SADUNs: Sharpness-Aware Deep Unfolding Networks for Image Restoration

ICLR 2026 Conference Submission17462 Authors

Keywords: linear inverse problems, deep unfolding networks, sharpness aware minimization

Abstract: The ability to improve model performance while preserving structural integrity represents a fundamental challenge in deep unfolding networks (DUNs), particularly when handling increasingly complex black-box priors. This paper presents a novel Sharpness-Aware Deep Unfolding Networks (SADUNs), that addresses these limitations by integrating Sharpness-Aware Minimization (SAM) principles with the proximal operator theory. By analyzing the gradient landscape of linear inverse problems, we develop the separable sharpness-aware perturbation and subgradient calculation modules that maintain original network structures while enhancing optimization. Our theoretical analysis demonstrates that SADUNs achieve linear convergence for sparse coding tasks under common assumptions. Crucially, our framework reduces training costs through fine-tuning compatibility and preserves inference speed by eliminating redundant gradient computations via proximal operator properties. Comprehensive experiments validate SADUNs across multiple domains. Moreover, we have validated the improvement of our framework on plug and play single image super resolution tasks, which means that our framework has the potential to expand to more types of deep unfolding networks.

Primary Area: interpretability and explainable AI

Submission Number: 17462