Scattered OAM mode recognition by an improved transfer learning method
Optics Letters · 15.11.2026 · 51 (22) · 6271
DOI: 10.1364/ol.612217 ↗Kurz gefasst
Leserbrief in der Zeitschrift Optics Letters, online erschienen am 09.11.2026. Verfasst von Junbao Hu und 5 weiteren Autor:innen. Beteiligte Einrichtung: Department of Optoelectronic Engineering, School of Microelectronics & Data Science, Anhui University of Technology, Ma'anshan, Anhui 243002, China. Die Arbeit nennt 21 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Natural Science Research Project of Anhui Educational Committee.
Abstract (Originalsprache)
Scattering environments severely distort orbital angular momentum (OAM) beams, compromising conventional mode recognition. Deep learning classifiers trained from scratch demand extensive labeled data, while standard transfer learning suffers from slow convergence and loss oscillations due to the substantial domain gap between natural images and random speckle patterns. We propose a two-phase transfer learning framework with lightweight adapter modules for parameter-efficient fine-tuning. Using an ImageNet-pretrained ResNet-50 backbone, we embed adapters in bottleneck layers and implement a progressive freeze-then-unfreeze strategy. The adapters first learn task-specific transformations with the backbone frozen, followed by joint fine-tuning of all parameters. Experiments on a dynamic scattering speckle dataset (four OAM modes) show that our method achieves ~99% test accuracy, outperforming standard deep learning (~50%) and conventional transfer learning (~90%). Convergence time is reduced to 8 minutes, compared with 56 minutes and 35 minutes for the baselines. The framework maintains stable training under varied scattering strengths and scales to 16 modes with 85% accuracy. This approach offers a lightweight and robust solution for intelligent OAM demodulation in resource-limited optical terminals.
Autor:innen und Institutionen
- Department of Optoelectronic Engineering, School of Microelectronics & Data Science, Anhui University of Technology, Ma'anshan, Anhui 243002, China.
- PhotonIQ Lab, School of Microelectronics & Data Science, Anhui University of Technology, Ma'anshan, Anhui 243002, China.
- Research Center for Humanoid Sensing, Zhejiang Lab, Hangzhou, Zhejiang 311100, China.
Förderung
- National Natural Science Foundation of China
- Natural Science Research Project of Anhui Educational Committee
Zitieren
APA
Hu, J., Hu, C., Feng, Y., Lei, Y., Mo, X., & Feng, F. (2026). Scattered OAM mode recognition by an improved transfer learning method. Optics Letters, 51(22), 6271. https://doi.org/10.1364/ol.612217
Vancouver
Hu J, Hu C, Feng Y, Lei Y, Mo X, Feng F. Scattered OAM mode recognition by an improved transfer learning method. Optics Letters. 2026;51(22):6271. doi:10.1364/ol.612217
BibTeX
@article{Hu2026,
author = {Hu, Junbao and Hu, Chunchen and Feng, Yuan and Lei, Yu and Mo, Xutao and Feng, Fu},
title = {Scattered OAM mode recognition by an improved transfer learning method},
journal = {Optics Letters},
year = {2026},
volume = {51},
number = {22},
pages = {6271},
doi = {10.1364/ol.612217}
}
EUVIMED