Feature based image registration using non-degenerate pixels

Abstract

Image registration (IR) aims to geometrically match one image to another. It is extensively used in many imaging applications. Among many existing IR methods, one widely used group of methods are feature-based. By a feature-based method, a number of relevant image features are first extracted from the two images, respectively, and then a geometric matching transformation is found to best match the two sets of features. However, proper identification and extraction of image features turns out to be a challenging task. Generally speaking, a good image feature extraction method should have the following two properties: (i) the identified image features should provide us proper information to approximate the geometric matching transformation accurately, and (ii) they should be easy to identify by a computer algorithm so that the entire feature extraction procedure is computer automatic. In this paper, a new type of image features is studied, which has the two properties described above. Together with the widely used thin plate spline (TPS) geometric transformation model, it is shown that our feature-based IR method works effectively in various cases.