Engineering a Lightweight Deep Joint Source-Channel Coding Based Semantic Communication System

Deep joint source-channel coding (DeepJSCC) has emerged as a novel technology in semantic communication, coinciding with the increasing demand for edge devices in the Internet of Things (IoT). Consequently, the deployment of DeepJSCC on edge devices has become a crucial research direction. However, DeepJSCC faces challenges related to channel fading. Moreover, implementing DeepJSCC on edge devices poses challenges due to the constrained computational resources as well as the compatibility issue between DeepJSCC and digital systems. In this paper, we devote to engineering the DeepJSCC system deployed on edge devices. First, we propose a method named DeepJSCC with Ensemble learning (DeepJSCC-ES) to resist the channel fading. Then, we present a pruning algorithm called the DeepJSCC SNR-Adaptive Pruning method (DJSAP) to make the DeepJSCC network lightweight, reducing the computational demands of on edge nodes. Further, we propose a method called the Simulated Fixed-Point Quantization training based on Soft Quantization function (SFPQSQ) to tackle the compatibility issue between DeepJSCC and digital systems. Finally, we deploy the whole DeepJSCC system on edge devices and conduct experiments to test the DeepJSCC system. The results of simulations show that the proposed DeepJSCC-ES system outperforms the baseline DeepJSCC, particularly excelling in low SNR conditions. Furthermore, the parameter size of the pruned model using DJSAP is compressed by 93.37% while the average structural similarity index metric (SSIM) decreases only by 0.92% compared with the baseline DeepJSCC. Additionally, the SFPQSQ works better than ordinary quantization methods in tackling the compatibility issue between DeepJSCC and digital systems. The experiment results also show that our proposed system can serve as a feasible solution for practical deployment on edge devices.