VLCnet: Deep learning based end-to-end visible light communication system

Ulkar M. G., Baykas T., Pusane A. E.

Journal of Lightwave Technology, vol.38, no.21, pp.5937-5948, 2020 (SCI-Expanded) identifier identifier

  • Publication Type: Article / Article
  • Volume: 38 Issue: 21
  • Publication Date: 2020
  • Doi Number: 10.1109/jlt.2020.3006827
  • Journal Name: Journal of Lightwave Technology
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, PASCAL, Aerospace Database, Communication Abstracts, Compendex, INSPEC, Metadex, Civil Engineering Abstracts
  • Page Numbers: pp.5937-5948
  • Keywords: Visible light communication, Deep learning, Lighting, Neural networks, Decoding, Training, Visible light communications, autoencoder, competitive learning, flicker reduction, input dependent noise
  • Istanbul Medipol University Affiliated: Yes


Visible light communication is a popular research area where proposed communication methods must satisfy the lighting related requirements as well. Suggested VLC modules should not only improve communication quality such as decreasing error rates but also comply with other lighting related constraints such as sustaining certain level of illumination. This increases the complexity of the optimization problem. Moreover, most of the time the suggested modules focus on a specific block of communication system which downgrades the system-wide performance on coming together. To solve this complex problem and jointly optimize the whole system, we suggest a deep learning based method, VLCnet. Despite the increasing number of neural network based channel decoders in the literature, few of them are addressing real-life application constraints. VLCnet is an error rate decreasing solution which takes into account, reducing flicker and sustaining certain illumination level. Moreover, our channel impulse response (CIR) is taken from reference CIRs for VLC and our study considers the input-dependent noise originated by the shot noise for the sake of generality. Flicker reducing activation units (FRAU) are the key part of VLCnet and the main novelty of this publication. FRAU is an example of a competitive layer and ensures run length limitation for flicker reduction. Both with input-independent and dependent noise, simulation results show performance superiority of the proposed VLCnet method. Although they have different setups, all results demonstrate the benefit of training with certain amount of noise. From the practicality perspective, proposed system is easy to be deployed since inference operation does not have iterations unlike most of the conventional detectors.