An Enhanced Transfer Learning-Based Automatic Modulation Recognition Method for OFDM-VLC System

In this paper, an enhanced transfer learning-based automatic modulation recognition (AMR) scheme is proposed and experimentally demonstrated in orthogonal frequency division multiplexing visible light communication (OFDM-VLC) systems. Within the proposed Gaussian transfer learning (GTL) framework, a ResNet deep convolutional neural network (DCNN) is constructed using Gaussian constellations as the training set. The performance of the proposed method is evaluated by using GoogLeNet, AlexNet and ImageNet transfer learning (ITL) as benchmarks. Experimental results show that, using only 120 training samples with a limited number of constellation points per sample, ResNet with ITL achieves accuracy improvements of 13% and 7% over GoogLeNet and AlexNet with ITL, respectively. Furthermore, ResNet with the proposed GTL achieves an additional 4% accuracy enhancement compared to ResNet utilising conventional ITL.