Noise-robust optical convolutional neural networks with learnable bit-plane weighted fusion
Optics Letters · 15.11.2026 · 51 (22) · 6359
DOI: 10.1364/ol.615383 ↗Kurz gefasst
Leserbrief in der Zeitschrift Optics Letters, online erschienen am 11.11.2026. Verfasst von Ruoxing Guo und 5 weiteren Autor:innen. Beteiligte Einrichtung: School of Electronic Engineering, Beijing University of Posts and Telecommunications, Beijing 100876, China. Die Arbeit nennt 24 Literaturangaben. Gefördert von: National Natural Science Foundation of China.
Abstract (Originalsprache)
Optical convolutional neural networks (OCNNs) are susceptible to modulation, alignment, and detection errors, which cause the accuracy of a deployed optical–electronic network to deviate from its numerical model. We propose a learnable bit-plane weighted-fusion (AWF) strategy for low-dimensional offline calibration after camera detection. Each 8-bit input image is decomposed into eight binary planes, convolved separately, and reconstructed using eight learnable normalized weights before electronic activation, pooling, and classification. With 5000 MNIST training images, the optical neural network (ONN) using AWF reaches 92.6% training accuracy, 1.9% points above the traditional binary weighted-fusion (TWF) method. The corresponding experimentally measured test accuracies are 90.30% and 88.69%, respectively. Under 5 dB added noise, AWF-ONN maintains 83.84% accuracy, 5.51% points above TWF, while its loss relative to the no-added-noise condition is 6.46% points compared with 10.36% points for TWF-ONN. A two-layer network retains an experimental advantage of 1.12% points. The results show that a low-dimensional calibration of bit-plane contributions can improve optical inference when labeled data or signal-to-noise ratio is limited.
Autor:innen und Institutionen
- School of Electronic Engineering, Beijing University of Posts and Telecommunications, Beijing 100876, China.
- School of Artificial Intelligence, Beijing Normal University, Beijing 100875, China.
ORCID-Profile: Yongpan Gao, Chuan Wang
Förderung
- National Natural Science Foundation of China
Zitieren
APA
Guo, R., Nie, Z., Zhou, Y., Liu, X., Gao, Y., & Wang, C. (2026). Noise-robust optical convolutional neural networks with learnable bit-plane weighted fusion. Optics Letters, 51(22), 6359. https://doi.org/10.1364/ol.615383
Vancouver
Guo R, Nie Z, Zhou Y, Liu X, Gao Y, Wang C. Noise-robust optical convolutional neural networks with learnable bit-plane weighted fusion. Optics Letters. 2026;51(22):6359. doi:10.1364/ol.615383
BibTeX
@article{Guo2026,
author = {Guo, Ruoxing and Nie, Zhiguo and Zhou, Yulu and Liu, Xinchang and Gao, Yongpan and Wang, Chuan},
title = {Noise-robust optical convolutional neural networks with learnable bit-plane weighted fusion},
journal = {Optics Letters},
year = {2026},
volume = {51},
number = {22},
pages = {6359},
doi = {10.1364/ol.615383}
}
EUVIMED