The Research Of Image Registration Method Based On SIFT Algorithm

As an important part of image processing technology,image registration technology is widely applied into many areas,such as computer vision,remote sensing imaging,pattern recognition,medical image processing and so on,and it's worth studying.The image registration algorithms based on feature points have received extensive attention because of their less computational time and high matching accuracy,of which the SIFT algorithm is the most significant one.Due to its invariance to image translation,zooming,changes in illumination,rotations and so on,SIFT algorithm has attracted a lot of attention and even becomes the standard of judging the performance of an algorithm.As the requirements of practical applications for the performance of image registration algorithms become higher and higher,the traditional SIFT algorithm could not meet it any more and the research on how to achieve high precision registration in a shorter time is of great significance.In this thesis,two improvements have been made on SIFT algorithm.First of all,to deal with the problem that there exit many false matches in SIFT algorithm,we have proposed an improved SIFT algorithm based on two-step purification and it can effectively improve the matching performance of SIFT algorithm through eliminating the false matches.Secondly,to overcome the time-consuming problem of SIFT algorithm,we have put forward a parallel processing method which will reduce the running time of SIFT algorithm through improving the parallelism of it.The major research work of this thesis includes the following respects:Firstly,we have made improvements on the symmetric Kullback-Leibler divergence algorithm.There is no clear boundary of the score between correct feature matches and the wrong ones in the traditional symmetric Kullback-Leibler divergence algorithm.Thus,it would unavoidably eliminate the correct feature matches when we use the traditional symmetric Kullback-Leibler divergence algorithm to refine the feature matches.When the initial feature matches are not too many,it will lead to the problem that the number of feature matches will be too small after purification because of the excessive purification.The improved symmetric Kullback-Leibler divergence algorithm in this thesis can make corresponding processing according to the number of initial feature matches.The intensity of purification of the improved algorithm will be reduced when the number of initial feature matches is small.As a result,enough feature matches will be kept.On the contrary,when the number of initial feature matches is large,the intensity of purification of the improved algorithm will be increased.Thus,the number of feature matches after purification will be smaller and the operating efficiency of the next stage of the SIFT algorithm will be improved.Secondly,to deal with the problem that there exit many false matches in SIFT algorithm,we have proposed an improved SIFT algorithm based on two-step purification.It was proved by the experiments that the improved SIFT algorithm could effectively eliminate the false matches and improved the matching accuracy of SIFT algorithm.Thirdly,to overcome the time-consuming problem of the SIFT algorithm,we have proposed a parallel processing method.We have done much parallel processing on SIFT algorithm in multiple levels and multiple stages.As a result,the efficiency of SIFT algorithm has been improved.In this thesis,we have designed comparative experiments for every improved algorithm and the effectiveness of every improved algorithm has been proved.The research work of this thesis has important theoretical significance and practical value.