A Hybrid Image Registration for Large Global and Non-Linear Local Deformed Images

This paper presents a process for improving image registration using a fusion of established image registration techniques, such as the Feature-Based Linear (FBL) and Demons methods, to overcome their limitations. The process leverages the FBL method to automatically extract control points between the fixed and moving images. The fixed image control points are used to create a Delaunay triangulation, which along with mapping functions in the x and y directions will generate initial displacement fields that warp the moving image to get an initial approximation of the fixed image. The Demons method is used to further improve the quality of the transformed moving image. The performance of the proposed method is evaluated using various similarity measures, including Mean Squared Error and Mutual Information, and is tested using both synthesized and natural images under different levels of linear, non-linear, and combined deformation types. The synthesized images are created using an algorithm that introduces free-form deformations controlled by affine geometric parameters, such as translation, rotation, and scaling, at a user-selected region of interest.