Motion compensation from limited data for reference-constrained image reconstruction

Abstract

When reconstructing images from limited (or sparsely sampled) data, reference (or template) images are useful for constraining image reconstruction for various applications. However, in order to be an effective constraint, the reference should be correctly aligned with the target image one wants to reconstruct. Conventional image registration methods assume that both the reference and target images are completely specified, but one usually has only limited data from the target. Therefore, these methods do not apply. This thesis addresses this new problem of registering a known high-resolution reference image to an unknown target image for which one has only limited measurements.

We solve this problem by introducing an intermediate image model that expresses the target image as a combination of a generalized series model and a residual component. This model allows the reference and target images to have different contrast and can be used with various motion models. It also makes use of all the available data to estimate the motion parameters. We propose practical algorithms to solve the optimization problems associated with motion parameter estimation. We also analyze the characteristics and performance of the proposed method by an estimation-theoretic approach and by computer simulations. We demonstrate accurate motion parameter estimation for an affine transformation model and a nonrigid deformable model.