LightSemCom: Edge-Guided Lightweight Semantic Communication for Image Restoration

Deploying high-performance semantic communication systems on resource-constrained end devices remains challenging, particularly for delay-sensitive image restoration. While existing CNN- and Transformer-based joint source-channel coding (JSCC) methods achieve compelling performance, their high parameter counts and computational complexity hinder deployment under strict memory and latency constraints. To address this challenge, we propose a lightweight semantic communication framework (LightSemCom) with edge-guided progressive distillation for efficient device-side encoding and decoding. Unlike conventional end-to-end training, our approach employs a multi-stage, feature-wise knowledge transfer strategy in which the edge server progressively guides different modules of the device encoder-decoder. This enables the lightweight model to capture task-relevant semantic features while maintaining robustness under varying channel conditions (e.g., noisy, low-SNR environments). Experimental results show that LightSemCom reduces model size by over 36.7% and inference latency by 66.9% compared to SwinJSCC and DeepJSCC-V, while achieving competitive PSNR and SSIM in image reconstruction. The proposed system offers a practical solution for semantic communication in resource-limited scenarios such as multi-robot collaboration and IoT-based monitoring.