SinDiff-denoise: Single image denoising based on contrastive diffusion model

Long Chen

doi:10.1016/j.neucom.2026.133847

SinDiff-denoise: Single image denoising based on contrastive diffusion model

2026-05-06

0

PRE

AI

DOI：10.1016/j.neucom.2026.133847

Original

Original request for help

Save

Abstract

En 中文

• We propose SinDiff-Denoise, a self-supervised framework that unlocks the generative potential of diffusion models for single-image denoising. It eliminates the reliance on clean reference data by learning robust restoration priors directly from the noisy observation. • A self-corruption and self-denoising strategy is introduced to construct a deterministic diffusion chain. By training the model to predict incrementally injected Gaussian noise, we enable implicit suppression of the unknown complex noise distribution. • We integrate contrastive learning with trajectory consistency regularization. This dual-constraint design aligns semantic features across augmented views to prevent overfitting and enforces structural stability along the reverse diffusion path. • An optimal step-matching strategy is developed to calibrate the reverse process, allowing for high-fidelity image restoration in a single inference step, effectively resolving the trade-off between diffusion generation quality and computational efficiency.

Keywords:

SinDiff-Denoise

single-image denoising

diffusion models

self-supervised learning

contrastive learning

SinDiff-denoise: Single image denoising based on contrastive diffusion model

Abstract

Journal

Organization