UC-STARJSCC: Unified Conditional Adaptation for Cross-Channel Image Semantic Transmission

Abstract

Image semantic transmission based on joint source-channel coding (JSCC) has shown promising performance in wireless communications. However, existing JSCC methods are typically optimized for a specific channel condition, resulting in limited adaptation across heterogeneous channels and often requiring model retraining. To address this issue, we propose UC-STARJSCC, a unified cross-channel image semantic transmission framework capable of adapting to diverse channel environments using a single model. Specifically, a unified conditional representation jointly encoding the signal-to-noise ratio (SNR) and channel type is introduced to enable channel-aware semantic adaptation. Furthermore, a Dual-phase Conditional Attention (DCA) module is developed to dynamically recalibrate multi-scale semantic features through conditional enhancement and fine-grained feature modulation. A hybrid-channel sampling strategy is further employed to facilitate unified training across heterogeneous channels. Experimental results show that UC-STARJSCC maintains competitive performance over AWGN channels while significantly enhancing robustness over Rayleigh fading channels. At SNR = 25 dB and CBR = 1/6 under Rayleigh fading, it achieves up to a 2.62 dB PSNR gain on Kodak and a 39.14% LPIPS reduction on CLIC over DeepJSCC-V.