EUVIMEDEUVIMED entdecken
Crossref

Log-Reparameterized Diffusion Priors for Blind Radial Defocus Deblurring

Chang Wang, Guangxu Li, Yaqiang Hu, Tohru Kamiya

IEEE Transactions on Computational Imaging · 2026 · 12 · 1008-1017

DOI: 10.1109/tci.2026.3695652 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Computational Imaging, 2026. Verfasst von Chang Wang und 3 weiteren Autor:innen. Die Arbeit nennt 55 Literaturangaben. Gefördert von: China Scholarship Council.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Radial defocus blur—a pervasive degradation in optical systems—exhibits spatially-variant patterns where blur intensity increases monotonically from center to periphery due to spherical aberration and field curvature. While common in systems, existing general-purpose deblurring methods often overlook this physics-informed structure, limiting their real-world performance. Although diffusion models show promise in image restoration, their application to radial defocus remains challenged by a lack of underlying physical modeling and optimization instabilities. To address these challenges, we propose Log-Diff, a physics-informed blinddeblurring framework. We first construct a comprehensive dataset of defocus kernels derived from parametric physical modeling to accurately capture spatially-variant and distance-dependent blur characteristics. Crucially, we introduce a log-space optimization strategy ($\boldsymbol{\sigma }^{\prime } = \log \boldsymbol{\sigma }$) within the diffusion process. This mechanism effectively resolves instability by enforcing non-negativity constraints and preventing gradient vanishing through adaptive scaling. By integrating this reparameterized physical model into the reverse sampling process, our architecture enables the joint estimation of the sharp image and the blur field with strict physical consistency. Extensive experiments on FFHQ demonstrate that Log-Diff outperforms baselines by over 39% in perceptual metrics (LPIPS 0.12) while maintaining high fidelity (PSNR 27.41 dB). Unlike normalized approximations, our approach generalizes to real-world projector scenarios, establishing a new paradigm for integrating physical constraints into generative priors.

Förderung

  • China Scholarship Council

Zitieren

APA

Wang, C., Li, G., Hu, Y., & Kamiya, T. (2026). Log-Reparameterized Diffusion Priors for Blind Radial Defocus Deblurring. IEEE Transactions on Computational Imaging, 12, 1008-1017. https://doi.org/10.1109/tci.2026.3695652

Vancouver

Wang C, Li G, Hu Y, Kamiya T. Log-Reparameterized Diffusion Priors for Blind Radial Defocus Deblurring. IEEE Transactions on Computational Imaging. 2026;12:1008-1017. doi:10.1109/tci.2026.3695652

BibTeX
@article{Wang2026,
  author = {Wang, Chang and Li, Guangxu and Hu, Yaqiang and Kamiya, Tohru},
  title = {Log-Reparameterized Diffusion Priors for Blind Radial Defocus Deblurring},
  journal = {IEEE Transactions on Computational Imaging},
  year = {2026},
  volume = {12},
  pages = {1008--1017},
  doi = {10.1109/tci.2026.3695652}
}
Kostenlos starten →