Semantic communication based on bi-level routing attention in IoT environment

Yan Xiao, Fan Xiumei, Kok-Lim Alvin Yau, Zhixin Xie, Rui Men

Abstract: Traditional communication methods rely on meticulously designed system modules to ensure accurate bit-level transmission. However, with the growing complexity and volume of data, these approaches are reaching the Shannon limit. Semantic communication (SC) emerges as a promising solution that emphasizes meaning rather than bit-level precision. Despite its potential, SC still faces challenges in efficiently extracting and preserving relevant features, especially in resource-constrained environments. This article proposes a SC method that utilizes bi-level routing attention (BRA), named BRASC. By capturing both region-to-region and token-to-token attention mechanisms, BRA captures broad meaning as well as specific details, retaining key semantic information while eliminating less relevant data, thereby enhancing feature representation and improving transmission accuracy. Experimental results demonstrate that BRASC consistently outperforms Original Images Direct Classification (OIDC), Deep Learning (ResNet18), and Baseline methods in image classification accuracy on the STL-10 dataset. BRASC shows superior performance across varying signal-to-noise ratios (SNR), maintaining high accuracy and stability, particularly excelling in low SNR conditions. This robustness across diverse channel conditions confirms BRASC’s effectiveness and adaptability for semantic communication in challenging communication scenarios.