Back to search

University of Minnesota

Advancing Untrained Deep Neural Networks

Abstract

dc:description.abstract

High-quality images are essential for many safety-critical and practical domains, ranging from commercial and medical to military applications. Nevertheless, real-world images are frequently degraded by nuisance factors, leading to noise, blur, and other distortions. Image restoration, the task of recovering clean images from corrupted observations, has long been a central research topic. Traditional restoration methods, which typically rely on hand-crafted priors, often produce unsatisfactory results with visible artifacts. More recently, deep neural network (DNN)-based approaches have significantly improved restoration quality but at the cost of requiring large (paired) training datasets and exhibiting poor generalization across diverse corruptions. The emergence of untrained deep neural networks (UDNNs) offers a compelling alternative. Unlike conventional or supervised deep learning methods, UDNNs rely solely on the corrupted image itself, obviating the need for external training data while delivering competitive restoration performance. Despite their successes, two major limitations---overfitting to noise and slow optimization---have hindered their widespread adoption. This thesis advances the development and practical deployment of UDNNs by addressing these limitations and extending their applicability. First, we propose a principled early-stopping strategy that adaptively halts optimization before overfitting occurs, thereby preserving restoration quality while avoiding superfluous computational overhead. Second, we introduce the Deep Random Projector (DRP), a method that substantially accelerates the optimization of UDNNs without compromising restoration quality. Beyond methodological improvements, we demonstrate the practical utility of UDNNs across both low-level and high-level vision tasks. For the low-level vision task, we tackle RAW image demosaicing and denoising, a cornerstone of the image processing pipeline in modern digital cameras, by proposing JDD-DoubleDIP, a joint optimization framework employing two UDNNs for superior restoration. For the high-level vision task, we address the challenge of robust image classification under common corruptions through DIP-RIC, which integrates UDNN-based denoising with student models and ensemble learning, achieving strong robustness across diverse perturbations. By addressing the fundamental challenges of UDNNs and demonstrating their versatility across a broad range of vision tasks, this thesis helps pave the way for their broader adoption in real-world applications, from digital imaging pipelines to robust perception systems. Consequently, these advancements have the potential to foster safer, more efficient, and more reliable technologies in critical domains. Furthermore, the methods proposed in this thesis enhance the robustness and trustworthiness of modern DNN-based models, thereby advancing both future research innovation and the practical deployment of artificial intelligence (AI) systems.

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Li, Taihui

Subjects

dc:subject × 6

Rights

Language dc:language.iso
en

Identifiers

dc:identifier.*
Dc Identifier Citation
Li, Taihui. (2026). Advancing Untrained Deep Neural Networks. Retrieved from the University Digital Conservancy, https://hdl.handle.net/11299/280356.

Chain of custody

source
Harvested from
University of Minnesota
Base URL
conservancy.umn.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Li, Taihui. Advancing Untrained Deep Neural Networks. 2026. https://hdl.handle.net/11299/280356