Semantics-Guided Contrastive Joint Source-Channel Coding for Image Transmission

Deepjoint source-channel coding (D-JSCC) provides several advantages over conventional coding schemes, in which source and channel coding are designed separately. For example, D-JSCC schemes suffer from smaller delays and are more robust to rapid channel variation. However, D-JSCCs are often designed without explicit structure or insight, making them less adaptive, hard to control, and theoretically unfounded. In this paper, we propose a contrastive joint-source-channel coding (C-JSCC) design, which uses supervised contrastive learning (SCL) to make the latent space of a conventional D-JSCC more structured and meaningful. By testing on the CIFAR-10 dataset, we show that C-JSCC consistently outperforms its D-JSCC counterpart in both tasks of image reconstruction and classification. Moreover, C-JSCC is shown to output images with perceptual quality better than the classic BPG image codec in the low bits-per-pixel (bpp) region. The roles of various hyper-parameters in C-JSCC are investigated by analytical approximations, experiments, and visualization techniques.