New domain adaptation method boosts PnP image reconstruction by 40%
Proximal mismatch quantified: new adaptation beats MSE in few-shot regimes.
Deep Dive
The paper introduces a domain adaptation technique for plug-and-play proximal gradient descent (PnP-PGD) image reconstruction. It defines 'proximal mismatch' between a deployed denoiser and the target-domain reference. The authors derive a stationarity bound that decays at O(1/K) with an additive mismatch term. Experiments on Gaussian deblurring and super-resolution show proximal matching adaptation yields significantly better reconstruction quality than MSE-based adaptation, especially in few-shot settings.
Key Points
- Defines 'proximal mismatch' to quantify domain shift in PnP-PGD denoisers
- Derives convergence bound of O(1/K) with additive mismatch term
- Proximal matching adaptation outperforms MSE-based adaptation by up to 40% in few-shot super-resolution experiments
Why It Matters
Enables robust medical and satellite image reconstruction when training data differs from deployment conditions.