Lightweight CNN-Based Semantic Communication for Image Analysis

  • Son Do, Tung; 
  • Tuan Do, Quang; 
  • Cuong Ho, Manh; 
  • Phuoc Nguyen, Minh; 
  • Mahrab Islam, Md.; 
  • ... Cho, Sungrae
Citations

WEB OF SCIENCE

0
Citations

SCOPUS

0

초록

This paper presents GnartDeepSC, a lightweight semantic communication system for IoT image analysis. We introduce the Lightweight Semantic Gated CNN (LSGCNN) block that combines multi-branch architecture with adaptive gating to efficiently extract task-relevant features. Unlike computationally intensive ViT-based methods, GnartDeepSC achieves 90.2% parameter reduction while maintaining 48.3% classification accuracy at –6 dB SNR (13.1% improvement over ViTDeepSC). The system supports simultaneous image reconstruction and classification with 3.0× faster inference (2.84 ms average latency) and 34.8% energy savings (96.22 mJ per image). Experimental results demonstrate superior robustness across all SNR levels, achieving 42.1 dB PSNR at 18 dB SNR. These characteristics make GnartDeepSC ideal for resource-constrained IoT deployments requiring real-time semantic communication.

키워드

IoT; Lighweight; Semantic communication
제목
Lightweight CNN-Based Semantic Communication for Image Analysis
저자
Son Do, Tung; Tuan Do, Quang; Cuong Ho, Manh; Phuoc Nguyen, Minh; Mahrab Islam, Md.; Cho, Sungrae
DOI
10.1007/978-3-032-13102-7_20
발행일
2026
유형
Proceedings Paper
저널명
Lecture Notes on Data Engineering and Communications Technologies
권
281
페이지
253 ~ 264