NICGM: A non-intrusive CSI-guided modulation method for image semantic communication

The incorporation of channel state information (CSI) has been proven effective in enhancing the robustness of transmission in image semantic communication systems. However, existing studies have largely overlooked the modal inconsistency between CSI and semantic features. Specifically, the intrusion of physical features into the semantic space can significantly degrade the reconstruction quality and hinder practical deployment under complex channel conditions. To address this issue, a novel non-intrusive CSI-guided modulation mechanism is proposed in this paper. The mechanism integrates affine modulation and gating control to structurally guide the fusion of CSI into the semantic feature space, effectively mitigating the destructive interference of CSI. Furthermore, a lightweight CSI encoder is introduced to accommodate feedback distortion and compression constraints. A modulation perturbation regularization term is designed to dynamically control the modulation intensity. During system training, perceptual loss and structural reconstruction loss are jointly optimized, while a modulation deviation monitoring indicator is employed to enhance non-intrusive behavior constraints. Experimental results demonstrate that, under low signal-to-noise ratio and distorted CSI scenarios, the proposed NICGM significantly outperforms baseline methods such as JSCC and JPEG in terms of PSNR and perceptual consistency. These findings validate the structural stability and performance advantages of the proposed model in non-ideal wireless environments.