A lightweight image transmission scheme for deep learning-based semantic communication

Abstract

In this study, a lightweight semantic communication scheme is proposed to address the issues of large storage overhead and limited adaptability caused by the dynamic signal-to-noise ratio (SNR) variations in wireless channels, which necessitate storing multiple SNR-specific models in conventional semantic communication schemes. The scheme integrates ghost convolution (GC) and a global attention mechanism (GAM) module. Experimental results demonstrate: a 30% reduction in parameters, FLOPs, and model storage while maintaining comparable performance to conventional convolutional models; under the 3dB Gaussian channel, the Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure(SSIM) reach 27dB and 0.9 respectively; under the Rayleigh channel at identical SNR, PSNR and SSIM achieve 25dB and 0.8. The approach significantly reduces deep learning model complexity while preserving image reconstruction quality, effectively solving the multi-SNR adaptation versus storage trade-off, and establishing a new idea for the image semantic communication scheme.