ResearchPod Summary
Plug-and-Play (PnP) image reconstruction methods use pre-trained denoisers as implicit priors to solve inverse problems. A major challenge is that these denoisers are often deployed on data distributions different from their training domain, leading to performance degradation. This paper investigates how this domain shift affects the convergence of PnP-PGD by defining a new metric called proximal mismatch—the discrepancy between the deployed denoiser and the ideal proximal map of the target domain's regularizer.
The authors derive a stationarity bound for PnP-PGD that explicitly links the convergence rate to this proximal mismatch. Based on this theoretical insight, they propose a domain adaptation strategy called proximal matching. Unlike standard MSE-based fine-tuning, which targets the conditional mean of the data, proximal matching aligns the denoiser's output with the target-domain proximal map, which is more directly relevant to the PnP-PGD optimization objective.
The study evaluates proximal matching adaptation using two structured denoiser families: Learned Proximal Networks (LPNs) and Gradient-Step (GS) denoisers. The experiments, conducted on Gaussian deblurring and super-resolution tasks under significant domain shift, reveal that proximal matching consistently yields higher reconstruction quality than MSE-based adaptation. The performance gains are particularly pronounced in the few-shot regime, where the method achieves improvements exceeding 10 dB in PSNR compared to traditional approaches.
This work bridges the gap between the theoretical convergence analysis of PnP methods and the practical reality of domain shift. By showing that MSE-based training is not necessarily optimal for PnP reconstruction, the authors provide a principled, mathematically grounded framework for adapting denoisers. This is especially valuable for real-world imaging applications where high-quality, domain-specific training data is scarce and must be used efficiently to ensure reliable reconstruction.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.