Position-Aided Semantic Communication for Efficient Image Transmission: Design, Implementation, and Experimental Results

Peiwen Jiang, Chao-Kai Wen, Shi Jin, Jun ZhangIEEE Trans. Wireless Communications 2026 · DOI

TL;DR

This paper introduces a novel Position-Aided Semantic Communication (PASC) framework that leverages location data and advanced foundation models (FMs) to drastically reduce image transmission overhead. By synthesizing a reference image based on position, PASC only transmits the differences, achieving significant bandwidth savings and robust performance, especially in challenging wireless environments.

What Problem Does This Paper Address?

Transmitting high-resolution images, especially in real-time scenarios like outdoor surveillance or autonomous driving, consumes substantial bandwidth and is susceptible to wireless channel impairments. Traditional methods like JPEG+LDPC struggle with efficiency and error resilience. Semantic communication aims to improve this by transmitting the meaning of the content rather than raw pixels, often relying on knowledge bases (KBs). However, existing semantic communication methods often require end-to-end training for KBs, limiting their adaptability to dynamic environments and failing to fully utilize readily available information at communication devices. This paper addresses the challenge of efficient and robust image transmission by integrating position information and advanced FMs into a semantic communication framework, making it more adaptable and efficient.

Methods

The core of the PASC framework lies in its three-level structure: effectiveness, semantic, and technical.

  • Effectiveness Level: This level focuses on ensuring the transmitted image is contextually relevant to its position. It uses FM-based image analysis to verify this relevance.
  • Semantic Level: This is where the novelty lies. The transmitter's position is used as a shared KB.
  • View Synthesis: An FM (specifically, a diffusion model) synthesizes a camera-view image (psyn) based on the known position and retrieved map data. This psyn serves as a shared reference for both transmitter and receiver.
  • Difference Calculation: The difference between the original image (p) and the synthesized image (psyn) is calculated. Only significant differences (pdiff) are transmitted, significantly reducing the data payload. A threshold (ε) controls how much of the difference is transmitted.
  • Semantic Encoder-Decoder: A specialized encoder-decoder network is trained to compress and decompress pdiff.
  • DM-based Semantic Reconstruction: At the receiver, the synthesized image (psyn) is combined with the reconstructed pdiff to generate the final image.
  • Technical Level: This level handles the physical transmission using an OFDM transceiver.
  • Adaptive Strategy: An LLM is employed to dynamically adjust transmission parameters (e.g., code rate, threshold ε) based on content and channel conditions, enhancing adaptability.
  • Hardware Implementation: A testbed is developed using NVIDIA Jetson Xavier NX, USRP devices, and a server to validate the real-time performance and challenges of FM integration.

Datasets

  • CVUSA and KITTI datasets: Used for training and testing, providing images with associated positions and satellite maps.
  • Offline mapping service: Used to retrieve bird's-eye view images based on positional data.

Key Insights and Ideas

  • Position as a Shared Knowledge Base: Leveraging the transmitter's location as a shared KB for image synthesis is a powerful way to reduce transmission overhead. This is particularly effective for outdoor scenes where position strongly correlates with visual content.
  • Synthesize-then-Transmit-Difference: Instead of transmitting raw pixels or even semantic features of the entire image, synthesizing a reference image and transmitting only the differences is a highly efficient strategy.
  • FM Adaptability: FMs are crucial for both synthesizing realistic reference images and for adaptively adjusting transmission parameters via an LLM, making the system robust to varying conditions.
  • Trade-offs in Semantic Communication: The paper highlights the trade-offs between bandwidth, complexity, and performance. For instance, higher thresholds (ε) save bandwidth but might sacrifice detail.
  • Real-time FM Integration Challenges: The significant processing time of FMs is a major hurdle for real-time communication. The paper proposes local processing for some modules (like UNet for view synthesis) and offloading complex FMs to a server.

Key Results

  • Significant Bandwidth Reduction: PASC demonstrates substantial bandwidth savings compared to conventional JPEG+LDPC and even other semantic communication methods like JSCC, especially at lower bandwidths (e.g., 2k bits).
  • Robustness under Low SNR: PASC maintains better performance than JSCC(16k) at low SNRs (below 0 dB), where JSCC(16k) degrades rapidly.
  • Superior Image Quality: Across various metrics (LPIPS, SSIM, CLIP-I), PASC consistently outperforms competing methods, particularly under challenging channel conditions. For example, at -5 dB SNR, PASC(2k) performs significantly better than JSCC(2k).
  • Adaptive Performance: The LLM-based adaptive strategy allows PASC to dynamically select optimal parameters, leading to better performance across different scenarios and bandwidth constraints.
  • Hardware Validation: The prototype testbed demonstrates the feasibility of PASC in a real-world setting, although it highlights the time cost of DM-based reconstruction.

Limitations and Future Work

  • FM Processing Latency: The primary limitation is the significant processing time of FMs, especially for DM-based reconstruction, which can take seconds. This hinders true real-time operation for some modules.
  • Future Work: The authors suggest exploring acceleration techniques like distillation to reduce diffusion steps and improve speed.
  • Mismatch Handling: While PASC is designed to be robust, mismatches between synthesized and actual images (due to outdated KBs or position errors) can still affect performance.
  • Future Work: The paper suggests simple local judgment strategies to detect mismatches and potentially fallback mechanisms (like JSCC) to ensure performance.
  • Generalizability of LLM Recommendations: The LLM's recommendations are based on available performance data. To improve future recommendations, new performance data from trained encoder-decoder pairs under novel settings should be incorporated.
  • Complexity of Implementation: Integrating FMs into communication systems requires careful consideration of computational resources and real-time constraints. The proposed testbed provides a step towards this, but further optimization is needed.

Summary updated

BibTeX
@article{jiang2026position,
  author = {Peiwen Jiang and Chao-Kai Wen and Shi Jin and Jun Zhang},
  title = {{Position-Aided Semantic Communication for Efficient Image Transmission: Design, Implementation, and Experimental Results}},
  journal = {IEEE Trans. Wireless Communications},
  year = {2026},
  doi = {10.1109/twc.2025.3614765},
}
Figure