Figure 1 Concept of PDPrior for eyeglass reflection removal.
Figure 2 Flowchart of PDPrior.
Figure 3 Experimental data collection setup and results and edge device prototype for online communication.
Announcing new publications from Opto-Electronic Advances. DOI 10.29026/oea.2026.250249.
Shannon, County Clare, Ireland, April 30, 2026 /EINPresswire.com/ — Announcing a new publication from Opto-Electronic Advances. DOI 10.29026/oea.2026.250249.
Rapid advances in network communication technology have achieved significant advances in high-definition real-time video transmission. These developments have fundamentally changed the way people interact online. Video conferencing, remote work, and online social platforms have experienced explosive growth, with globally recognized services such as Zoom, Microsoft Teams, Tencent Meeting, and Feishu gaining widespread adoption. At the same time, advances in deep learning algorithms are making facial recognition-based biometrics the dominant identity verification approach in critical sectors such as government and finance. However, in scenarios such as video conferencing, facial recognition, and group photo shoots, glasses wearers frequently encounter reflection artifacts. Specular reflections on the surface of glasses can obscure key features of the face, reduce visual recognition, confuse the user experience, and significantly increase the error rate of facial recognition systems.
Reflections in glasses typically consist of specular highlights, reflections of the surrounding scene, or a combination of both. The captured image can be thought of as a superposition of a transparent layer and a reflective layer. Therefore, the goal of reflection removal is to effectively separate these two components to recover clear and reliable facial information. Existing reflection removal approaches can be broadly divided into single-image methods and multi-image methods. Single-image methods lack additional cues, so they typically rely on image statistics or data-driven deep learning models. In contrast, multi-imaging methods exploit changes in illumination, perspective, or polarization to easily disentangle reflective and transmissive layers.
Current single-image reflection removal methods rely heavily on large pairs of training data and often have limited generalization ability under unknown lighting conditions, limiting their practical applicability. Because there are inherent differences in the polarization properties of reflected and transmitted light, polarization filtering can partially separate the reflected and transmitted components. In recent years, the rapid development of portable polarization imaging devices, such as focal plane split polarization sensors and emerging metasurface-based polarization imaging techniques, has enabled the acquisition of polarization information in real-world scenarios, further accelerating research in polarization-based reflection cancellation. Nevertheless, traditional polarization filtering techniques are still limited in practice and provide optimal performance only under ideal conditions, such as when the angle of incidence approaches the Brewster angle and the optical path contains only simple single-pass reflection/transmission processes.
To address the above challenges, this study proposes a combined polarization and generation mechanism to eliminate eyeglass reflections. The proposed method, called PDPrior, is a polarization-induced diffusion prior model that requires neither training data nor ground-truth images, enabling artifact-free eyeglass reflection removal.
The core idea of PDPrior is to take full advantage of the inherent priors embedded in the diffusion generation model while incorporating polarization cues as constraints and guidance into the generation process. Unlike traditional supervised learning frameworks, PDPrior relies only on the acquired polarimetry observations and does not require training data or additional annotations, significantly reducing the cost of data acquisition and model training. Specifically, in this study, we construct a self-supervised loss function based on a physical imaging model and freeze the U-Net model parameters. During the inverse generation process, the reflection and transmission variables are updated alternately, resulting in an output that is not only visually realistic but also physically interpretable. By combining generative modeling and image physics, PDPrior remains effective even under unknown lighting conditions. Experimental results show that PDPrior can reliably remove eyeglass reflections without introducing artifacts across a variety of scenarios, including indoor and outdoor environments, polarized and non-polarized illumination, different facial appearances, and different types of glasses. Moreover, PDPrior achieves higher scores on facial image quality evaluation benchmarks such as CR-FIQA and CLIB-FIQA, demonstrating the effectiveness of PDPrior in improving downstream face-related tasks.
Future research will focus on deploying PDPrior on resource-constrained platforms such as edge devices to enable efficient and real-time reflection removal. This includes model compression and one-step diffusion inference to speed up processing while maintaining visual quality. Additionally, extending PDP to night-time imaging scenarios by leveraging infrared polarization data represents a promising direction toward all-day reflection removal.
PDPrior has wide application potential in areas such as video conferencing systems, facial recognition and ID authentication, intelligent security, mobile photography, and augmented reality. In video conferencing, clear, reflection-free facial images greatly improve communication quality and user immersion. In high-security scenarios such as financial services and government applications that rely on facial recognition, reliable reflection cancellation can effectively reduce recognition errors and improve system security and reliability. As a result, this technology is expected to have a positive impact on daily work, remote collaboration, and the digital operation of society.
The source code is publicly available at https://github.com/THUHoloLab/PDPrior.
The test dataset is available at https://cloud.tsinghua.edu.cn/f/a49e0f59a8a54c4eb14d/?dl=1.
Keywords: eyeglass reflection removal, diffusion model, untrained learning, polarization-guided optimization
# # # # # #
Liangcai Cao received the BA/MA and PhD degrees from Harbin Institute of Technology and Tsinghua University in 1999/2001 and 2005, respectively. Afterwards, he became an assistant professor at Tsinghua University’s School of Precision Instruments. He is currently a tenured professor and director of the Institute of Optoelectronics at Tsinghua University. He was a visiting scholar at the University of California, Santa Cruz and MIT in 2009 and 2014, respectively. His research interests include holographic imaging and holographic displays. He is a Fellow of Optica and SPIE. Research group homepage: http://www.holoddd.com
# # # # # #
Opto-Electronic Advances (OEA) is a high-impact, open access, peer-reviewed SCI journal with an Impact Factor of 22.4 (Journal Citation Reports 2024). OEA is indexed in the SCI, EI, DOAJ, Scopus, CA, and ICI databases and has expanded its editorial board to 41 members from 17 countries.
# # # # # #
More information: http://www.oejournal.org/oea
Editorial board: http://www.oejournal.org/oea/editorialboard/list
All issues are available in the online archive at http://www.oejournal.org/oea/archive.
Submissions to OEA can be made using ScholarOne (https://mc03.manuscriptcentral.com/oea).
ISSN: 2096-4579
CN: 51-1781/TN
Contact: oea@ioe.ac.cn
Twitter: @OptoElectronAdv (https://twitter.com/OptoElectronAdv?lang=ja)
WeChat: OE_Journal
# # # # # #
Chen YT, Kao LC. Polarization-induced diffusion preprocessing for eyeglass reflection removal. Opto-Electron Adv 9, 250249 (2026). DOI: 10.29026/oea.2026.250249
Optoelectronics Advances Editorial Office
Advances in optoelectronics
+86 591 2286 1222
please email here
Legal disclaimer:
EIN Presswire provides this news content “as is” without warranty of any kind. Our company does not assume any responsibility or liability
The accuracy, content, images, videos, licenses, completeness, legality, or reliability of the information contained herein;
article. If you have any complaints or copyright issues related to this article, kindly contact the author above.
