visible image matching is an important research topic in the field of computer vision.Infrared and visible image matching has been widely applied in the fields of satellite remote sensing imaging,drone flight navigation,and car safe driving.Matching methods for homologous visible images have been extensively studied and achieve excellent performance under the conditions of varying illuminations,scales and rotations etc.Similar to homologous image matching methods,infrared and visible image matching methods have to solve the problem of feature invariance under different imaging conditions.However,due to the different imaging mechanisms of infrared and visible images,there are large differences in intensities,gradients,and textures between the two types of images.For the matching of infrared and visible images,existing methods still face great challenges in terms of the feature detection repeatability and the feature matching accuracy.In order to improve the performance of infrared and visible image matching methods,this thesis utilizes the similarity of global structures of infrared and visible images to design feature matching methods.By studying point feature matching method,region feature matching method and iterative matching method,this thesis improves the matching performance in terms of the feature detection repeatability and the feature matching accuracy.The main research content of this thesis includes:(1)Feature point match based on the global structure edge and feature point analysis based on structured prediction are studied.For the problems of the low detection repeatability and the low matching accuracy in feature point matching of infrared and visible images,a feature point matching and analysis method is proposed based on the global structure edge.Firstly,according to the similarity of the global structure edge of infrared and visible images,common feature points are extracted by the edge constraint to improve the repeatability of feature points.Secondly,feature point descriptions are established by edge attributes including length and direction to improve the matching accuracy of feature points.Finally,in order to solve the problem of the low discriminability of features caused by the difference of gradients and textures between infrared and visible images,structured prediction method is used to evaluate the discriminability of feature points.Based on the similarity of the global structure edge of infrared and visible images,the proposed method improves the matching performance in terms of the feature detection repeatability and the feature matching accuracy.(2)Region detection based on the edge response and region match based on the self-similar of region shapes are studied.In order to leverage the global structure of images further,a region feature matching method is proposed based on the similarity of shapes of homogeneous regions in infrared and visible images.Firstly,consistent edge responses of infrared and visible images are computed and region features are extracted to solve the problem of the low repeatability caused by the intensity difference of the boundary of homogeneous regions in infrared and visible images.Secondly,according to the similarity of region shapes in infrared and visible images,local self-similar descriptions of region shapes are calculated and the redundant descriptions of smooth patches and low discriminative patches are removed to obtain efficient descriptions.By utilizing the global structure further,the proposed method improves the matching accuracy of infrared and visible images.(3)Feature saliency analysis and iterative match based on feature combination are studied.According to the characteristics that large number and extraction efficiency of feature points and high accuracy and small number of region features,a feature combination iterative matching method is proposed based on saliency analysis.In the initialization stage,the region match is used to estimate the projection transformation to avoid the low matching accuracy and the imprecise transformation model in the feature point match of the iterative closest point method.In the meantime,the concept of matching region is presented to quantify the error of the transformation relationship to make use of the position constraints of feature points.In the iterative matching stage,updating matching regions of feature points,evaluating feature salience measurements and expanding the match set are performed iteratively until the estimation of the image transformation relationship is completed.Experimental results show that the proposed method can estimate the image transformation relationship when the percentage of correct matches is low and the number of correct matches is small.