A Convergent Image Fusion Algorithm Using Scene-Adapted Gaussian-Mixture-Based Denoising.

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• The plug-and-play (PnP) framework offers flexibility in image processing by treating denoisers as black boxes within iterative algorithms.
• Analyzing the convergence of PnP methods can be challenging due to the lack of explicit objective functions in many advanced denoisers.

Purpose of the Study:

• To develop a novel image fusion technique inspired by the PnP framework.
• To address the convergence analysis challenges in PnP-based image fusion.
• To achieve state-of-the-art performance in image fusion tasks.

Main Methods:

• We propose integrating a scene-adapted denoiser into the alternating direction method of multipliers (ADMM) iterative algorithm.
• The scene-adapted denoiser acts as a proxy for the proximity operator, enabling tailored denoising for specific imaging scenarios.
• The proposed method is evaluated on hyperspectral fusion/sharpening and fusion of blurred-noisy image pairs.

Main Results:

• The proposed scene-adapted denoiser within ADMM yields state-of-the-art results in image fusion.
• The integration of a scene-adapted denoiser allows for rigorous convergence analysis of the fusion algorithm.
• Demonstrated effectiveness on diverse image fusion challenges, including hyperspectral data and blurred/noisy image pairs.

Conclusions:

• The proposed scene-adapted denoiser-ADMM framework provides a robust and theoretically sound approach to image fusion.
• This method enhances flexibility and performance compared to traditional PnP techniques.
• The approach offers a promising direction for advanced image fusion applications requiring guaranteed convergence.