TranGDeepSC: Leveraging ViT knowledge in CNN-based semantic communication system
Tung Son, Thanh Phung Truong, Quang Tuan, Sungrae Cho
Abstract
This paper introduces TranGDeepSC, a lightweight CNN-based deep semantic communication (DeepSC) system that leverages Vision Transformer (ViT) knowledge through co-training to enhance image transmission. Evaluated on CIFAR-100 across various SNRs, TranGDeepSC demonstrates competitive performance with ViTDeepSC, and outperforms SemViT and ADJSCC-V in image quality, particularly in low-SNR environments. Notably, it offers substantial gains in efficiency: 92.8% fewer parameters than ADJSCC-V, 72.0% lower energy use, and 48% faster processing than ViTDeepSC. These advantages make TranGDeepSC well-suited for resource-constrained applications in next-generation communication systems, including 6G, IoT, and real-time multimedia streaming.