Image Registration Algorithm Based on Manifold Regularization with Thin-Plate Spline Model

In this paper, we propose a new method based on manifold regularization technology with a thin-plate spline for image registration, which is used to remove the outlier by approximating the transformation function. Under a Bayesian framework, we use a latent variable to indicate whether a correspondence is an inlier that should satisfy a mapping function; and then, we formulate the problem as optimizing a posterior problem related to the mapping function. The initial correspondences discard some mismatching points because of the similarity constraint, which may contain important information; however, the manifold regularization (MR) term utilizes all of the feature points and preserve this information as a constraint. In addition, we use the thin-plate spline (TPS) model, which is composed of a global affine transformation and local bending function, to construct the transformation function. Finally, we obtain the solution using the expectation-maximization algorithm. Extensive experiments show that our method outperforms with other comparable state-of-art methods.