UDD: Unsupervised denoising diffusion for noisy multi-focus image fusion.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Multi-focus image fusion (MFIF) aims to create all-in-focus images from varied focal planes.
• Existing MFIF methods degrade significantly in noisy conditions, and supervised denoising requires paired clean-noisy data, posing acquisition challenges.

Purpose of the Study:

• To develop a noise-robust multi-focus image fusion framework that does not require noise-specific training.
• To address the limitations of current MFIF and denoising techniques in handling noisy inputs.

Main Methods:

• Proposes an unsupervised denoising diffusion (UDD) framework incorporating a noise-aware conditional diffusion (NACD) module for pixel-space denoising.
• Introduces a self-collaboration enhancement fusion (SCEF) module to leverage complementary information during diffusion for unified denoising and fusion.
• Employs a unified unsupervised approach for robust performance across various noise levels.

Main Results:

• The UDD framework demonstrates superior performance compared to state-of-the-art denoise-then-fuse pipelines.
• Achieved a 2.17 dB PSNR and 0.58 MI improvement on the Gaussian Lytro-N dataset (σ=0.2).
• Reduced PIQE from 45.72 to 30.29 on the real-world RLLMF dataset, indicating enhanced fusion quality.

Conclusions:

• The proposed UDD framework offers effective noise-robust multi-focus image fusion without reliance on paired training data.
• UDD provides a generalizable and robust solution for MFIF in diverse and challenging noise conditions.
• The method significantly improves image quality metrics and reduces perceived image quality differences.