Abstract
Plug-and-Play (PnP) algorithms are a class of iterative algorithms that address image inverse problems by combining a physical model and a deep neural network for regularization. Even if they produce impressive image restoration results, these algorithms rely on a non-standard use of a denoiser on images that are less and less noisy along the iterations, which contrasts with recent algorithms based on Diffusion Models (DM), where the denoiser is applied only on re-noised images. We propose a new PnP framework, called Stochastic deNOising REgularization (SNORE), which applies the denoiser only on images with noise of the adequate level. It is based on an explicit stochastic regularization, which leads to a stochastic gradient descent algorithm to solve ill-posed inverse problems. A convergence analysis of this algorithm and its annealing extension is provided. Experimentally, we prove that SNORE is competitive with respect to state-of-the-art methods on deblurring and inpainting tasks, both quantitatively and qualitatively.
Introduction
Image restoration tasks, such as deblurring or inpainting, aim to recover a clean image from a degraded observation. Modern Plug-and-Play (PnP) methods leverage pre-trained deep neural networks—specifically denoisers—as priors to guide this reconstruction. However, PnP methods often suffer from domain shift. As iterations progress, the image estimate becomes cleaner, eventually forcing the denoiser to process images far outside its training distribution (which typically consists of noisy images). This can lead to unwanted artifacts.
Conversely, Diffusion Models (DMs) maintain the denoiser within its trained domain by always applying it to re-noised versions of the image. While DMs yield high perceptual quality, they often lack the rigorous convergence guarantees of optimization-based PnP. SNORE (Stochastic deNOising REgularization) bridges this gap by introducing a theoretically grounded framework that combines the reliability of PnP with the generative power of diffusion-style denoising.
Methodology
The core of SNORE is a new regularization term, $\mathcal{R}_\sigma(x)$, representing the average value of a smoothed log-prior. Instead of direct application, the algorithm follows a stochastic process:
- Noise Injection: A noisy version $\tilde{x}_k = x_k + \sigma\epsilon$ is created.
- Denoising: The denoiser processes this noisy version, ensuring it operates "at home" within its trained domain.
- Stochastic Update: The difference between the noisy image and the denoised output informs a Stochastic Gradient Descent (SGD) step.
The authors provide a convergence analysis proving the algorithm reaches a critical point of the objective function. An Annealed SNORE variant is also introduced, which gradually reduces the noise level $\sigma$ to transition from coarse global structures to fine local details, mirroring the reverse process in Diffusion Models.
Experimental Setup
SNORE was evaluated on standard benchmarks including CBSD68 and CBSD10. For deblurring, the Levin et al. Kernels (8 real-world camera shake kernels) were used to simulate realistic degradations.
The framework was also tested on SAR-speckle free data for despeckling. The underlying denoiser, DRUNet, was trained on a composite dataset including CBSD, Waterloo Exploration, DIV2K, and Flick2K.
Results and Discussion
SNORE consistently outperforms standard PnP and Regularization by Denoising (RED) methods in perceptual metrics such as LPIPS and BRISQUE, even when traditional PSNR values are comparable. In inpainting tasks with 50% missing pixels, SNORE recovers sharp structures and realistic textures where other methods often produce blur.
In deblurring experiments, SNORE achieved state-of-the-art results competitive with specialized solvers like DiffPIR. While SNORE offers superior quality and theoretical stability, it is computationally more intensive. For example, deblurring an image took approximately 67 seconds with SNORE compared to 5 seconds for RED, due to the many iterations required by the annealing schedule. Future work will focus on accelerating convergence using more efficient stochastic optimizers like ADAM.
Conclusion
SNORE successfully integrates stochastic denoising into a formal optimization framework. By keeping the denoiser within its trained domain through intentional noise injection, it provides a provably convergent path to high-quality image restoration, effectively bridging the gap between PnP methods and Diffusion Models.
Summary updated