Efficient least squares multimodal registration with a globally exhaustive alignment search

Area of Science:

• Computer Vision
• Medical Imaging
• Remote Sensing

Background:

• Image registration is challenging with large initial misalignments, especially for multimodal images.
• Local optimization methods often fail to find the global optimum in such scenarios.
• Region of Interest (ROI) analysis further complicates registration.

Purpose of the Study:

• To develop an efficient multimodal rigid-body image registration method.
• To address challenges posed by large initial misalignments within ROIs.
• To provide a robust solution for multimodal image registration problems.

Main Methods:

• An exhaustive search over all integer translations combined with local search for rotations.
• Utilizes the Fast Fourier Transform (FFT) for efficient computation of the sum of squared differences (SSD) cost function.
• Employs linear regression to model intensity relationships between images for each translation.

Main Results:

• The method successfully registers images with significant initial misalignments.
• Demonstrates consistent performance across medical imaging, remote sensing, and forensic science applications.
• Achieves close registration, enabling subsequent fine-tuning by local optimization methods.

Conclusions:

• The developed method offers an efficient and robust solution for multimodal rigid-body image registration with large initial misalignments.
• It effectively overcomes limitations of traditional local optimization techniques.
• The approach provides a reliable pre-processing step for various image analysis tasks.