Wireless Vision-Centered Semantic Communication for Smart City Environment: Pretrained Network and Quantization

This paper introduces a Vision-Centered Semantic Communication (VCSC) system tailored for efficient image transmission in smart city environments, where bandwidth is limited and channels are subject to severe noise. Unlike conventional text-centered or classical compression approaches, VCSC leverages a pretrained latent encoder-decoder network to extract compact, semantically rich representations directly from images. An innovative attention-based quantization strategy is employed to selectively allocate higher precision to critical regions, thereby reducing the overall bit rate while preserving essential semantic details. The quantized latent codes are robustly transmitted over wireless channels modeled with additive white Gaussian noise and Rayleigh fading. An end-to-end training framework minimizes both reconstruction and perceptual losses, ensuring high-fidelity image recovery even under adverse conditions. Extensive simulations demonstrate that VCSC outperforms traditional methods in preserving fine-grained details and semantic integrity, offering a promising solution for real-time surveillance, transportation, and infrastructure monitoring in smart cities.