The Research Of Composite Registration Technology Based On Features Of Multi-sensor Images

Due to different imaging mechanism and imaging conditions of visible image,infraredimage and SAR image,the characteristics of different images are obviously distinct,includinggray distribution,statistical properties and noise,so registration methods based on intensity donot well achieve registration of different sensor image.Multi-sensor images registration isrealized mainly by using registration method based on feature.In the spatial domain,although gray difference of multi-sensor images is prominent,theedges of the targets of two images are relatively stable,and remain unchanged,and possesscorrelation.In the transform domain,low frequency image is the approximate image of thesource image,and describes the basic trend of the source image,and reduces the influence ofnoise of detail,so that low frequency images of two images show the similarity.Highfrequency image reflects the details of the source image,and the feature points extracted inhigh frequency image have the very good robustness.Therefore,on the basis of the analysis ofthe existing feature extraction techniques,this paper is orientated towards visible image andinfrared image,visible image and SAR image,studying the composite registration technologyin spatial domain and transform domain.The main contents of this paper are as follows:(1)Multi-source images composite registration method based on edge and SURF featurein spatial domain is proposed.Firstly,SURF feature points are extracted in two Canny edgeimages.Secondly,SURF feature points are matched by using a both way and two levelsmatching method.Then the optimal affine transformation parameters are estimated based onthe accurate matching results.Finally resampling and bilinear interpolation are employed tofinish the registration of different sensors images.Experimental results show the proposedalgorithm has high registration precision and high speed.(2)Multi-source images composite registration method based on sub-images of NSCTand SURF features in transform domain is presented.Firstly,two original images aredecomposed by NSCT,when gray difference of two low frequency images is very big,lowfrequency image of real-time image is inverted.Secondly,SURF features are extracted in lowfrequency images,and edge feature points are extracted in high frequency images.Then SURF features are accurately matched by using a bothway and two levels matchingmethod,and the edge points are matched by normalized cross correlation andRANSAC.Finally, the repeated points of precise matching points which come from the spatialdomain and transform domain are removed,and remaining matching points are used tocalculate affine transformation parameters.The experimental results demonstrate that thismethod has good performance of image registration.(3)Considering respective characteristics of image registration in the spatial domain andtransform domain,and their mutual compensation,a composite registration method combiningthe spatial domain and transform domain is put forward.Firstly,the repeated points of accuratematching points which come from the spatial domain and transform domain areremoved.Then the rest of precise matching points are employed to compute affinetransformation model,and image registration is achieved.The experimental results show theeffectiveness of the proposed algorithm.