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

Multi-focus image fusion (MFIF) generates all-in-focus images from multiple source images captured at different focal planes. However, existing MFIF methods suffer severe performance degradation under noise conditions due to their reliance on high-quality input assumptions, while supervised denoising approaches depend heavily on noisy training data, limiting generalizability and creating challenges in acquiring paired clean-noisy images. We propose an unsupervised denoising diffusion (UDD) framework for noise-robust fusion without relying on noise-specific training. UDD introduces three key components: (1) an unsupervised pixel-space noise-aware conditional diffusion (NACD) that introduces a pixel-space conditional formulation, eliminating the need for explicit degradation matrices and their transformed-domain decomposition and achieving noise-agnostic denoising through conditional Gaussian scaling; (2) a self-collaboration enhancement fusion (SCEF) module that exploits complementary information along the diffusion trajectory generated by NACD through residual learning and spatially adaptive enhancement for unified optimization of denoising and fusion; and (3) a unified unsupervised framework achieving robust performance across diverse noise conditions. For example, UDD consistently outperforms state-of-the-art denoise-then-fuse pipelines, with PSNR improving by 2.17 dB and MI by 0.58 on the Gaussian Lytro-N (σ=0.2) dataset, and reducing PIQE from 45.72 to 30.29 on the real-world RLLMF dataset.