Non-rigid image registration is often performed using cross-correlation; however, this approach struggles with outliers, particularly in regions lacking distinctive patterns. Keypoint matching methods provide reliable point correspondences, yet they may not sufficiently capture a dense deformation field.
In this talk, I will introduce a novel approach to modeling deformation fields as a geometric transformation field, where transformations vary smoothly across space. Additionally, I will present a sparse kernel regression technique that utilizes a blending of geometric transforms to estimate the deformation field from a limited number of point correspondences. To demonstrate the effectiveness of this approach, I will showcase several examples of non-rigid deformation modeling in histological section images and CT lung images.