Image Registration And Mosaic For Free Multi-View Scene

With the vigorous development of virtual reality,visualization navigation,lowaltitude aerial photography,video conference & monitoring and so on,the image stitching technology,as the core part,has become one of the research enthusiasts in the field of computer vision.The application constraint of traditional fixed-viewpoint image mosaic is relatively stringent,which requires that the optical axis center of the camera is parallel or intersect at one point.However,the usual way of shooting is freer and the scene image may have plenty planar structures.If the traditional global single transform registrat io n model is used,it will cause serious mismatch and leakage matching,resulting in projection distortion which induces shape distortion in stitched result.Therefore,image registration and mosaic with perspective freedom are studied in this thesis.It mainly studies the detection of scene planes and the refinement of feature matching,the local registration model based on multi-homography transform,and the efficient fusion algorithm of panoramic image.In addition,a method of using wide-angle lens to reduce the hardware cost of panoramic image acquisition is provided,and color correction via white balance and barrel distortion correction of wide-angle lens are deeply studied in pre-process stage.The main work of this thesis is summarized as follows:(1)In the period of pre-process,improved in grid-template calibration method,an adaptive correction method based on polynomial piecewise optimization model is proposed to correct wide angle distortion.The automatic extraction of control point pairs is realized by using the subarea directional search algorithm.Then the optimal distortion center is estimated by point-distance search method.The piecewise optimal error model,which is adapted to fit the global radial distortion,is constructed to achieve high-precis io n correction effect.(2)Aiming at multi-plane structure scene images,a feature matching hierarchica l algorithm based on J-Linkage clustering is proposed.Several improved algorithms for SIFT feature extraction is given to increase the number of matching pairs in large parallax images.Then KNN matching search procedure is accelerated by the K-D tree.Finally,the J-Linkage clustering algorithm is applied to the robust estimation of mult ip le homography transform models.By establishing similar concept space of matching pairs,feature pairs with similar preference sets are continuously clustered,so as to classify and refine feature matching results on different planes,improve matching accuracy,and increase the total number of matching pairs.(3)To solve the projection distortion and dislocation in traditional single transform registration methods,a registration optimization model based on local mesh-projectio n transform and deformation constraint is given.Each image is divided into grid blocks,and Moving DLT transform is used to calculate the local projection matrix.Then the registration error function is established with constraints of position,shape preserving and similarity conditions.Thus,the global registration accuracy is improved,the distortion of projection is reduced,and the original angle of view is retained as much as possible.(4)A multi-resolution image fusion method based on stitching-seam searching is proposed.Firstly,exposure compensation and wave correction are processed,and then graph cut/maximum flow algorithm is used to search for the optimal seam.Finally,mult iresolution Pyramid fusion algorithm is applied to image fusion in overlap regions.This method can effectively eliminate stitching-seam and avert fringe or ghost phenomenon caused by registration errors or moving targets,so that the stitching result is more natural.