Some Experiments in Relaxation Image Matching Using Corner Features

Abstract

For the improvement of navigation accuracy, image matching is considered.

A relaxation method based on patterns of local features is used to find matches between pairs of images or subimages that differ in position or orientation. A local operator is applied to the two images to detect two sets of “corners” C₁,..., C_m and D₁,..., D_n, each of them characterized by position, orientation, contrast, and “sharpness” (of the angle). For each pair (C_i, D_j), a figure of merit is computed, and a relaxation process is used to iteratively adjust these figures of merit, based on the merits of other pairs in approximately corresponding positions. After a few iterations of this process, “good” matches (pairs having much better merit than their next best choices) are clustered, yielding sets of transformation parameters (shift vectors or rotation angles) under which many corners correspond. This method has yielded good results for TV images of objects such as tools and industrial parts, as well as for aerial images of terrain.