Research On Image Matching Technology Based On Feature Description

Image matching is an important technology in the field of computer vision,pattern recognition and image processing,which develops rapidly.At present,it is widely used in image stitching,image retrieval,3D reconstruction,moving objects tracking and object recognition,etc.This thesis focuses on the existing problems of image matching technology based on feature,and does research on image local feature descriptor and image matching technology.It summarizes existing research results,and analyzes the advantages and disadvantages of tradtional algorithms.Based on this,the improved algorithms are proposed.The main contents and contributions of this thesis are as follows:(1)To avoid the disadvantages of high computational cost and large memory cost from SIFT,and the less discriminative power and robustness from binary-valued descriptors such as BRIEF,ORB,BRISK and FREAK,a novel binary local feature descriptor based on SIFT is proposed.Extensive experiments on large databases demonstrate the strong discriminative power and robustness of our quantization methods.In addition,the proposed algorithm reduces the memory cost and computation cost.(2)In the coarse matching process,the descriptor is divided into two parts according to how much information the pixels around keypoint contribute to it.Then,they are matched respectively.Some wrong matching point pairs can be removed through initial matching.And,the matching time is reduced.(3)In the refinement of key points,to avoid the problem of the RANSAC algorithm,image matching algorithm based on Delaunay triangulation is proposed.In this algorithm,the convex hull of the matching points in the coarse matching result is made a plane triangulation,and Photographic invariant of Photographic geometry is used to refine.According to the characteristics of the data set,the Delaunay triangulation takes into account the spatial location information of the key points,and the constructed Delaunay triangulation is unique.Extensive experiments on ZuBuD dataset demonstrate that the proposed method has higer matching accuracy and less time-consuming than the RANSAC algorithm.