Trustworthy Image Super-Resolution via Generative Pseudoinverse

2025-05-18

AI Summary

This work addresses trustworthy image super-resolution (SR), focusing on two key consistency requirements: physical consistency with the degradation model and statistical asymptotic consistency with the low-resolution (LR) observations. To this end, we propose a generative pseudo-inverse framework that formulates SR as a prior-driven constrained optimization problem, explicitly incorporating the degradation operator. We design a degradation-aware network architecture and introduce an asymptotic consistency regularizer, ensuring the reconstructed solution converges statistically to the true high-resolution (HR) image. To our knowledge, this is the first generative SR method to rigorously guarantee both measurement consistency and asymptotic unbiasedness. Consequently, it substantially suppresses reconstruction artifacts and improves uncertainty calibration. Extensive experiments demonstrate state-of-the-art performance across multiple benchmarks, with notable advances in observation fidelity, physical interpretability, and overall reliability.

Technology Category

• Computer Vision: Learning & Optimization for CV
• Machine Learning: Calibration & Uncertainty Quantification
• Constraint Satisfaction and Optimization: Constraint Satisfaction

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Responsible Web: Consent frameworks and practices on the web
• Security and Privacy: Large-scale security measurements

Abstract

We consider the problem of trustworthy image restoration, taking the form of a constrained optimization over the prior density. To this end, we develop generative models for the task of image super-resolution that respect the degradation process and that can be made asymptotically consistent with the low-resolution measurements, outperforming existing methods by a large margin in that respect.

Problem

• Constrained optimization for trustworthy image restoration
• Generative models respecting image degradation process
• Asymptotically consistent super-resolution with low-resolution inputs

Innovation

• Constrained optimization over prior density
• Generative models respecting degradation process
• Asymptotically consistent with low-resolution measurements