Effects of Training Images on CNN-Based Demodulation for Digital Signage and Image Sensor-Based VLC

Abstract

This paper studies a visible light communication (VLC) system using a digital signage and an image sensor. The authors have focused on the demodulation part of the communication system, which modulates data signals without disturbing the visual information on the digital signage, and have proposed a novel concept that uses machine learning to demodulate the data signals from images received by the image sensor. However, it has not been fully clarified which parameters of the training images contribute to the performance of the machine learningbased demodulation. This paper extends the convolutional neural network (CNN)-based demodulation method and clarifies how much the number of parallelized data signals and the number of patterns of data signals in the training images contribute to the demodulation performance. The results show that the performance improves with the number of parallelized data signals in the training images, and that half of the signal patterns are sufficient for learning.