Ghost Module and Transformer-Based Lightweight End-to-End Communication Without Pilots

Abstract

Research on pilot-free communication has attracted significant attention for its potential to reduce pilot overhead and enhance spectral efficiency. Despite their competitive performance compared with traditional methods, existing pilot-free end-to-end (E2E) schemes often suffer from high computational complexity and limited BER performance. In this latter, we propose a deep learning-based lightweight pilot-free E2E communication system for frequency-selective fading channels to address these challenges. The proposed network employs the Ghost module to lighten the network design, develops an effective feature extractor by integrating the Dense Block with bilinear production, and adopts a modified transformer-based decoder to achieve precise decoding. Experimental results show that our method achieves the best BER performance and the lowest computational complexity compared to existing pilot-free approaches.