Plug-and-Play image restoration with Stochastic deNOising REgularization

2026-06-12 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, extended

Summary

SNORE (Stochastic deNOising REgularization) is a novel Plug-and-Play (PnP) framework designed for image inverse problems such as deblurring and inpainting. It improves upon traditional PnP algorithms by applying the denoiser exclusively to images with an appropriate noise level, aligning with practices in Diffusion Models. This approach, based on explicit stochastic regularization, results in a stochastic gradient descent algorithm. The framework includes an annealing extension and offers a comprehensive convergence analysis. Experimentally, SNORE demonstrates competitive quantitative and qualitative performance against leading methods on deblurring and inpainting tasks. The implementation utilizes a Gradient-Step denoiser, running for 1500 iterations across 16 annealing levels, with specific initial parameters for σ and α and a fixed step-size of 0.1. The associated code is publicly available on GitHub.

Key takeaway

For Machine Learning Engineers developing image restoration solutions, SNORE provides a robust framework for deblurring and inpainting. If you are seeking methods with strong theoretical convergence guarantees and competitive empirical performance, consider integrating SNORE's stochastic regularization. Its approach of applying denoisers to appropriately noisy images, coupled with an annealing schedule, can yield superior perceptual and quantitative results compared to existing PnP or DDM techniques, especially for ill-posed inverse problems.

Key insights

SNORE improves PnP image restoration by applying denoisers to appropriately noisy images, enhancing performance and theoretical guarantees.

Principles

• Denoisers perform better on images with noise levels matching their training domain.
• Explicit stochastic regularization can lead to robust stochastic gradient descent.
• Annealing procedures can improve convergence in non-convex optimization.

Method

SNORE optimizes a variational objective using a stochastic gradient descent algorithm, applying a deep denoiser to a noisy version of the image at each iteration, optionally with an annealing schedule for σ and α.

In practice

• Implement SNORE with a Gradient-Step denoiser for deblurring and inpainting.
• Use an annealing schedule for σ and α to enhance practical efficiency.
• Initialize with observation and fixed step-size (e.g., δ=0.1).

Topics

• Image Restoration
• Plug-and-Play Algorithms
• Stochastic Regularization
• Deblurring
• Inpainting
• Diffusion Models

Code references

• Marien-RENAUD/SNORE

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• An Analysis of Regularization and Fokker-Planck Residuals in Diffusion Models for Image Generation — Simpler regularizers can achieve Fokker-Planck regularization benefits in diffusion models with less computational cost.
• Beyond MMSE: Enhancing PnP Restoration with ProxiMAP — ProxiMAP enhances PnP image restoration by aligning denoiser noise schedules to prevent "cartoon-like" artifacts from inexact learned scores.
• SNNF: An SNN-based Near-Sensor Noise Filter for Dynamic Vision Sensors — SNNF efficiently filters DVS noise using a compact binary representation and a hardware-optimized Spiking Neural Network.
• IG-Diff: Complex Night Scene Restoration with Illumination-Guided Diffusion Model — Illumination-guided diffusion models effectively restore complex night scenes by integrating illumination priors and synthetic data.
• RelativeFlow: Taming Medical Image Denoising Learning with Noisy Reference — RelativeFlow improves medical image denoising by learning relative noise mappings from heterogeneous noisy references.
• Adapting Diffusion Language Models for Lossless Pixel-Level Image Transmission — DDM-SSCC adapts diffusion language models for lossless pixel-level image transmission, achieving superior exact recovery via novel coding and denoising strategies.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.