Bidirectional composition on lie groups for gradient-based image alignment.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Computational Geometry

Background:

• Parametric image alignment is crucial for medical imaging and computer vision.
• Existing gradient-based methods often combine image gradients prematurely, limiting performance.
• There is a need for advanced alignment techniques that leverage gradient information more effectively.

Purpose of the Study:

• To introduce a new formulation for parametric gradient-based image alignment using bidirectional composition on Lie groups (BCL).
• To propose two novel methods, BCL and projected BCL (PBCL), based on this formulation.
• To analyze the performance and theoretical underpinnings of these new methods compared to state-of-the-art approaches.

Main Methods:

• Development of the bidirectional composition on Lie groups (BCL) formulation.
• Implementation of the BCL method using a compositional framework for error minimization.
• Introduction of the projected BCL (PBCL) as an approximation of the BCL method.
• Comparative analysis of computational complexity, convergence rate, and frequency of convergence.

Main Results:

• The BCL method effectively utilizes gradients from both template and current images without prior combination.
• Numerical experiments demonstrate performance improvements, especially under asymmetric noise levels.
• The proposed methods show competitive or superior performance compared to existing gradient-based techniques.

Conclusions:

• The BCL formulation offers a novel and effective approach to parametric gradient-based image alignment.
• The PBCL method provides a computationally efficient approximation with strong performance.
• These methods advance the field of image registration, particularly for challenging datasets with varying noise characteristics.