Semantic Feature Compression and Adaption for Latency-Sensitive Semantic Communication

IEEE Globecom Workshops (GC Wkshps), pp. 1–6

Abstract

Nowadays, semantic communication has been considered as a promising field on improving the communication efficiency by transmitting the meaningful content of information. Despite great progress, there remain significant challenges in communication scenarios with sensitive latency constraints. In this paper, we investigate a deep learning-based semantic communication framework for latency-constrained semantic information transfer and consider image classification as the example semantic task. The proposed framework starts with a semantic feature extractor, that extracts a captured image to a feature map. The low latency requirement forces the application of finite blocklength (FBL) codes, such that the transmission reliability highly relies on transmitted data amount and available symbols resulting from latency constraints. To deal with this challenge, we introduce compression module including SparseConv block and ultra-low precision quantization, achieving compressed feature map with 128× compression ratio, and propose two approaches for robust feature adaptions, allowing adaptive feature selection and coding rate designs with respect to different latency demands. In particular, space-based and channel-based feature adaption training are performed to learn the optimal feature information in limited blocklength. At the semantic receiver, these symbols carrying feature information will be reconstructed and decoded for labelling the image. Finally, numerical results verify the promising performance of our proposed framework for latency-constrained semantic communication, as well as the high adaptability of our proposed design with respect to low latency requirements.

Authors 4

1. Huanyu Zhang Aachen INDA Chair

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University,INDA Chair,Germany

2. Yulin Hu

   Wuhan University

   Affiliation as printed

   Wuhan University,School of Electronic Information,China

3. Xiaopeng Yuan Aachen INDA Chair

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University,INDA Chair,Germany

4. Anke Schmeink Aachen INDA Chair

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University,INDA Chair,Germany