Image Matching Based on ORB with Nonlinear Scale Space

Abstract

To solve the problems of image precision loss and edge blur caused by SIRB (traditional improved ORB) algorithm when build multi-scale space using linear Gaussian pyramid, NORB (nonlinear scale space ORB) algorithm, which is based on the FED (Fast Explicit Diffusion) framework of nonlinear scale space structure, is proposed. Also, in order to overcome the limitation that after normalizing multi-scale space, there is a large number of repeatable (means multiple keypoints in the same image point) and unstable keypoints, the two-steps removing method is to put forward to filter the keypoints. Experimental results show that NORB exhibits scale invariance which is not possessed by the original ORB, effectively solves SIRB’s problem of accuracy loss, and manages to remove the unstable keypoints and repeatable keypoints, greatly improving the matching accuracy. Compared with the ORB and SIRB, although NORB slightly increases the time consumption, it obtains a higher matching accuracy and retains the high speed of the ORB algorithm at the same time, therefore overall, the detection efficiency is superior to other algorithms.