Neural Network Training System for Marker Encoding
- Authors: Wang L.P.1, Grinchuk O.V.2, Tsurkov V.I.2,3
-
Affiliations:
- Nanjing University of Aeronautics and Astronautics
- Moscow Institute of Physics and Technology
- Federal Research Center Computer Science and Control, Russian Academy of Sciences
- Issue: Vol 58, No 3 (2019)
- Pages: 434-440
- Section: Artificial Intelligence
- URL: https://journals.rcsi.science/1064-2307/article/view/220387
- DOI: https://doi.org/10.1134/S1064230719030195
- ID: 220387
Cite item
Abstract
In this paper, we propose a training system for visual markers that provides the generation and subsequent recognition (under real-world conditions) of stylized images that contain information encoded by a sequence of bits. New types of neural network layers that make the recognition process tolerant to external noise are developed. The training process is based on the end-to-end principle, which enables automatic training for the intermediate stages of the model. The experimental results that demonstrate the performance of the system in encoding and decoding artificial markers are presented.
About the authors
L. P. Wang
Nanjing University of Aeronautics and Astronautics
Email: tsurkov@ccas.ru
China, Nanjing
O. V. Grinchuk
Moscow Institute of Physics and Technology
Author for correspondence.
Email: oleg.grinchuk@phystehc.edu
Russian Federation, Dolgoprudny, 141701
V. I. Tsurkov
Moscow Institute of Physics and Technology; Federal Research Center Computer Science and Control, Russian Academy of Sciences
Author for correspondence.
Email: tsurkov@ccas.ru
Russian Federation, Dolgoprudny, 141701; Moscow, 119991