Overview
Field: Computer Vision Authors: Haley Duba-Sullivan, Emma J. Reid, Charles A. Bouman, Gregery T. Buzzard Published: 2026-11-06 arXiv: 2611.09845
Summary
Plug-and-play (PnP) methods reconstruct images by alternating a data-fitting step with an image denoiser. Their performance depends on two parameters: the step size \(\gamma\) and the measurement noise level \(\sigma\). In practice, \(\gamma\) is often tuned by grid search over validation images and must be retuned whenever the problem setting changes, and \(\sigma\) is assumed known or estimated separately.
AdaptivePnP is a PnP framework that jointly estimates \(\sigma\) and \(\gamma\) at every iteration from the data-fitting residual. To do so, the authors apply the PnP virtual model to the data-fitting proximal map and show that the residual is zero-mean Gaussian with covariance parameterized by \(\gamma\) and \(\sigma\).
However, when one parameter is much larger than the other, the residual carries little information about the smaller one, and using a plain maximum likelihood (ML) estimator can degrade PnP iterations. The authors therefore introduce an identifiable maximum likelihood (IML) estimator, which restricts the ML estimate to a window where the normalized Cramér-Rao bound of each parameter is bounded by a user-specified tolerance. For the virtual model, this window is an interval of the ratio \(\rho = \gamma/\sigma\), so computing the IML estimate reduces to a one-dimensional search over \(\rho\).
Results
- AdaptivePnP is applied to two standard PnP methods, producing algorithms that require neither tuning of \(\gamma\) nor knowledge of \(\sigma\).
- On image deblurring and super-resolution tasks, AdaptivePnP matches or exceeds PnP baselines tuned with ground-truth \(\sigma\).
Original Abstract (excerpt)
> Plug-and-play (PnP) methods reconstruct images by alternating a data-fitting step with an image denoiser. Their performance depends on two parameters: the step size \(\gamma\) and the measurement noise level \(\sigma\). In practice, \(\gamma\) is often tuned by grid search over validation images and must be retuned whenever the problem setting changes, and \(\sigma\) is assumed known or estimated separately. We propose AdaptivePnP, a PnP framework that jointly estimates \(\sigma\) and \(\gamma\) at every iteration from the data-fitting residual. To do so, we apply the PnP virtual model to the data-fitting proximal map and show that the residual is zero-mean Gaussian with covariance parameterized by \(\gamma\) and \(\sigma\). However, when one parameter is much larger than the other, the residual carries little information about the smaller one ...
---
*Collected automatically on 2026-11-10. Source: arXiv:2611.09845*