Spatiotemporal fringe pattern analysis using physics-informed deep learning
Photonics Research · 01.05.2026 · 14 (5) · 1893
DOI: 10.1364/prj.584978 ↗Kurz gefasst
Fachartikel in der Zeitschrift Photonics Research, online erschienen am 24.04.2026. Verfasst von Xinran Zhou und 7 weiteren Autor:innen. Die Arbeit wurde bisher 5-mal zitiert (Crossref) und nennt 45 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Fundamental Research Funds for the Central Universities, National Key Laboratory of Shock Wave and Detonation Physics.
Abstract (Originalsprache)
In optical metrology, particularly in phase-measuring profilometry techniques, achieving both high measurement accuracy and high temporal resolution remains a long-standing challenge. For high-speed dynamic measurements, many existing methods still adopt a frame-by-frame processing strategy, tending to overlook the rich temporal information embedded across adjacent frames. This constraint makes it difficult to maintain robust performance under ultra-short exposure conditions. Taking fringe projection profilometry (FPP) as a representative example, we propose a physics-informed spatiotemporal fringe pattern analysis (PIST-FPA) framework that unifies physical modeling with deep learning for high-speed, high-fidelity three-dimensional (3D) reconstruction. Rather than processing each frame independently, PIST-FPA jointly analyzes a short sequence of fringe images within a temporal window, exploiting both spatiotemporal correlations and physics-based priors. PIST-FPA incorporates a background-estimation network for illumination correction, a physics-guided optical-flow alignment for motion compensation, and an attention-enhanced U-Net for spatiotemporal phase demodulation, forming a unified and interpretable learning-based processing pipeline. This hybrid paradigm effectively decouples object motion from surface morphology, enabling accurate and temporally consistent phase retrieval even under ultra-short exposure and low-signal-to-noise conditions. Both synthetic and real experiments at 10,000 frames per second demonstrate that PIST-FPA substantially outperforms both the conventional Fourier transform method and single-frame deep-learning method in terms of accuracy, robustness, and spatiotemporal consistency. By bridging physical modeling and data-driven learning in a unified spatiotemporal framework, PIST-FPA establishes a generalizable paradigm for physics-informed dynamic optical metrology, paving the way for next-generation high-speed phase-measuring profilometry and intelligent optical instrumentation.
Förderung
- National Natural Science Foundation of China
- Fundamental Research Funds for the Central Universities
- National Key Laboratory of Shock Wave and Detonation Physics
- China Postdoctoral Science Fund
- Open Research Fund of Jiangsu Key Laboratory of Spectral Imaging & Intelligent Sense
- Postgraduate Research & Practice Innovation Program of Jiangsu Province
Zitieren
APA
Zhou, X., Liu, Y., Tang, J., Xiao, Y., Liu, Y., Feng, S., Chen, Q., & Zuo, C. (2026). Spatiotemporal fringe pattern analysis using physics-informed deep learning. Photonics Research, 14(5), 1893. https://doi.org/10.1364/prj.584978
Vancouver
Zhou X, Liu Y, Tang J, Xiao Y, Liu Y, Feng S, et al. Spatiotemporal fringe pattern analysis using physics-informed deep learning. Photonics Research. 2026;14(5):1893. doi:10.1364/prj.584978
BibTeX
@article{Zhou2026,
author = {Zhou, Xinran and Liu, Yiheng and Tang, Jie and Xiao, Yutong and Liu, Yifan and Feng, Shijie and Chen, Qian and Zuo, Chao},
title = {Spatiotemporal fringe pattern analysis using physics-informed deep learning},
journal = {Photonics Research},
year = {2026},
volume = {14},
number = {5},
pages = {1893},
doi = {10.1364/prj.584978}
}
EUVIMED