A Semantic Coding Scheme for Robust Image Transmission over Noisy Channels

Separation-based pipelines for wireless image transmission suffers a cliff effect at low signal-to-noise ratios. While deep joint source-channel coding may alleviate this drawback, most of the designs ignore readily available side information during decoding. To overcome these drawbacks, we introduce a semantic JSCC scheme in which a convolutional encoder maps an image to a spatial latent vector that traverses through an AWGN channel, and a decoder that is conditioned on class labels and instantaneous SNR values via learnable embeddings. To this end, we trained our model end-to-end with a mixed loss combining L1 fidelity and structural similarity. Our method offers a threefold contributions such as label- and SNR-conditioned decoding, preservation of spatial latent structure, and a simple, compute-efficient objective yielding robust reconstructions under AWGN. Evaluated on STL-10 across a range of SNRs, the proposed method consistently improves both perceptual and distortion metrics over a non-conditioned JSCC baseline. At 5 dB SNR, it improves PSNR value by 4.86 percent and improves LPIPS value by 28 percent. Qualitative reconstructions show reduced over-smoothing and sharper class-relevant details at low SNR. Overall, the results indicate that we provide practical gains especially for compute-constrained noisy wireless links which are used to transmit visual information.