A Novel Lightweight Deep Joint Source-Channel Coding Framework: Using 1D-CNN for SNR and Compression Rate Adaptation

Deep Joint Source-Channel Coding (DeepJSCC) has emerged as a promising paradigm in semantic communication, driven by the growing demands of the Internet of Things (IoT). Considering the resource constraints of IoT devices and the dynamic characteristics of wireless environments, it is crucial to develop a lightweight and adaptive DeepJSCC framework. However, most existing DeepJSCC methods either rely on complex designs to handle adaptability to varying SNR and compression rate (CR), or overlook this issue entirely, which significantly hinders their practical applicability. To address these challenges, we propose a lightweight semantic communication framework with SNR and CR adaptation (LSCF-SCA), leveraging 1D-CNN to achieve an effective trade-off between system performance and complexity in DeepJSCC. The proposed framework incorporates an adaptive SNR module based on 1D-CNN, which dynamically adjusts semantic features to varying channel conditions. For CR adaptation, predictors are generated via 1D-CNN, enabling instance-wise bandwidth allocation through differentiable sparsity constraints. Additionally, depthwise separable convolutions are employed in the feature extraction stage to reduce model complexity. Experimental results demonstrate that our proposed LSCF-SCA reduces parameters by $\mathbf{7 8. 6 1 \%}$ compared to conventional networks, while preserving a mere drop of under 3% in PSNR. It effectively eliminates the “cliff effect” in the separation coding scheme and achieves an optimal balance between PSNR and bandwidth under varying SNR.