Research On Image Registration Algorithm Based On AKAZE Feature

Image registration is an important branch of image processing,which is widely used in various fields,such as computer vision,automatic driving,3D reconstruction,image stitching,augmented reality and so on.With the development of computer vision and image processing technology,the speed and accuracy of image registration is also increasing.The quality of image registration results has a decisive factor for the subsequent image processing.Therefore,image registration is still one of the most important research directions.Because of the different viewpoints,there will be a certain affine relationship between the images.When the images are rotated or affine transformed,AKAZE(A-KAZE)algorithm can not get the correct matching point pairs.In order to solve this problem,this paper proposes an affine invariant AKAZE image registration algorithm,which simulates the affine transformation of the input image in different longitude and latitude directions,then extracts the feature points of the affine transformation image by using AKAZE algorithm,and describes the feature points by using MLDB(Modified-Local Difference Binary)descriptors.In view of the large computation of traditional GMS(Grid-based Motion Statistics for Fast)algorithm,this paper uses a motion statistics algorithm based on the five palace feature score to preliminarily screen the feature point,and then uses the random consistency algorithm to delete the mismatch point.Experiments show that the improved algorithm can effectively get the correct matching points when the image is affine transformed,and complete the image registration while ensuring the matching accuracy.The traditional AKAZE algorithm mainly takes the gray information of the image as the input,and does not take the color information and illumination information of the image into account,resulting in the mismatch of the image in the areas with similar gray levels but different colors.In this paper,we add color compensation and illumination compensation to the traditional gray information,which overcomes the disadvantage that the algorithm does not use the color and illumination information to distinguish the identical point.AKAZE algorithm extracts many redundant feature points while extracting feature points.The existence of redundant points not only affects the efficiency of the algorithm,but also causes mismatches to a large extent.To solve this problem,a redundant point detection and deletion algorithm is proposed.The algorithm uses Euclidean distance and the repetition rate of the feature area as constraints to determine the feature points with high degree of overlap,and then compares the response strength to delete the feature points with small value as redundant points.After matching,the random consistency algorithm is used to eliminate the mismatches,and the whole algorithm is accelerated by parallel operation.Experiments show that the proposed algorithm can not only effectively identify the identical point,but also remove the redundant points in the image,which improves the operation efficiency.The improved algorithm is applied to the image stitching of UAV.The experiment shows that the improved algorithm can complete the image registration efficiently while ensuring the matching accuracy.Finally,the registration image is fused by the gradual in and gradual out method After fusion,there is no splicing gap,the overall brightness changes evenly and there is no obvious ghosting,and the splicing effect is better.