Research On Image Feature Points Matching Algorithm Based On SIFT

Image registration is a fundamental problem in the field of digital image processing.It is used to compare or fuse images acquired under different conditions for the same object,including object or scene recognition,3D modeling according to multiple images,stereo matching,and motion tracking,etc.It is the important step in the process of the image from getting into the analysis.Image registration technology is widely used in many fields,such as machine vision,medical image processing,military field and material mechanics.This paper mainly studies the image feature point matching algorithm based on SIFT.The feature points extracted by the SIFT algorithm have no tolerance to the image similarity transformation,and have high robustness to light,noise and micro angle change.However,the SIFT feature point descriptor has the disadvantages of complicated computation,high dimensionality and slow matching speed.In this paper,we propose an approximate search algorithm based on vector inner product.Firstly,introduce the appropriate reference vector with the dynamic change of the point to be retrieved.The inner product of the sample set and the reference vector is obtained by directly extracting the data in the feature point description matrix and the inner product is ordered.Secondly,calculate the inner product of the point to be retrieved and the reference vector,find the position in the corresponding inner product sequence and determine the approximate search range centered on the position.Using the approximate nearest neighbor search in the range instead of the full range search in the sample set,the search speed is greatly increased at the expense of a smaller amount of accuracy.To verify the effectiveness and advantages of this algorithm,the algorithm and the KD-BBF algorithm of the approximate nearest neighbor search are used to test a image test library of Oxford University VGG Lab.And the performance of the two algorithms is evaluated in terms of matching speed,correct matching quantity,and accuracy.The experimental results show that the selection of the appropriate approximate search range can make the proposed algorithm improve the search speed compared with the KD-BBF algorithm,but only lose a small number of correct matches and the accuracy.Then,we use the algorithm proposed in this paper to carry out the feature matching and image fusion experiment.A good fusion result is obtained in the case of two images with large amplitude translation,rotation and scaling transformations,and a small amount of affine and projection transformation.