An Approach to Image Matching Based on SIFT and ORB Algorithms

A four-stage approach to the matching of medical images based on the use of SIFT and ORB algorithms has been considered. At the first stage, keypoints and their corresponding descriptors are identified using the SIFT or ORB algorithm for a certain pair of images. At the second stage, the identified descriptors corresponding to keypoints have been clustered in order to determine the optimal number of clusters and, if necessary, identify target clusters among them. In this case, clustering can be performed using some algorithm which allows to determine the centroids or medoids of the clusters. It is proposed to exclude from further consideration those keypoints whose descriptors are significantly removed from the centroids (medoids) of their clusters. At the third and fourth stages, the sequential identification of “good” matches based on the Lowe’s ratio test and the homography method for the remaining keypoints, for example, of target clusters with the rejection of matches which can be classified as noise has been realized. A practical example which realizes the matching of images from the field of medical diagnostics and confirms the effectiveness of the proposed approach has been given.