U-Net-Based Generative Joint Source-Channel Coding for Wireless Image Transmission

A research team from multiple institutions has submitted a paper proposing two novel AI-driven methods for transmitting images over unreliable wireless channels. The work, titled "U-Net-Based Generative Joint Source-Channel Coding for Wireless Image Transmission," addresses a key bottleneck in modern communication: sending high-quality visual data through noisy environments. Traditional and even some deep learning-based Joint Source-Channel Coding (JSCC) methods often sacrifice perceptual quality for pixel-level accuracy or vice versa, and can be computationally heavy. The team's new approaches leverage generative AI architectures to overcome these limitations, aiming for a better balance between clarity, realism, and efficiency.

The researchers developed two specific models. The first, G-UNet-JSCC, employs a U-Net-based generator as its decoder. The U-Net's signature skip connections allow for multi-scale feature fusion, integrating both low-level details and high-level semantics to improve reconstruction. It's trained using a weighted combination of Structural Similarity (SSIM) and Mean-Squared Error (MSE) losses. Building on this, the more advanced cGAN-JSCC incorporates adversarial training. It uses the same encoder but pairs the U-Net generator with a patch-based discriminator in a two-stage training process, combining adversarial and distortion losses. Simulation results show both methods achieve superior fidelity and perceptual quality. Notably, cGAN-JSCC demonstrates better performance and greater robustness to channel noise for low-resolution images, a critical finding for real-world applications like mobile video calls or IoT devices. This work represents a significant step toward more resilient and visually pleasing wireless multimedia systems.