Lightweight Semantic Communication for Wireless Image Transmission

Semantic communication has emerged as a disruptive technology to enhance spectral efficiency through effective information extraction and reconstruction. Although state-of-the-art designs have achieved precise semantic extraction through various advanced models, their significant demand for computational resources renders them unsuitable for resource-constrained devices. To address this challenge, a lightweight joint source-channel coding network is proposed. Specifically, a shift operation is introduced as an alternative to spatial convolutional operation, effectively decoupling feature extraction from computational heavy matrix multiplications. Consequently, a dual shift-based network is designed to enhance high-dimensional semantic extraction capability while significantly reducing computational complexity. Experimental results demonstrate that the proposed scheme balances network lightweightness and reconstruction quality. Compared to convolutional neural network-based and lightweight benchmarks, the proposed model reduces the parameters by 48% and 31%, respectively, while its peak signal-to-noise ratio (PSNR) performance is very close. Moreover, the learned perceptual image patch similarity (LPIPS) achieved by the proposed model outperforms the lightweight baseline scheme.