Test-Time Adaptation Improves Inverse Problem Solving With Patch-Based Diffusion Models

Jason Hu

doi:10.1109/TCI.2025.3587407

2025-01-01

PRE

DOI：10.1109/TCI.2025.3587407

原文链接

原文求助

摘要

En 中文

扩散模型因具备学习强大图像先验的能力，在解决反问题方面取得了显著成功，但现有方法需要大量与测试数据集同分布的训练图像数据。在实际应用中，可用训练数据集的大小可能从不存在到非常大不等。在某些情况下，由于先验学习不足，从有限数据训练的传统扩散模型可能导致重建结果较差。一种潜在的改进方法是，先使用可能分布不匹配的可用训练数据训练扩散模型，然后在重建时通过网络微调来弥补分布差异。本研究探讨了该网络微调过程对训练数据分布差异程度各异的扩散模型的影响。具体而言，我们采用自监督损失来使学习到的扩散网络适应测试数据，同时帮助网络输出与测量值保持一致性。理论和实验均表明，基于图像块的扩散先验的测试时适应，相较于传统全图像扩散模型的测试时微调，能带来更高质量的重建结果。大量实验显示，在各类反问题中，当训练与测试数据分布存在显著领域偏移时，测试时适应能显著提升图像重建质量。有趣的是，即使在同分布情况下，测试时适应也能显著改善重建质量。

Keyword:

Deep learning

diffusion models

image processing

inverse problems

期刊

IEEE Transactions on Computational Social Systems

IF:

4.9

论文数:

635

被引数:

6.8K

机构

University of Michigan

学者数:

6.4W

论文数: 5.3W

被引数: 124

引用论文

Dynamic CT Reconstruction from Limited Views with Implicit Neural Representations and Parametric Motion Fields

2021 IEEE/CVF International Conference on Computer Vision (ICCV)

IF0

2021-10-01

OAAI

Albert W. Reed; Hyojin Kim; Rushil Anirudh; K. Aditya Mohan; Kyle Champley; Jingu Kang; Suren Jayasuriya

Image denoising with block-matching and 3D filtering

SPIE Proceedings

IF0

2006-02-02

PREAI

Kostadin Dabov; Alessandro Foi; Vladimir Katkovnik; Karen Egiazarian

Provable Preconditioned Plug-and-Play Approach for Compressed Sensing MRI Reconstruction 可证明预条件化即插即用方法用于压缩感知MRI重建

IEEE Transactions on Computational Imaging

IF0

2024-01-01

PREAI

Hong,Tao; Xu,Xiaojian; Hu,Jason; Fessler,Jeffrey A.

A Style-Based Generator Architecture for Generative Adversarial Networks 生成对抗网络的基于样式的生成器体系结构

IEEE Transactions on Pattern Analysis and Machine Intelligence

IF0

2021-12-01

PREAI

Tero Karras; Samuli Laine; Timo Aila

Restoring Vision in Adverse Weather Conditions With Patch-Based Denoising Diffusion Models 使用基于补丁的去噪扩散模型在恶劣天气条件下恢复视力

IEEE Transactions on Pattern Analysis and Machine Intelligence

IF0

2023-08-01

OAAI

Ozan Özdenizci; Robert Legenstein

The Lung Image Database Consortium, (LIDC) and Image Database Resource Initiative (IDRI): A Completed Reference Database of Lung Nodules on CT Scans 肺图像数据库联盟 (LIDC) 和图像数据库资源倡议 (IDRI): ct扫描上完整的肺结节参考数据库

MEDICAL PHYSICS

IF3.2

2011-01-24

1.8K

OAAI

Armato, Samuel G., III; McLennan, Geoffrey; Bidaut, Luc; McNitt-Gray, Michael F.; Meyer, Charles R.; Reeves, Anthony P.; Zhao, Binsheng; Aberle, Denise R.; Henschke, Claudia I.; Hoffman, Eric A.; Kazerooni, Ella A.; MacMahon, Heber; van Beek, Edwin J. R.; Yankelevitz, David; Biancardi, Alberto M.; Bland, Peyton H.; Brown, Matthew S.; Engelmann, Roger M.; Laderach, Gary E.; Max, Daniel; Pais, Richard C.; Qing, David P-Y; Roberts, Rachael Y.; Smith, Amanda R.; Starkey, Adam; Batra, Poonam; Caligiuri, Philip; Farooqi, Ali; Gladish, Gregory W.; Jude, C. Matilda; Munden, Reginald F.; Petkovska, Iva; Quint, Leslie E.; Schwartz, Lawrence H.; Sundaram, Baskaran; Dodd, Lori E.; Fenimore, Charles; Gur, David; Petrick, Nicholas; Freymann, John; Kirby, Justin; Hughes, Brian; Casteele, Alessi Vande; Gupte, Sangeeta; Sallam, Maha; Heath, Michael D.; Kuhn, Michael H.; Dharaiya, Ekta; Burns, Richard; Fryd, David S.; Salganicoff, Marcos; Anand, Vikram; Shreter, Uri; Vastagh, Stephen; Croft, Barbara Y.; Clarke, Laurence P.

An Educated Warm Start for Deep Image Prior-Based Micro CT Reconstruction

IEEE Transactions on Computational Imaging

IF0

2022-01-01

OAAI

Riccardo Barbano; Johannes Leuschner; Maximilian Schmidt; Alexander Denker; Andreas Hauptmann; Peter Maass; Bangti Jin

Boosting the Performance of Plug-and-Play Priors via Denoiser Scaling

2020 54th Asilomar Conference on Signals, Systems, and Computers

IF0

2020-11-01

OAAI

Xiaojian Xu; Jiaming Liu; Yu Sun; Brendt Wohlberg; Ulugbek S. Kamilov