Image registration of simulated synthetic aperture sonar images using SIFT

Abstract

The scale-invariant feature transform (SIFT) algorithm is investigated for registration of two synthetic aperture sonar images. The image registration geometry for two ideal sonar tracks is presented. The sonar track parameters (and thus the registration transform) can be fully determined from two exact non-degenerate correspondences and the altitude of one of the tracks. SIFT yielded 99% inlier correspondences and demonstrated sub-pixel accuracy. RANSAC was used to select inliers within a squared pixel error tolerance. The proposed feature-based algorithm yielded estimations of the tracks corresponding to a misregistration up to 0.17 pixels throughout the scene.

A SIFT‐Like Feature Detector and Descriptor for Multibeam Sonar Imaging

“…Researchers have proposed various methods to improve SIFT's performance in such cases. Depending on the application, some researchers removed or modified some steps of the algorithm to improve its performance [19,20]. Others suggested denoising or filtering the images during preprocessing to reduce the influence of speckle noise [21,22].…”

“…Considering the gradient definition of existing algorithms [9][10][11][19][20][21][22], in this study, a new gradient definition adapted to MBS images is presented. This new gradient definition makes magnitude and orientation robust to speckle noise.…”