A maximum likelihood approach to joint image registration and fusion

Area of Science:

• Computer Vision
• Signal Processing
• Remote Sensing

Background:

• Image registration and fusion are conventionally treated as separate estimation problems.
• Existing methods may not optimally tune registration parameters for fusion quality.

Purpose of the Study:

• To develop a unified maximum likelihood framework for joint image registration and fusion.
• To optimize registration accuracy and fusion performance simultaneously.

Main Methods:

• A maximum likelihood approach is proposed, formulating joint image registration and fusion as a single estimation problem.
• The Expectation-Maximization algorithm is employed for solving the joint optimization.
• The Cramer-Rao bound (CRB) is derived for performance analysis.

Main Results:

• Experimental evaluation using visual, IR thermal, and hyperspectral images demonstrates performance.
• The proposed method achieves mean square error for registration parameters close to the CRBs.
• Fusion performance, measured by edge preservation (Q(AB/F)), is improved compared to Laplacian pyramid fusion.

Conclusions:

• The joint optimization approach effectively enhances both image registration accuracy and fusion quality.
• The method offers a robust solution for multi-modal and multi-sensor image fusion applications.
• The proposed technique provides a significant advancement over conventional separate processing methods.