Research On Infrared And Visible Image Registration

Infrared and visible light image registration can achieve geometric alignment of the two images and fully extract the complementary information of the two images,which is widely used in military reconnaissance,video surveillance,intelligent driving and other fields.Although the infrared and visible light image registration technology has made great progress in various fields in recent years,due to the uncertain relationship between the imaging resolution and pixel gray value between the infrared and visible light images,how to improve the registration accuracy of the two still remains.It is a technical difficulty and has important research value.In this paper,the feature-based infrared and visible image registration is studied for feature description,feature matching and geometric transformation parameter estimation.The main research work is as follows:(1)Aiming at the problems of poor robustness of existing feature descriptors and low effective matching ratio of infrared and visible light images,a multi-level matching method based on CR＿SIFT and offset consistency is proposed.Firstly,based on the relative stability of the contour between infrared and visible light images,the method uses the Canny algorithm to obtain the respective edge contour maps,and performs SIFT feature point detection and localization on the contour maps.Then the centroid method and concentric circles are introduced to construct their own feature descriptors for each feature point.Finally,based on the similarity of the descriptors and the offset consistency between the correct matching point pairs,the constraints on the matching point pair set are enhanced,and the obvious wrong matching is removed through multi-level matching,and the final correct matching point pair set is obtained.(2)Aiming at the problem that it is difficult to construct the correct geometric correspondence between infrared and visible light images,a registration method based on multi-level matching and Gaussian mixture model is proposed.Based on the set of matching point pairs obtained by the multi-level matching method,the geometric transformation model between images is firstly constructed,and the joint probability density of the matching point pairs is calculated by using the Gaussian mixture model.Then the parameters of the geometric transformation model of infrared and visible light images are iteratively solved by Bayesian criterion and expectation maximization algorithm.Finally,the alignment and registration of infrared and visible light images are completed through geometric transformation and linear interpolation.Experimental results show that the CR＿SIFT descriptor has good robustness in changing scenes,and the number of effective matches obtained by the multi-level matching method increases significantly.The registration method based on multi-level matching and Gaussian mixture model has higher performance than several other advanced methods,the matching accuracy can be maintained at about 90%,and the positioning error is maintained at about 1 pixel,which can achieve infrared and visible light images.Subpixel level registration.