Towards generative knowledge-collaborative semantic communication for extreme traffic environments

While Semantic Communication (SC) is a promising paradigm for Intelligent Transportation Systems (ITS) by prioritizing task-relevant information, existing frameworks utilizing static knowledge bases or discriminative decoding often struggle to adapt to extreme traffic environments (e.g., nighttime, rain, and snow), where degraded visual inputs and low-SNR wireless channels cause severe semantic impairment and irreversible information loss. To overcome this perception bottleneck, this paper proposes the Generative Knowledge-Collaborative Semantic Communication (GKC-SC) framework. First, an environment-adaptive Semantic Knowledge Base (SKB) is constructed by integrating Retinex physics decoupling, diffusion-based texture restoration, and the Segment Anything Model (SAM) to extract high-precision semantic masks. Second, the transmitter leverages these SKB-derived masks for Adaptive Semantic Compression (ASC) alongside a Bi-level Routing Attention (BRA) module, effectively suppressing background redundancy to focus on critical traffic regions. Finally, a Guided Diffusion Receiver (GDR) is designed to jointly decode the transmitted discrete indices and SKB generative priors, enabling the active completion of high-frequency details and semantic structures. Extensive experiments on the BDD100K and ACDC datasets across a wide SNR range (-10 dB to 25 dB) demonstrate GKC-SC’s superiority. The framework achieves a mean Average Precision (mAP@0.5) of 0.918 on BDD100K and 0.935 on ACDC, while significantly reducing Learned Perceptual Image Patch Similarity (LPIPS) scores compared to baseline methods.