Research On Remote Sensing Image Matching Method Based On Local Invariant Features

In recent years,local invariant features in images have been widely used in image feature matching algorithms due to their high discrimination,strong invariance and good robustness.However,the current local invariant feature matching algorithm will affect the matching accuracy due to the transformation of the image,such as geometry,illumination,and perspective.Therefore,it is of great significance to study more effective local invariant feature matching algorithms.To this end,this paper firstly analyzes and summarizes the existing local invariant feature matching algorithms,on this basis,the image matching method based on the fusion of the improved ORB(Oriented Fast and Rotated Brief)algorithm and the matching strategy and the image matching method based on the improved Radiation-variation Insensitive Feature Transform(RIFT)are studied.The improved ORB algorithm and the improved RIFT algorithm are tested on Mikolajczyk dataset,UAV image data and multimodal remote sensing image dataset.The main research work and conclusions are as follows:(1)A remote sensing image matching method based on the fusion of improved ORB algorithm and matching strategy is studied.Aiming at the problem that ORB algorithm has no scale information and low matching accuracy in image matching,firstly extract feature points in multi-scale space,use gray centroid method to find the direction of feature points,and construct Root SIFT descriptor in the feature description stage to improve feature points stability;then combine the Flann and bidirectional matching strategies to perform feature matching,and use cosine similarity constraint matching to obtain high-quality matching points;finally,the Random Sampling Consistency Algorithm is used to eliminate false matches,and the algorithm is tested by standard datasets and aerial images.The results show that the improved ORB algorithm shows good performance in image scale,rotation and illumination changes.Compared with the traditional ORB algorithm,the matching accuracy rate is increased by about13%,and the matching accuracy is about doubled.(2)A remote sensing image matching method based on improved RIFT algorithm is studied.Aiming at the problems of high computational cost,long time-consuming and small number of feature points after image rotation in the RIFT algorithm,this paper firstly obtains corner and edge point features based on the phase consistency map;then the maximum index map constructed by the Log-Gabor convolution sequence is normalized,and an affine circular region shape descriptor is constructed for feature description to improve the matching efficiency;finally,NBCS(Normalized Barycentric Coordinate System)is used to remove gross errors to improve the matching accuracy.The results show that the improved RIFT algorithm is robust to nonlinear radiation differences,and the computational efficiency is approximately 2 times higher than that of the RIFT algorithm in terms of accuracy.With the increasing rotation angle,the average number of correct matching points of the improved RIFT algorithm is slightly higher than that of the RIFT algorithm by 2,and the distribution of feature points is more uniform,which can better complete the multi-modal remote sensing image matching task.There are 46 figures,3 tables and 80 references.