Computation and transmission adaptive semantic communication for reliability-guarantee image reconstruction in IoT

Abstract

Semantic communication can significantly compress source data, improving transmission efficiency. However, semantic communication systems under varying channel condition have not been well studied, especially for tasks that require reliability guarantee. This paper focus on the reliability-guarantee image reconstruction tasks, and study the computation and transmission adaptive semantic communication. First, a computation and transmission adaptive semantic communication (CTASC) system is proposed. It is able to adjust the computation load and transmission load of an image reconstruction task adaptively while guaranteeing the reconstruction reliability. Specifically, a semantic encoder with multiple convolutional neural network (CNN) slices with different network depths is designed to adjust the transmission load and computation load. Second, a joint computation and transmission resource allocation problem aimed at minimizing the maximum delay within system is formulated. To solve this problem, we decompose it into two nested sub-problems and propose a Simulated Annealing with Re-perturbation Mechanism (SA-RPM) algorithm and an Alternating Optimization (AO) algorithm to solve these sub-problems, respectively. Simulation results demonstrate that compared to variable code length enabled DeepJSCC (DeepJSCC-V), our system can achieve higher compression ratio(CR) with similar LPIPS performance. Simulation results also show that our resource allocation scheme can obtain an approaching value to the optimal maximum delay, with an average difference not exceeding 0.2%.