Research On Scale Invariant Feature And Its Application In Image Registration

Image registration is the fundamental and critical task in a lot of fields such as digitalimage processing and computer vision. For the sake of image fusion in some practicalapplication, image registration of different images including a same scene which wasacquired by different equipment in different time and different viewing angle is oftennecessary. Therefore image registration is widely researched and used in engineeringapplication such as object identification, image fusion, three-dimensional reconstruction,medical image processing etc. Because of the excellent performances such as highprecision, high EFT, strong robustness, wide using range of application, image registrationbased on image feature play a very important role in Existing methods of image registration.An effective and precise extraction method of image features is very important forimage registration based on feature. At present, Scale Invariant Feature Transform(SIFT) isa very effective feature detection algorithm for its outstanding performance in imagerotation, image scaling, zooming, illumination change, noise, affine, viewing angle etc.Therefore, SIFT algorithm is certainly considered to extract the feature points for imageregistration. However, SIFT algorithm have to create a lot of scale spaces of image whichwas used to detect feature points. What’s more, a great many of convolutions andoperations of weighted histogram have been computed in this process and a large numberof floating-point calculations has to be computed when matching feature point which isdescribed by128-dimensional vector. As a result, the time cost of image registration in bigresolution image becomes to be unacceptable. Aiming at the high time cost of feature pointmatching in SIFT, two improved algorithm have been proposed, one is called a new SIFTmatching algorithm based on distance to reference point, DRP-SIFT for short, and the otherone is called a new fast algorithm of self-adaptive search scope for SIFT matching,AutoARV&DP-SIFT for short. In term of DRP-SIFT, Firstly, computing the distancebetween every feature point which should be matched with the reference point one by one,then ordering the distance results and save it as Dist_Order. Secondly, computing thedistance between query feature point with the reference point, then searching the nearestneighbor of this distance in the Dist_Order with binary search and returning theIndex_Center. Finally, searching the nearest neighbor of query feature point one by one in acertain range whose center is Index_Center. In term of AutoARV&DP-SIFT, Firstly, an appropriate reference vector is computed based on the feature vectors set. Secondly, aself-adaptive search scope is determined by this reference vector. Finally, the matching ofSIFT is executed in a small feature vectors set which is filtered by norm.In the end, a lot of experiments have been done to verify the effectiveness of newimproved algorithms. Experimental Results show that both DRP-SIFT algorithm andAutoARV&DP-SIFT algorithm can decrease the time cost of SIFT feature points matchingeffectively with no loss of matching performances compared with the classical BBF searchalgorithm.