Image Registration And Mosaic Techniques Based On Feature Points

Image registration technique is a major part of computer version and is widely used invirtual reality、defense security and remote sensing. Along with the application, find a betterway to extract the points、increase the precision of registration and reduce the work time are themain questions to be solved. In this paper, we study the basic theory of image mosaic based onfeature points and it can be divided as follows:(1) Introduce the key techniques of feature-based mosaic systematically, including featureextraction and work out the descriptor、calculation of the homography matrix and image fusion.We observe the result of linear matching between a set of data obtained by RANSAC andLMedS respectively and find that LMedS can reject the wrong pairs better when the number ofinlines is more than a half.(2) Aimed at the susceptible with scale change of Harris corner, we present anauto-adjusted algorithm of image size based on phase-correlation. First, we detect the zoomrelationship and translation factors between the images and then modulate the unregistratedimage’s scale to the same level as the original image. While easy to location the key points andinheriting the fast speed of Harris，the scale difference between the images could enlarge from1.8to4.7and can eventually obtain a clear and stable mosaic result.(3) Considering that the traditional registration methods are almost rely on thetime-consuming descriptor such as SIFT and SURF, we discussed a binary descriptor based onFAST and BRIEF which called ORB. A main direction is added to FAST operator and thecorrelation of BRIEF is limited about0.5, so ORB is rotation invariant and resistant to noise. Wedemonstrate through experiments how ORB is at two orders of magnitude faster than SIFT, and make it better to adapt the scale difference of images.(4) We propose a semi-automatic way to mosaic images which have a obvious plane fall.First make a division among the different planes to get several ROI, then extract key points andmake registration independently. The main advantage is that we can exploit the self-takeinformation in images and can adopt the parallel mechanism of computers to cut the time down.In this paper, a6to10times of speed-up testified the validity of the algorithm while using SIFTto get the feature descriptors.