Resource-Aware Sparse Federated Image Semantic Communication with Dual-Side U-Net Fusion

Abstract

This paper proposes a resource-aware sparse federated image semantic communication framework with dual-side U-Net fusion to improve reconstruction quality and reduce federated training overhead in distributed image transmission scenarios. Firstly, lightweight U-Net modules are introduced at both the transmitter and receiver to extract and exploit useful semantic and structural information for feature fusion and reconstruction compensation. An information noise-contrastive estimation (InfoNCE)-based feature-level consistency constraint is further incorporated, extending the optimization from pixel-level reconstruction to feature-level semantic representation preservation. Secondly, to address the communication burden caused by parameter-intensive semantic communication models in federated training, a resource-aware regularization term is designed to guide local optimization toward communication-efficient update distributions while preserving reconstruction-relevant parameter changes. Thirdly, a sparse federated optimization mechanism is developed by combining importance-guided update control, magnitude-aware sparse transmission, dense warm-up, and residual accumulation, thereby reducing both uplink and downlink communication costs. Experimental results show that the proposed framework improves reconstruction quality and feature-level semantic consistency while achieving more efficient federated training.