SADJSCC: Semantic-Aware DJSCC With Attribute Fidelity Evaluation for Channel-Adaptive Image Transmission

Conventional deep joint source-channel coding (DJSCC) methods exhibit semantic misalignment in noisy environments due to their reliance on pixel-level metrics and inability to preserve interpretable semantic attributes, particularly under varying low signal-to-noise ratio (SNR) conditions. Semantic misalignment poses significant risks for safety-critical applications such as telemedicine and autonomous systems, where semantic misinterpretations may result in operational failures. To address these challenges, we propose a semantic-aware deep joint source-channel coding with attribute fidelity evaluation for channel-adaptive image transmission (SADJSCC), a unified framework to address semantic misalignment: semantic feature distortion, latent space displacement, and semantic undetectability. Our SADJSCC introduces a Gaussian discriminant analysis-based semantic-aware module (SAM) that generates coordinated semantic labels and semantic latent vectors through expectation-maximization optimization, resolving latent space displacement while enhancing feature dispersion. The semantic labels guide both the encoder and decoder to achieve semantic alignment. We further incorporate an attention fusion module (AFM) employing multi-head attention to dynamically weight semantic and feature latent spaces, thereby mitigating noise-induced distortions during transmission. To enable quantifiable semantic assessment independent of pixel corruption, we further develop a Semantic Attribute Fidelity Evaluation (SAFE) module that leverages Bayesian counterfactual mechanisms. Moreover, the framework integrates an SNR adaptive module (SNRAM) that dynamically recalibrates features in response to channel fluctuations. Experimental validation across CelebA, CIFAR-10, and MNIST datasets demonstrates consistent performance gains for SADJSCC. On CelebA, SADJSCC achieves an 8.4 dB peak signal-to-noise ratio (PSNR) improvement over conventional DJSCC at SNR =-4 dB, with structural similarity index measure (SSIM) enhanced by 0.16 and learned perceptual image patch similarity (LPIPS) reduced by 46%, while maintaining high computational efficiency of 12.45 dB/G. Visualizations confirm the ability to preserve semantic coherence independent of pixel-level distortions, validating its utility for safety-sensitive applications.