IoT‐Oriented Lightweight Joint Source‐Channel Coding for Image Transmission via Filter Pruning in Semantic Communication

ABSTRACT

Traditional communication systems follow Shannon's bit‐level paradigm, prioritizing reliable bit transmission without regard to semantic content. Semantic Communications shift the focus to conveying task‐relevant meaning, offering substantial efficiency gains in resource‐constrained scenarios such as the Internet of Things (IoT) and edge computing. The Attention‐based Deep Joint Source–Channel Coding (ADJSCC) network achieves robust image reconstruction over wide Signal to Noise ratio (SNR) ranges by adaptively allocating resources between source compression and channel protection. However, its high computational complexity prevents deployment on energy‐ and memory‐limited devices. This paper proposes a Lightweight ADJSCC model obtained through structured Filter Pruning guided by information‐capacity and independence metrics, followed by multi‐objective fine‐tuning using a composite loss that balances pixel‐wise distortion (measured by Mean Square Error (MSE)), structural fidelity (measured by Structural Dissimilarity Index Measure (DSSIM)), and perceptual quality (measured by Learned Perceptual Image Patch Similarity (LPIPS)) with Adaptive Gradient Balancing (AGB). Results on the CIFAR‐10 dataset over Additive White Gaussian Noise (AWGN) channels show that the original ADJSCC model reconstructs images with the highest perceptual quality, maintaining sharp details and natural colours across varying SNRs. The pruned model fine‐tuned with MSE‐only loss exhibits noticeable degradation, including amplified noise, colour distortion, and texture loss, especially at lower SNRs. The proposed Lightweight ADJSCC model produces reconstructions that are qualitatively superior to the MSE‐only baseline, demonstrating better noise suppression and smoother gradients, particularly in low‐SNR conditions. Visual examples indicate that the proposed model's multi‐objective fine‐tuning strategy helps preserve structural integrity and natural statistics, resulting in reconstructions that are perceptually closer to the original, high‐fidelity model. These findings highlight the effectiveness of perception‐ and structure‐aware recovery in mitigating pruning‐induced semantic degradation, making the proposed lightweight design suitable for edge‐intelligent semantic image transmission in future 6G systems. The proposed framework mitigates pruning‐induced semantic degradation in ADJSCC networks through information‐aware pruning and perceptual‐structural fine‐tuning.